Follow-up experiment · 30 October 2026

Put the topic first.
Then check the meaning.

The shorter topic-first prompt improved front-loading in this sample, with fewer observed ellipsis endings. It still produced one missing title and did not make every collection consistent. Three models, three prompts, 24 requests. This follow-up tests a shorter topic-first override against both the SDK default and the previous action-first override. GPT-6 Luna joins through a ChatGPT subscription with low reasoning requested.

Read the unchanged first experiment ↗

Front-loading improves. Collection structure still varies.

432 generation attempts: 24 cases × three prompts × two repeats × three models. 431 returned a displayed title. Gemini returned no usable title once for the topic-first long-input case; that missing title remains visible below. The quality charts exclude it and show their actual successful-title denominators. No combined “quality score” hides the differences.

Default

117/144front-loading passes

2 fidelity failures · 12 ellipsis endings · 0 missing titles

Previous override

106/144front-loading passes

6 fidelity failures · 22 ellipsis endings · 0 missing titles

Topic-first override

133/143front-loading passes

1 fidelity failures · 9 ellipsis endings · 1 missing titles

Green: pass. Amber: partial. Red: fail. Connotation applies only to feedback and the neutral database comparison. Fidelity checks whether the title still describes the requested work.

Per-title criterion	Default	Previous override	Topic-first override
Match how people parse	143 pass · 1 partial / 144	143 pass · 1 partial / 144	130 pass · 13 partial / 143
Put distinguishing words first	117 pass · 26 partial · 1 fail / 144	106 pass · 38 partial / 144	133 pass · 10 partial / 143
Match connotation to purpose	29 pass · 13 partial / 42	36 pass · 6 partial / 42	34 pass · 8 partial / 42
Requested-work fidelity	142 pass · 2 fail / 144	138 pass · 6 fail / 144	142 pass · 1 fail / 143

Scoring audit: deduplicating identical text did not prevent inconsistent treatment of equivalent wording across AI batches. Before publication, nine unique grades were adjudicated under the fixed rubric: six inspect/test fidelity grades and three clipped-detail parse grades. The tables use those corrected grades. Original blind judgments · Every correction.

Pooled descriptive counts only. The models use different transports and sampling controls; inspect each model before generalizing.

Haiku 4.5 · all cases

Per-title criterion	Default	Previous override	Topic-first override
Match how people parse	48 pass / 48	47 pass · 1 partial / 48	40 pass · 8 partial / 48
Put distinguishing words first	37 pass · 11 partial / 48	38 pass · 10 partial / 48	43 pass · 5 partial / 48
Match connotation to purpose	10 pass · 4 partial / 14	10 pass · 4 partial / 14	9 pass · 5 partial / 14
Requested-work fidelity	48 pass / 48	43 pass · 5 fail / 48	47 pass · 1 fail / 48

Gemini 3 Flash Preview · all cases

Per-title criterion	Default	Previous override	Topic-first override
Match how people parse	48 pass / 48	48 pass / 48	45 pass · 2 partial / 47
Put distinguishing words first	43 pass · 5 partial / 48	40 pass · 8 partial / 48	42 pass · 5 partial / 47
Match connotation to purpose	9 pass · 5 partial / 14	12 pass · 2 partial / 14	12 pass · 2 partial / 14
Requested-work fidelity	46 pass · 2 fail / 48	48 pass / 48	47 pass / 47

GPT-6 Luna · low · all cases

Per-title criterion	Default	Previous override	Topic-first override
Match how people parse	47 pass · 1 partial / 48	48 pass / 48	45 pass · 3 partial / 48
Put distinguishing words first	37 pass · 10 partial · 1 fail / 48	28 pass · 20 partial / 48	48 pass / 48
Match connotation to purpose	10 pass · 4 partial / 14	14 pass / 14	13 pass · 1 partial / 14
Requested-work fidelity	48 pass / 48	47 pass · 1 fail / 48	48 pass / 48

Separate familiar examples from fresh ones.

The 16 reused requests informed the prompt revision. Eight fresh requests have no outputs in the first study. They were still designed by the same researcher after seeing that study, so they are not an independent external benchmark.

Reused cases

Per-title criterion	Default	Previous override	Topic-first override
Match how people parse	96 pass / 96	95 pass · 1 partial / 96	86 pass · 9 partial / 95
Put distinguishing words first	88 pass · 8 partial / 96	80 pass · 16 partial / 96	91 pass · 4 partial / 95
Match connotation to purpose	18 pass · 6 partial / 24	20 pass · 4 partial / 24	22 pass · 2 partial / 24
Requested-work fidelity	94 pass · 2 fail / 96	92 pass · 4 fail / 96	94 pass · 1 fail / 95

Fresh cases

Per-title criterion	Default	Previous override	Topic-first override
Match how people parse	47 pass · 1 partial / 48	48 pass / 48	44 pass · 4 partial / 48
Put distinguishing words first	29 pass · 18 partial · 1 fail / 48	26 pass · 22 partial / 48	42 pass · 6 partial / 48
Match connotation to purpose	11 pass · 7 partial / 18	16 pass · 2 partial / 18	12 pass · 6 partial / 18
Requested-work fidelity	48 pass / 48	46 pass · 2 fail / 48	48 pass / 48

Collections have their own denominators.

Important tradeoff: full consistency passes fell from 10/12 default groups to 4/12 topic-first groups; the previous override passed 9/12. Topic-first subjects did not produce a shared grammatical pattern reliably. Consistency is scored across auth and upgrade siblings. Meaningful variation also includes feedback and backup-operation groups. Each group belongs to one model, prompt, and repeat. Do not add these counts to the per-title scores.

Family-level criterion	Default	Previous override	Topic-first override
Consistent structure	10 pass · 1 partial · 1 fail / 12	9 pass · 3 partial / 12	4 pass · 4 partial · 4 fail / 12
Meaningful variation	14 pass · 10 partial / 24	14 pass · 8 partial · 2 fail / 24	12 pass · 12 partial / 24

Every family judgment

Haiku 4.5 · Default · run 1 · auth

Consistency: partial · variation: partial

Haiku 4.5 · Default · run 1 · upgrade

Consistency: pass · variation: pass

Parallel Upgrade and Run Regression Tests titles distinguish urllib3, Pydantic, and Pillow.

Haiku 4.5 · Default · run 1 · feedback

Consistency: not_applicable · variation: partial

Subjects remain distinct, but chess sentiment is truncated and notification feedback shifts to analysis.

Haiku 4.5 · Default · run 1 · operations

Consistency: not_applicable · variation: pass

Backup permission investigation and safe restore testing reflect the two different requests.

Haiku 4.5 · Default · run 2 · auth

Consistency: fail · variation: partial

Troubleshoot, Debug, and Error Investigation use different structures for equivalent investigations; subjects remain clear.

Haiku 4.5 · Default · run 2 · upgrade

Consistency: pass · variation: pass

All three use Upgrade and Run Regression Tests with distinct package names.

Haiku 4.5 · Default · run 2 · feedback

Consistency: not_applicable · variation: partial

Feedback subjects remain distinct, but notification sentiment is omitted and chess criticism is truncated.

Haiku 4.5 · Default · run 2 · operations

Consistency: not_applicable · variation: pass

Investigating the backup permission error and testing a restore safely are justified action differences.

Haiku 4.5 · Previous override · run 1 · auth

Consistency: pass · variation: partial

All use Investigating with distinct auth subjects; session timing loses the specific five-minute symptom.

Haiku 4.5 · Previous override · run 1 · upgrade

Consistency: pass · variation: pass

Parallel Upgrading package patch titles distinguish all three packages; latest patch wording is compatible.

Haiku 4.5 · Previous override · run 1 · feedback

Consistency: not_applicable · variation: fail

Energy complaints become an investigation, and offline praise becomes Praising rather than summarizing customer feedback.

Haiku 4.5 · Previous override · run 1 · operations

Consistency: not_applicable · variation: pass

Permission-error investigation and safe restore testing preserve distinct operational tasks.

Haiku 4.5 · Previous override · run 2 · auth

Consistency: pass · variation: pass

Parallel Investigating titles retain post-reset login, five-minute sessions, and OAuth 400 distinctions.

Haiku 4.5 · Previous override · run 2 · upgrade

Consistency: pass · variation: pass

All use Upgrading with package-specific patch wording; no unsupported differences are introduced.

Haiku 4.5 · Previous override · run 2 · feedback

Consistency: not_applicable · variation: fail

Refund complaints become an investigation; Praising screen-reader labels loses customer attribution and changes the task.

Haiku 4.5 · Previous override · run 2 · operations

Consistency: not_applicable · variation: pass

Nightly permission-error investigation and safe restore testing are meaningfully distinct.

Haiku 4.5 · Topic-first override · run 1 · auth

Consistency: fail · variation: partial

Haiku 4.5 · Topic-first override · run 1 · upgrade

Consistency: pass · variation: pass

Each package uses the same upgrade and test pattern; ampersand differences do not alter structure.

Haiku 4.5 · Topic-first override · run 1 · feedback

Consistency: not_applicable · variation: partial

Subjects are distinct, but screen-reader praise becomes generic feedback and the chess complaint is truncated.

Haiku 4.5 · Topic-first override · run 1 · operations

Consistency: not_applicable · variation: pass

Failed-backup investigation and disposable-directory restore testing preserve the requested operational distinction.

Haiku 4.5 · Topic-first override · run 2 · auth

Consistency: fail · variation: partial

Investigation noun, investigate clause, and bare OAuth symptom use inconsistent structures while retaining distinct subjects.

Haiku 4.5 · Topic-first override · run 2 · upgrade

Consistency: pass · variation: pass

urllib3, Pydantic, and Pillow share the upgrade and test pattern without arbitrary differences.

Haiku 4.5 · Topic-first override · run 2 · feedback

Consistency: not_applicable · variation: partial

Subjects and praise remain identifiable, but truncation obscures the negative sides of chess and notification feedback.

Haiku 4.5 · Topic-first override · run 2 · operations

Consistency: not_applicable · variation: pass

Backup permission investigation and safe restore testing retain different subjects and justified actions.

Gemini 3 Flash Preview · Default · run 1 · auth

Consistency: pass · variation: partial

All use Debugging with distinct auth subjects; session wording drops the five-minute detail retained by other symptoms.

Gemini 3 Flash Preview · Default · run 1 · upgrade

Consistency: pass · variation: pass

All use Upgrade and Run Tests; Tests and Regression Tests are equivalent abbreviations here.

Gemini 3 Flash Preview · Default · run 1 · feedback

Consistency: not_applicable · variation: partial

All six subjects remain distinct, but screen-reader praise and mixed chess and notification sentiment become generic feedback.

Gemini 3 Flash Preview · Default · run 1 · operations

Consistency: not_applicable · variation: pass

Debugging the failed backup and testing a restore in a disposable directory follow the different inputs.

Gemini 3 Flash Preview · Default · run 2 · auth

Consistency: pass · variation: partial

Parallel Debugging titles distinguish the auth subjects, but session specificity drops the five-minute symptom.

Gemini 3 Flash Preview · Default · run 2 · upgrade

Consistency: pass · variation: pass

Parallel Upgrade and Run Tests titles preserve package distinctions; Regression Tests is an equivalent expansion.

Gemini 3 Flash Preview · Default · run 2 · feedback

Consistency: not_applicable · variation: partial

Parallel summary tasks distinguish subjects, but chess and notification titles omit their mixed sentiment.

Gemini 3 Flash Preview · Default · run 2 · operations

Consistency: not_applicable · variation: pass

Backup-error debugging and disposable-directory restore testing are distinct actions supported by the requests.

Gemini 3 Flash Preview · Previous override · run 1 · auth

Consistency: partial · variation: partial

OAuth switches from Investigating to Debugging; session timing is less specific, though all auth subjects remain clear.

Gemini 3 Flash Preview · Previous override · run 1 · upgrade

Consistency: partial · variation: partial

urllib3 uses Updating while Pydantic and Pillow use Upgrading; all retain testing and distinct package names.

Gemini 3 Flash Preview · Previous override · run 1 · feedback

Consistency: not_applicable · variation: partial

Summary subjects and most sentiments are clear, but chess loses its explicitly mixed feedback distinction.

Gemini 3 Flash Preview · Previous override · run 1 · operations

Consistency: not_applicable · variation: pass

Permission-error investigation and restore testing in a temporary directory reflect distinct requested work.

Gemini 3 Flash Preview · Previous override · run 2 · auth

Consistency: pass · variation: partial

All use Investigating and distinguish auth subjects; session expiration timing omits the five-minute detail.

Gemini 3 Flash Preview · Previous override · run 2 · upgrade

Consistency: partial · variation: partial

Updating urllib3 differs unnecessarily from Upgrading the other packages; testing and package distinctions remain visible.

Gemini 3 Flash Preview · Previous override · run 2 · feedback

Consistency: not_applicable · variation: partial

Subjects and complaint/praise distinctions remain clear, but chess feedback omits its mixed sentiment.

Gemini 3 Flash Preview · Previous override · run 2 · operations

Consistency: not_applicable · variation: partial

Investigation and explanation are justified, but explaining restoration omits the requested restore-test focus.

Gemini 3 Flash Preview · Topic-first override · run 1 · auth

Consistency: fail · variation: partial

Gemini 3 Flash Preview · Topic-first override · run 1 · upgrade

Consistency: partial · variation: partial

Pydantic and Pillow add Patch to upgrade and test while urllib3 does not, despite equivalent requests.

Gemini 3 Flash Preview · Topic-first override · run 1 · feedback

Consistency: not_applicable · variation: partial

Subjects remain distinct and chess retains mixed feedback, but notification sentiment is generalized and energy drops summary framing.

Gemini 3 Flash Preview · Topic-first override · run 1 · operations

Consistency: not_applicable · variation: pass

Backup failure investigation and restore verification in a disposable directory preserve the different requested tasks.

Gemini 3 Flash Preview · Topic-first override · run 2 · auth

Consistency: fail · variation: partial

Gemini 3 Flash Preview · Topic-first override · run 2 · upgrade

Consistency: partial · variation: partial

Only Pydantic and Pillow explicitly say patch; the equivalent urllib3 task uses a shorter upgrade-and-test structure.

Gemini 3 Flash Preview · Topic-first override · run 2 · feedback

Consistency: not_applicable · variation: partial

Subjects remain distinct, but notification preferences loses explicit mixed sentiment and energy complaints lose summary framing.

Gemini 3 Flash Preview · Topic-first override · run 2 · operations

Consistency: not_applicable · variation: pass

Backup failure investigation and disposable-directory restore testing remain distinct and supported by the inputs.

GPT-6 Luna · low · Default · run 1 · auth

Consistency: pass · variation: pass

Parallel Investigate titles preserve password-reset login, five-minute session expiration, and OAuth HTTP 400.

GPT-6 Luna · low · Default · run 1 · upgrade

Consistency: pass · variation: pass

All three package titles use Upgrade and run regression tests with meaningful package-name differences.

GPT-6 Luna · low · Default · run 1 · feedback

Consistency: not_applicable · variation: partial

Subjects remain distinct, but offline praise loses attribution and truncation obscures chess and notification criticism.

GPT-6 Luna · low · Default · run 1 · operations

Consistency: not_applicable · variation: pass

Nightly backup-error investigation and disposable-directory restore testing reflect the distinct requests.

GPT-6 Luna · low · Default · run 2 · auth

Consistency: pass · variation: pass

All use Investigate and retain the distinct login-reset, five-minute session, and OAuth HTTP 400 symptoms.

GPT-6 Luna · low · Default · run 2 · upgrade

Consistency: pass · variation: pass

Parallel Upgrade and run regression tests titles distinguish each package without arbitrary variation.

GPT-6 Luna · low · Default · run 2 · feedback

Consistency: not_applicable · variation: partial

Subjects remain clear, but offline praise loses attribution and notification scheduling/sound feedback omits mixed sentiment.

GPT-6 Luna · low · Default · run 2 · operations

Consistency: not_applicable · variation: pass

Investigating a backup permission error and documenting safe restore testing follow the investigation and explanation requests.

GPT-6 Luna · low · Previous override · run 1 · auth

Consistency: pass · variation: pass

Parallel Investigating titles distinguish login after password reset, five-minute sessions, and OAuth HTTP 400.

GPT-6 Luna · low · Previous override · run 1 · upgrade

Consistency: pass · variation: pass

All three use Upgrading and running regression tests with distinct package names.

GPT-6 Luna · low · Previous override · run 1 · feedback

Consistency: not_applicable · variation: pass

Parallel Summarizing titles distinguish six subjects and correctly preserve complaints, customer praise, and mixed feedback.

GPT-6 Luna · low · Previous override · run 1 · operations

Consistency: not_applicable · variation: pass

Nightly permission-error investigation and disposable-directory restore testing preserve the two different tasks.

GPT-6 Luna · low · Previous override · run 2 · auth

Consistency: pass · variation: pass

Parallel Investigating titles preserve meaningful login, session, and OAuth symptom differences.

GPT-6 Luna · low · Previous override · run 2 · upgrade

Consistency: pass · variation: pass

Upgrading and running regression tests stays parallel across urllib3, Pydantic, and Pillow.

GPT-6 Luna · low · Previous override · run 2 · feedback

Consistency: not_applicable · variation: pass

All six summary titles distinguish their subjects and preserve complaint, praise, and mixed sentiment.

GPT-6 Luna · low · Previous override · run 2 · operations

Consistency: not_applicable · variation: pass

Backup-error investigation and safe restore testing retain the justified operational differences despite truncated safety detail.

GPT-6 Luna · low · Topic-first override · run 1 · auth

Consistency: pass · variation: pass

Each uses subject plus Investigate and preserves reset login, five-minute sessions, and OAuth HTTP 400.

GPT-6 Luna · low · Topic-first override · run 1 · upgrade

Consistency: pass · variation: pass

All use package Upgrade and Regression Tests; punctuation and emoji do not change the parallel structure.

GPT-6 Luna · low · Topic-first override · run 1 · feedback

Consistency: not_applicable · variation: pass

All six subjects remain identifiable, with complaint, praise, and mixed feedback accurately differentiated.

GPT-6 Luna · low · Topic-first override · run 1 · operations

Consistency: not_applicable · variation: pass

Nightly backup permission-failure investigation contrasts appropriately with safe restore testing.

GPT-6 Luna · low · Topic-first override · run 2 · auth

Consistency: partial · variation: partial

GPT-6 Luna · low · Topic-first override · run 2 · upgrade

Consistency: partial · variation: partial

Test Regressions differs from Regression Tests and only Pydantic specifies Patch; package distinctions remain clear.

GPT-6 Luna · low · Topic-first override · run 2 · feedback

Consistency: not_applicable · variation: pass

Distinct subjects retain complaints, customer praise, and mixed chess/notification feedback without invented work.

GPT-6 Luna · low · Topic-first override · run 2 · operations

Consistency: not_applicable · variation: pass

Backup permissions investigation and safe restore testing preserve the requested operational distinction.

All 144 matched comparisons.

Each row contains three actual displayed SDK outputs for the same synthetic request. Exact duplicate request/title pairs share one blind judgment, even when different models produced them.

Model Cases 144 comparisons

Haiku 4.5 · run 1 · auth-login · reused

Default

🔒️ Login Failure After Password Reset

frontload: pass fidelity: pass

37 characters · 1.43s

Previous override

🔍 Investigating login failures after reset

frontload: pass fidelity: pass

42 characters · 0.88s

Topic-first override

🔍 Login failure after password reset

frontload: pass fidelity: pass

36 characters · 0.79s

Original synthetic request

Investigate why login fails after a password reset. No cause is confirmed yet.

Haiku 4.5 · run 1 · auth-session · reused

Default

🔒️ Session Expiration Investigation

frontload: pass fidelity: pass

Session expiration leads; clear investigative noun phrase.

35 characters · 0.59s

Previous override

🔍 Investigating session expiry timing

frontload: pass fidelity: pass

Session appears second; expiry timing remains under investigation.

37 characters · 0.78s

Topic-first override

🔍 Session expiry investigate timeouts

frontload: pass fidelity: pass

Session expiry leads, but the unpunctuated transition to investigate is awkward.

37 characters · 0.79s

Original synthetic request

Investigate why sessions expire after five minutes. No cause is confirmed yet.

Haiku 4.5 · run 1 · auth-oauth · reused

Default

🐛 OAuth 400 Error Investigation

frontload: pass fidelity: pass

OAuth 400 leads a clear investigation without asserting a cause.

31 characters · 0.87s

Previous override

🔍 Investigating OAuth callback 400 errors

frontload: pass fidelity: pass

OAuth appears second; investigation retains uncertainty about the 400 errors.

41 characters · 0.85s

Topic-first override

🔍 OAuth callbacks debug 400 errors

frontload: pass fidelity: pass

OAuth callbacks leads, but callbacks debug creates an ambiguous actor/action construction.

34 characters · 0.87s

Original synthetic request

Investigate why OAuth callbacks return HTTP 400. No cause is confirmed yet.

Haiku 4.5 · run 1 · upgrade-urllib3 · reused

Default

📦️ Upgrade urllib3 and Run Regression Tests

frontload: pass fidelity: pass

urllib3 appears second; requested upgrade and regression tests are retained.

43 characters · 0.92s

Previous override

🔄 Upgrading urllib3 to latest patch

frontload: pass fidelity: pass

urllib3 appears second; latest patch upgrade matches the requested work.

35 characters · 1.05s

Topic-first override

🔄 urllib3 upgrade and test

frontload: pass fidelity: pass

urllib3 leads; upgrade and testing are clearly represented.

26 characters · 1.04s

Original synthetic request

Upgrade urllib3 to the latest compatible patch release and run its regression tests.

Haiku 4.5 · run 1 · upgrade-pydantic · reused

Default

📦️ Upgrade Pydantic and Run Regression Tests

frontload: pass fidelity: pass

Pydantic appears second; upgrade and regression tests are explicit.

44 characters · 0.72s

Previous override

🔄 Upgrading Pydantic to latest patch

frontload: pass fidelity: pass

Pydantic appears second; patch upgrade matches the requested work.

36 characters · 1.05s

Topic-first override

🔧 Pydantic upgrade & test

frontload: pass fidelity: pass

Pydantic leads; concise upgrade and test task is understandable.

25 characters · 1.04s

Original synthetic request

Upgrade Pydantic to the latest compatible patch release and run its regression tests.

Haiku 4.5 · run 1 · upgrade-pillow · reused

Default

📦️ Upgrade Pillow & Run Regression Tests

frontload: pass fidelity: pass

Pillow appears second; upgrade and regression tests are explicit.

40 characters · 0.76s

Previous override

🔄 Upgrading Pillow patch

frontload: pass fidelity: pass

Pillow appears second; patch upgrading matches the task despite omitted regression tests.

24 characters · 0.72s

Topic-first override

🖼️ Pillow upgrade and test

frontload: pass fidelity: pass

Pillow leads; upgrade and testing preserve the requested work.

26 characters · 0.75s

Original synthetic request

Upgrade Pillow to the latest compatible patch release and run its regression tests.

Haiku 4.5 · run 1 · filler-cache · reused

Default

🗃️ Redis Cache Eviction Policies

frontload: pass fidelity: pass

Redis cache leads a clear descriptive topic; broader policy wording is compatible.

32 characters · 0.79s

Previous override

📚 Explaining Redis allkeys-lru

frontload: pass fidelity: pass

Redis appears second; explaining allkeys-lru retains the specific requested topic.

30 characters · 0.77s

Topic-first override

🗄️ Redis eviction explain allkeys-lru

frontload: pass fidelity: pass

Redis eviction leads, but eviction explain allkeys-lru is grammatically awkward.

37 characters · 0.82s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Redis cache eviction, especially the allkeys-lru policy.

Haiku 4.5 · run 1 · filler-build · reused

Default

⚡ Rust Incremental Compilation & Cache Misses

frontload: pass fidelity: pass

Rust incremental leads a clear descriptive compilation topic.

45 characters · 0.91s

Previous override

📚 Explaining Rust incremental compilation cache...

frontload: pass fidelity: pass

Rust appears second; explanation of incremental compilation remains clear despite clipping.

50 characters · 0.85s

Topic-first override

🦀 Rust incremental compilation cache misses

frontload: pass fidelity: pass

Rust incremental leads; the descriptive phrase captures compilation and cache misses.

43 characters · 0.77s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Rust incremental compilation and why cache misses slow builds.

Haiku 4.5 · run 1 · feedback-complaint · reused

Default

🐛 Lesson Energy Limit Complaints

frontload: pass fidelity: pass

Lesson energy leads; complaints preserve the negative customer feedback.

32 characters · 0.85s

Previous override

🔍 Investigating lesson energy limit interruptions

frontload: partial fidelity: fail

Energy limit appears after Investigating lesson; investigation replaces the requested complaint summary.

49 characters · 0.73s

Topic-first override

📚 Lesson energy limit interrupts practice

frontload: pass fidelity: pass

Lesson energy leads; interruptions preserve the complaint's negative substance.

41 characters · 0.72s

Original synthetic request

Summarize customer complaints: users say the lesson energy limit interrupts practice and forces them to wait. Do not turn this into praise.

Haiku 4.5 · run 1 · feedback-praise · reused

Default

✨ Offline Lessons Customer Praise Summary

frontload: pass fidelity: pass

Offline lessons lead; customer praise is explicitly being summarized.

41 characters · 0.70s

Previous override

🚂 Praising offline lessons for train practice

frontload: pass fidelity: fail

Offline appears second and praise stays positive, but Praising changes reported praise into agent endorsement.

45 characters · 0.72s

Topic-first override

📱 Offline lessons praised for train practice

frontload: pass fidelity: pass

Offline lessons lead; praised reports positive reception and its train-practice benefit.

44 characters · 0.68s

Original synthetic request

Summarize customer praise: users love the offline lessons because they can practice on the train. Do not invent a complaint.

Haiku 4.5 · run 1 · feedback-mixed · reused

Default

🐛 Chess Lesson Feedback: Puzzles Good, Checkmat...

frontload: pass fidelity: pass

Chess lesson leads; puzzle praise is visible, but checkmate criticism is truncated away.

50 characters · 0.89s

Previous override

♟️ Reviewing chess lessons: puzzles praised, ch...

frontload: pass fidelity: pass

Chess appears second; truncation leaves the second feedback clause incomplete and hides the negative side.

50 characters · 0.89s

Topic-first override

♟️ Chess lessons: puzzles good, checkmate rule ...

frontload: pass fidelity: pass

Chess lessons lead; puzzle praise is visible, but the checkmate complaint is not.

50 characters · 0.79s

Original synthetic request

Summarize mixed feedback about chess lessons: users enjoy the puzzles but complain that the checkmate rule sometimes rejects valid moves.

Haiku 4.5 · run 1 · neutral-compare · reused

Default

🗃️ SQLite vs PostgreSQL for Offline-First App

frontload: pass fidelity: pass

SQLite leads; versus expresses a neutral comparison for the offline-first app.

45 characters · 0.76s

Previous override

🔍 Comparing SQLite vs PostgreSQL

frontload: pass fidelity: pass

SQLite appears second; comparison expresses no database preference.

32 characters · 0.63s

Topic-first override

📊 Database comparison: SQLite vs PostgreSQL

frontload: partial fidelity: pass

Specific databases follow Database comparison; wording remains neutral and makes no selection.

43 characters · 0.74s

Original synthetic request

Compare SQLite and PostgreSQL for a small offline-first app. I have not selected either database.

Haiku 4.5 · run 1 · negation · reused

Default

👷 Docker Disk Usage Inspection Guide

frontload: pass fidelity: pass

Docker disk leads; inspection guide accurately captures the explanatory request.

36 characters · 0.97s

Previous override

🐳 Inspecting Docker disk usage

frontload: pass fidelity: pass

Docker appears second; disk inspection does not imply deletion.

30 characters · 0.73s

Topic-first override

🐳 Docker disk usage inspect

frontload: pass fidelity: pass

Docker disk leads, but trailing inspect makes the grammar awkward.

27 characters · 0.72s

Original synthetic request

Explain how to inspect Docker disk usage. Do not delete containers, images, or volumes.

Haiku 4.5 · run 1 · uncertainty · reused

Default

🔒 TLS Certificate Rotation & Webhook Failures

frontload: pass fidelity: pass

TLS certificate leads; juxtaposing rotation and failures does not assert causation.

45 characters · 0.79s

Previous override

🔍 Checking TLS rotation for webhook failures

frontload: pass fidelity: pass

TLS appears second; checking rotation preserves the causal hypothesis.

44 characters · 0.74s

Topic-first override

🔍 TLS certificate rotation - investigate webhoo...

frontload: pass fidelity: pass

TLS certificate leads; investigation remains visible and does not assert the hypothesis as fact.

50 characters · 0.69s

Original synthetic request

Check whether TLS certificate rotation might explain the intermittent webhook failures. This is a hypothesis, not a confirmed cause.

Haiku 4.5 · run 1 · completed · reused

Default

🐛 Write Release Note for Login Fix

frontload: partial fidelity: pass

Clear release-note task, but the distinguishing login topic appears later.

34 characters · 0.72s

Previous override

📝 Writing release note for login fix

frontload: partial fidelity: pass

Release-note writing is clear, but login appears after the first two words.

36 characters · 0.85s

Topic-first override

✅ Login fix release note

frontload: pass fidelity: pass

24 characters · 0.72s

Original synthetic request

The login fix has already merged and all regression tests passed. Write a short release note about the completed fix.

Haiku 4.5 · run 1 · long-tail · reused

Default

🐛 Investigate Zephyr's Unicode CSV Parser Crash

frontload: pass fidelity: pass

Zephyr appears second; investigation follows the actual parser-crash request.

47 characters · 0.65s

Previous override

🔍 Investigating Zephyr Unicode CSV parser crash

frontload: pass fidelity: pass

Zephyr appears second; investigation targets the Unicode CSV parser crash.

47 characters · 0.93s

Topic-first override

🔍 Zephyr Unicode CSV crash

frontload: pass fidelity: pass

Clear crash topic led by Zephyr Unicode; retains the actual long-input request.

26 characters · 0.75s

Original synthetic request

Reference background only; the actual request follows at the end.
The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder. The project uses a small command-line interface and stores its examples in a dedicated folder.
ACTUAL REQUEST: Investigate Zephyr's Unicode CSV parser crash. Do not work on the CLI examples.

Haiku 4.5 · run 1 · backup-status · fresh

Default

👷 Investigate Nightly Backup Permission Error

frontload: partial fidelity: pass

Investigate Nightly precedes Backup; the permission error remains investigative.

45 characters · 0.78s

Previous override

🔍 Investigating backup permission error

frontload: pass fidelity: pass

Backup appears second; investigating the permission error preserves the requested work.

39 characters · 0.72s

Topic-first override

🔍 Backup failed permission error investigate

frontload: pass fidelity: pass

Backup failed leads; trailing investigate makes the clause awkward, while failure status remains accurate.

44 characters · 0.76s

Original synthetic request

The nightly backup failed with a permission error. Investigate it; do not claim that a backup completed.

Haiku 4.5 · run 1 · backup-restore · fresh

Default

✅ Test Backup Restore Safely

frontload: partial fidelity: pass

Test Backup delays restore beyond two words; safe procedure does not assert completion.

28 characters · 0.80s

Previous override

🔄 Testing backup restore safely

frontload: partial fidelity: pass

Testing backup delays restore beyond two words; safe testing remains compatible.

31 characters · 0.69s

Topic-first override

🔄 Backup restore testing in disposable dir

frontload: pass fidelity: pass

Backup restore leads; disposable-directory testing matches the request.

42 characters · 0.79s

Original synthetic request

Explain how to test a backup restore in a disposable directory. Do not overwrite the production database.

Haiku 4.5 · run 1 · feedback-refunds · fresh

Default

👔 Customer Refund Processing Delays Report

frontload: pass fidelity: pass

Refund appears second; processing delays omit explicit customer complaints or frustration.

42 characters · 0.74s

Previous override

📋 Reporting slow refund processing complaints

frontload: partial fidelity: pass

Reporting slow delays refund processing; complaints retain negative sentiment.

45 characters · 1.04s

Topic-first override

🔄 Refund delays customer frustration

frontload: pass fidelity: pass

Refund delays lead; customer frustration preserves negative attributed sentiment.

36 characters · 0.79s

Original synthetic request

Summarize customer complaints that refund processing takes weeks. Report their frustration without endorsing or dismissing it.

Haiku 4.5 · run 1 · feedback-accessibility · fresh

Default

💄 Screen-Reader Labels Customer Praise Summary

frontload: pass fidelity: pass

Screen-reader labels lead; praise is explicitly attributed to customers.

46 characters · 0.71s

Previous override

🎯 Customers praising screen-reader labels

frontload: partial fidelity: pass

Screen-reader appears later; customers are correctly identified as the source of praise.

41 characters · 0.64s

Topic-first override

📋 Customer feedback on screen-reader labels

frontload: partial fidelity: pass

Screen-reader topic appears later; generic customer feedback loses the explicit praise.

43 characters · 0.74s

Original synthetic request

Summarize customer praise for the new screen-reader labels. The praise is from customers, not a request for your endorsement.

Haiku 4.5 · run 1 · feedback-notifications · fresh

Default

✨ Feedback Analysis: Scheduling Love, Sound Issues

frontload: partial fidelity: pass

Generic Feedback Analysis leads; later scheduling love and sound issues preserve both sentiments.

50 characters · 0.83s

Previous override

📢 Reviewing notification feedback

frontload: pass fidelity: pass

Notification appears second; generic feedback omits the mixed polarity.

33 characters · 0.81s

Topic-first override

📊 Customer feedback: notifications mixed

frontload: partial fidelity: pass

Notifications appears after Customer feedback; explicitly mixed preserves the overall sentiment.

40 characters · 0.81s

Original synthetic request

Summarize mixed customer feedback: users like notification scheduling but dislike the default notification sound.

Haiku 4.5 · run 1 · proposal-not-approved · fresh

Default

👷 Jenkins to Buildkite Migration Proposal

frontload: pass fidelity: pass

Jenkins and Buildkite lead the meaningful words; migration remains a proposal.

41 characters · 0.78s

Previous override

📋 Drafting Jenkins-to-Buildkite proposal

frontload: pass fidelity: pass

Jenkins-to-Buildkite appears second; drafting a proposal implies no started migration.

40 characters · 0.81s

Topic-first override

🔄 Jenkins replacement proposal draft

frontload: pass fidelity: pass

Jenkins replacement leads; proposal draft does not imply migration approval.

36 characters · 0.73s

Original synthetic request

Draft a proposal to replace Jenkins with Buildkite. No migration has been approved or started.

Haiku 4.5 · run 1 · suspected-leak · fresh

Default

⚡️ Investigate Image Buffer Memory Growth

frontload: pass fidelity: pass

Image appears second; investigation preserves uncertainty about buffer-related memory growth.

41 characters · 0.75s

Previous override

🔍 Investigating image buffer retention

frontload: pass fidelity: pass

Image appears second; investigating retention does not confirm a leak.

38 characters · 0.75s

Topic-first override

🔍 Image buffers memory growth—investigate

frontload: pass fidelity: pass

Image buffers leads, but the noun sequence before investigate is awkward; no confirmed leak is claimed.

41 characters · 0.77s

Original synthetic request

Check whether retained image buffers might cause the memory growth. This is only a hypothesis; do not report a confirmed leak.

Haiku 4.5 · run 1 · docs-after-revert · fresh

Default

🚀 Update Incident Timeline After Deployment Revert

frontload: pass fidelity: pass

Incident appears second; updating the timeline after revert preserves the documentation task.

50 characters · 0.78s

Previous override

🔄 Updating incident timeline after revert

frontload: pass fidelity: pass

Incident appears second; updating its timeline after revert preserves the documentation task.

41 characters · 0.83s

Topic-first override

🔄 Deployment Incident Update Timeline

frontload: pass fidelity: pass

Deployment incident leads; timeline update remains the requested work.

37 characters · 0.62s

Original synthetic request

The deployment was reverted yesterday. Update the incident timeline; do not redeploy or reopen the rollout.

Haiku 4.5 · run 2 · auth-login · reused

Default

🐛 Troubleshoot Post-Password Reset Login Failure

frontload: partial fidelity: pass

48 characters · 0.74s

Previous override

🔍 Investigating login failures post-reset

frontload: pass fidelity: pass

41 characters · 0.76s

Topic-first override

🔍 Login reset failure investigation

frontload: pass fidelity: pass

35 characters · 0.75s

Original synthetic request

Investigate why login fails after a password reset. No cause is confirmed yet.

Haiku 4.5 · run 2 · auth-session · reused

Default

🔧 Debug 5-Minute Session Expiration Issue

frontload: partial fidelity: pass

Debug and 5-Minute precede Session; debugging remains investigative.

41 characters · 0.77s

Previous override

🔍 Investigating session expiry at 5 min

frontload: pass fidelity: pass

Session appears second; investigation preserves uncertainty.

39 characters · 0.80s

Topic-first override

🔍 Sessions expire - investigate cause

frontload: pass fidelity: pass

Sessions expire leads; investigating the cause preserves uncertainty.

37 characters · 0.80s

Original synthetic request

Investigate why sessions expire after five minutes. No cause is confirmed yet.

Haiku 4.5 · run 2 · auth-oauth · reused

Default

🐛 OAuth 400 Callback Error Investigation

frontload: pass fidelity: pass

OAuth leads; investigation does not assert a cause or repair.

40 characters · 0.82s

Previous override

🔍 Investigating OAuth callbacks returning 400

frontload: pass fidelity: pass

OAuth appears second; investigating callbacks returning 400 matches the request.

45 characters · 0.87s

Topic-first override

🔍 OAuth callbacks returning 400 errors

frontload: pass fidelity: pass

OAuth callbacks lead; the reported 400 errors are described without a causal claim.

38 characters · 0.66s

Original synthetic request

Investigate why OAuth callbacks return HTTP 400. No cause is confirmed yet.

Haiku 4.5 · run 2 · upgrade-urllib3 · reused

Default

📦️ Upgrade urllib3 & Run Regression Tests

frontload: pass fidelity: pass

urllib3 appears second; upgrade and regression testing are explicit.

41 characters · 0.77s

Previous override

🔄 Upgrading urllib3 to latest patch

frontload: pass fidelity: pass

urllib3 appears second; latest patch upgrade matches the requested work.

35 characters · 0.77s

Topic-first override

🔧 urllib3 upgrade and test

frontload: pass fidelity: pass

urllib3 leads a clear upgrade and testing task.

26 characters · 0.87s

Original synthetic request

Upgrade urllib3 to the latest compatible patch release and run its regression tests.

Haiku 4.5 · run 2 · upgrade-pydantic · reused

Default

📦 Upgrade Pydantic & Run Regression Tests

frontload: pass fidelity: pass

Pydantic appears second; upgrade and regression tests are explicit.

41 characters · 0.75s

Previous override

🔧 Upgrading Pydantic patch

frontload: pass fidelity: pass

Pydantic appears second; patch upgrading matches the task despite omitted tests.

26 characters · 0.73s

Topic-first override

🔧 Pydantic upgrade & test

frontload: pass fidelity: pass

Pydantic leads; concise upgrade and test task is understandable.

25 characters · 0.82s

Original synthetic request

Upgrade Pydantic to the latest compatible patch release and run its regression tests.

Haiku 4.5 · run 2 · upgrade-pillow · reused

Default

📦️ Upgrade Pillow & Run Regression Tests

frontload: pass fidelity: pass

Pillow appears second; upgrade and regression tests are explicit.

40 characters · 1.05s

Previous override

🔄 Upgrading Pillow patch version

frontload: pass fidelity: pass

Pillow appears second; patch upgrade is accurate despite omitted tests.

32 characters · 1.48s

Topic-first override

🖼️ Pillow upgrade and test

frontload: pass fidelity: pass

Pillow leads; upgrade and testing preserve the requested work.

26 characters · 0.82s

Original synthetic request

Upgrade Pillow to the latest compatible patch release and run its regression tests.

Haiku 4.5 · run 2 · filler-cache · reused

Default

⚡️ Redis Cache Eviction & allkeys-lru Policy

frontload: pass fidelity: pass

Redis cache leads; eviction and allkeys-lru identify the requested explanation topic.

44 characters · 1.05s

Previous override

📚 Explaining Redis cache eviction allkeys-lru

frontload: pass fidelity: pass

Redis appears second; cache eviction and allkeys-lru retain the explanation topic.

45 characters · 0.81s

Topic-first override

🔄 Redis cache eviction explain

frontload: pass fidelity: pass

Redis cache leads, but trailing explain lacks a clear grammatical connection.

30 characters · 0.98s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Redis cache eviction, especially the allkeys-lru policy.

Haiku 4.5 · run 2 · filler-build · reused

Default

👷 Rust Incremental Compilation & Cache Misses

frontload: pass fidelity: pass

Rust incremental leads; descriptive phrase retains compilation and cache misses.

45 characters · 0.82s

Previous override

📚 Explaining Rust incremental compilation cache...

frontload: pass fidelity: pass

Rust appears second; explanation of incremental compilation remains clear despite clipping.

50 characters · 0.84s

Topic-first override

🦀 Rust incremental compilation cache misses

frontload: pass fidelity: pass

Rust incremental leads; the descriptive phrase captures compilation and cache misses.

43 characters · 0.85s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Rust incremental compilation and why cache misses slow builds.

Haiku 4.5 · run 2 · feedback-complaint · reused

Default

🐛 Customer Complaints: Energy Limit Issues

frontload: partial fidelity: pass

Energy limit appears after Customer Complaints; negative reported feedback remains clear.

42 characters · 0.72s

Previous override

📋 Summarizing lesson energy limit complaints

frontload: partial fidelity: pass

Summarizing lesson delays energy limit; complaints remain explicitly negative.

44 characters · 0.81s

Topic-first override

🎓 Lesson limits interrupt practice, users report

frontload: pass fidelity: pass

Lesson limits lead; users report attributes the negative practice interruptions.

48 characters · 0.95s

Original synthetic request

Summarize customer complaints: users say the lesson energy limit interrupts practice and forces them to wait. Do not turn this into praise.

Haiku 4.5 · run 2 · feedback-praise · reused

Default

✨ Summarize Offline Lessons Customer Praise

frontload: pass fidelity: pass

Offline appears second; summarizing customer praise preserves source and sentiment.

43 characters · 0.85s

Previous override

🚂 Praising offline lessons for train practice

frontload: pass fidelity: fail

Offline appears second and praise stays positive, but Praising changes reported praise into agent endorsement.

45 characters · 0.72s

Topic-first override

📱 Offline lessons praised for train practice

frontload: pass fidelity: pass

Offline lessons lead; praised reports positive reception and its train-practice benefit.

44 characters · 0.72s

Original synthetic request

Summarize customer praise: users love the offline lessons because they can practice on the train. Do not invent a complaint.

Haiku 4.5 · run 2 · feedback-mixed · reused

Default

🐛 Chess Lesson Feedback: Puzzles Good, Checkmat...

frontload: pass fidelity: pass

Chess lesson leads; puzzle praise is visible, but checkmate criticism is truncated away.

50 characters · 0.87s

Previous override

♟️ Reviewing chess feedback

frontload: pass fidelity: pass

Chess appears second; reviewing feedback is compatible, but neither sentiment is expressed.

27 characters · 0.85s

Topic-first override

♟️ Chess Lessons: puzzles liked, checkmate rule...

frontload: pass fidelity: pass

Chess lessons lead; puzzles liked is visible, but the checkmate complaint is truncated away.

50 characters · 0.74s

Original synthetic request

Summarize mixed feedback about chess lessons: users enjoy the puzzles but complain that the checkmate rule sometimes rejects valid moves.

Haiku 4.5 · run 2 · neutral-compare · reused

Default

🗃️ SQLite vs PostgreSQL for Offline App

frontload: pass fidelity: pass

SQLite leads; versus expresses a neutral database comparison.

39 characters · 0.85s

Previous override

🔍 Comparing SQLite vs PostgreSQL

frontload: pass fidelity: pass

SQLite appears second; comparison expresses no database preference.

32 characters · 0.72s

Topic-first override

🗄️ Database Comparison: SQLite vs PostgreSQL

frontload: partial fidelity: pass

Database Comparison delays the named databases; comparison remains neutral.

44 characters · 0.81s

Original synthetic request

Compare SQLite and PostgreSQL for a small offline-first app. I have not selected either database.

Haiku 4.5 · run 2 · negation · reused

Default

👷 Docker Disk Usage Inspection Guide

frontload: pass fidelity: pass

Docker disk leads; inspection guide accurately captures the explanatory request.

36 characters · 0.72s

Previous override

🐳 Inspecting Docker disk usage

frontload: pass fidelity: pass

Docker appears second; disk inspection does not imply deletion.

30 characters · 0.79s

Topic-first override

🐳 Docker disk usage inspection

frontload: pass fidelity: pass

Docker disk leads a clear inspection topic without deletion.

30 characters · 0.65s

Original synthetic request

Explain how to inspect Docker disk usage. Do not delete containers, images, or volumes.

Haiku 4.5 · run 2 · uncertainty · reused

Default

🔒️ Investigate TLS Certificate Rotation & Webhooks

frontload: pass fidelity: pass

TLS appears second; investigate preserves the hypothetical relationship.

50 characters · 0.75s

Previous override

🔍 Investigating TLS rotation in webhooks

frontload: pass fidelity: pass

TLS appears second; investigating rotation does not assert a confirmed cause.

40 characters · 0.80s

Topic-first override

🔍 TLS certificate rotation investigation

frontload: pass fidelity: pass

TLS certificate leads; investigation preserves uncertainty despite omitting webhook detail.

40 characters · 0.77s

Original synthetic request

Check whether TLS certificate rotation might explain the intermittent webhook failures. This is a hypothesis, not a confirmed cause.

Haiku 4.5 · run 2 · completed · reused

Default

🐛 Write Release Note for Login Fix

frontload: partial fidelity: pass

Clear release-note task, but the distinguishing login topic appears later.

34 characters · 0.87s

Previous override

✅ Writing release note for login fix

frontload: partial fidelity: pass

Release-note task is clear, but login appears later than the first two words.

36 characters · 0.72s

Topic-first override

🔧 Login fix released with passing tests

frontload: pass fidelity: fail

39 characters · 0.77s

Original synthetic request

The login fix has already merged and all regression tests passed. Write a short release note about the completed fix.

Haiku 4.5 · run 2 · long-tail · reused

Default

🐛 Investigate Zephyr's Unicode CSV Parser Crash

frontload: pass fidelity: pass

Zephyr appears second; investigation follows the actual parser-crash request.

47 characters · 0.76s

Previous override

🔍 Investigating Zephyr Unicode CSV parser crash

frontload: pass fidelity: pass

Zephyr appears second; investigation targets the Unicode CSV parser crash.

47 characters · 0.62s

Topic-first override

🔍 Zephyr Unicode CSV crash investigate

frontload: pass fidelity: pass

Zephyr Unicode leads; trailing investigate is awkward but retains the crash investigation.

38 characters · 0.80s

Original synthetic request

Haiku 4.5 · run 2 · backup-status · fresh

Default

👷 Investigate Nightly Backup Permission Error

frontload: partial fidelity: pass

Investigate Nightly precedes Backup; the permission error remains investigative.

45 characters · 0.88s

Previous override

🔍 Investigating nightly backup permission error

frontload: partial fidelity: pass

Backup follows Investigating nightly; the permission-error investigation is accurate.

47 characters · 0.79s

Topic-first override

🔍 Backup permission error investigation

frontload: pass fidelity: pass

Backup permission leads; investigation retains the reported failure.

39 characters · 0.66s

Original synthetic request

The nightly backup failed with a permission error. Investigate it; do not claim that a backup completed.

Haiku 4.5 · run 2 · backup-restore · fresh

Default

✅ Testing Backup Restore Safely

frontload: partial fidelity: pass

Testing Backup delays restore beyond the first two words; no completion is asserted.

31 characters · 0.74s

Previous override

🧪 Testing backup restore safely

frontload: partial fidelity: pass

Testing backup delays the distinguishing restore topic until the third word.

31 characters · 0.72s

Topic-first override

🔄 Backup restore testing safely

frontload: pass fidelity: pass

Backup restore leads; the descriptive testing topic is compatible with explaining safe restore tests.

31 characters · 0.72s

Original synthetic request

Explain how to test a backup restore in a disposable directory. Do not overwrite the production database.

Haiku 4.5 · run 2 · feedback-refunds · fresh

Default

👔 Customer Refund Processing Delays

frontload: pass fidelity: pass

Refund appears second; customer processing delays preserve the complaint's negative subject.

35 characters · 0.71s

Previous override

🔄 Investigating slow refund processing

frontload: partial fidelity: fail

Refund appears third; investigating processing replaces summarizing customer complaints.

38 characters · 0.79s

Topic-first override

🔍 Customer refund delays reported

frontload: pass fidelity: pass

Refund appears second; reported delays omit explicit complaints or frustration.

33 characters · 0.70s

Original synthetic request

Summarize customer complaints that refund processing takes weeks. Report their frustration without endorsing or dismissing it.

Haiku 4.5 · run 2 · feedback-accessibility · fresh

Default

✨ Screen-Reader Labels Customer Praise

frontload: pass fidelity: pass

Screen-reader labels lead; customer praise explicitly preserves positive reported feedback.

38 characters · 0.70s

Previous override

🎤 Praising screen-reader labels

frontload: pass fidelity: fail

Screen-reader appears second and positivity remains, but Praising replaces summarizing customer praise.

31 characters · 0.72s

Topic-first override

📊 Customer praise for screen-reader labels

frontload: partial fidelity: pass

Screen-reader labels appears after Customer praise; attribution and positive sentiment remain clear.

42 characters · 0.73s

Original synthetic request

Summarize customer praise for the new screen-reader labels. The praise is from customers, not a request for your endorsement.

Haiku 4.5 · run 2 · feedback-notifications · fresh

Default

📝 Summarize Customer Feedback: Notifications

frontload: partial fidelity: pass

Notifications follows Summarize Customer Feedback; generic feedback omits the mixed sentiment.

44 characters · 0.84s

Previous override

🔔 Summarizing notification feedback

frontload: pass fidelity: pass

Notification appears second; generic feedback loses the mixed sentiment distinction.

35 characters · 0.79s

Topic-first override

🔔 Notification feedback: scheduling liked, soun...

frontload: pass fidelity: pass

Notification leads; scheduling liked is visible, but the sound complaint is truncated away.

50 characters · 0.80s

Original synthetic request

Summarize mixed customer feedback: users like notification scheduling but dislike the default notification sound.

Haiku 4.5 · run 2 · proposal-not-approved · fresh

Default

👷 Jenkins to Buildkite Migration Proposal

frontload: pass fidelity: pass

Jenkins and Buildkite lead the meaningful words; migration remains a proposal.

41 characters · 0.75s

Previous override

📋 Drafting Jenkins replacement proposal

frontload: pass fidelity: pass

Jenkins appears second; drafting a proposal preserves its unapproved status.

39 characters · 0.71s

Topic-first override

📋 Jenkins Replacement Proposal Draft

frontload: pass fidelity: pass

Jenkins leads; proposal draft does not imply an approved migration.

36 characters · 0.63s

Original synthetic request

Draft a proposal to replace Jenkins with Buildkite. No migration has been approved or started.

Haiku 4.5 · run 2 · suspected-leak · fresh

Default

⚡️ Investigate Image Buffer Memory Growth

frontload: pass fidelity: pass

Image appears second; investigation preserves uncertainty about buffer-related memory growth.

41 characters · 0.73s

Previous override

🔍 Investigating image buffer memory growth

frontload: pass fidelity: pass

Image appears second; investigation avoids confirming a leak.

42 characters · 0.69s

Topic-first override

🔍 Memory buffers: check leak hypothesis

frontload: pass fidelity: pass

Memory buffers leads; leak remains a hypothesis rather than a confirmed diagnosis.

39 characters · 0.69s

Original synthetic request

Check whether retained image buffers might cause the memory growth. This is only a hypothesis; do not report a confirmed leak.

Haiku 4.5 · run 2 · docs-after-revert · fresh

Default

🚀 Update Incident Timeline After Deployment Revert

frontload: pass fidelity: pass

Incident appears second; updating the timeline after revert preserves the documentation task.

50 characters · 0.75s

Previous override

🔄 Updating incident timeline after revert

frontload: pass fidelity: pass

Incident appears second; updating its timeline after revert preserves the documentation task.

41 characters · 0.85s

Topic-first override

🔄 Deployment Revert: Update Timeline

frontload: pass fidelity: pass

Deployment revert leads; updating the timeline does not reopen rollout.

36 characters · 0.85s

Original synthetic request

The deployment was reverted yesterday. Update the incident timeline; do not redeploy or reopen the rollout.

Gemini 3 Flash Preview · run 1 · auth-login · reused

Default

🐛 Debugging Login Failure After Password Reset

frontload: pass fidelity: pass

46 characters · 1.44s

Previous override

🔍 Investigating login failure after reset

frontload: pass fidelity: pass

41 characters · 0.99s

Topic-first override

🔑 Login Failure Investigation

frontload: pass fidelity: pass

29 characters · 0.97s

Original synthetic request

Investigate why login fails after a password reset. No cause is confirmed yet.

Gemini 3 Flash Preview · run 1 · auth-session · reused

Default

👔 Debugging Session Expiration Timeout

frontload: pass fidelity: pass

Session appears second; debugging expiration preserves investigation without claiming a repair.

38 characters · 2.55s

Previous override

🕵️ Investigating session expiration timing

frontload: pass fidelity: pass

Session appears second; investigating expiration timing preserves the task.

42 characters · 1.43s

Topic-first override

🕵️ Session Expiry: Investigate Timeout Issue

frontload: pass fidelity: pass

Session expiry leads; timeout remains under investigation.

44 characters · 1.41s

Original synthetic request

Investigate why sessions expire after five minutes. No cause is confirmed yet.

Gemini 3 Flash Preview · run 1 · auth-oauth · reused

Default

🐛 Debugging OAuth 400 Callback Errors

frontload: pass fidelity: pass

OAuth appears second; debugging preserves investigation of 400 errors.

37 characters · 0.95s

Previous override

🔍 Debugging OAuth callback 400 errors

frontload: pass fidelity: pass

OAuth appears second; debugging callback 400 errors preserves investigation.

37 characters · 2.31s

Topic-first override

🔑 OAuth Callbacks: Investigate 400 errors

frontload: pass fidelity: pass

OAuth callbacks lead; investigation of 400 errors matches the request.

41 characters · 1.14s

Original synthetic request

Investigate why OAuth callbacks return HTTP 400. No cause is confirmed yet.

Gemini 3 Flash Preview · run 1 · upgrade-urllib3 · reused

Default

📦️ Upgrade urllib3 and Run Tests

frontload: pass fidelity: pass

urllib3 appears second; upgrade and testing are both retained.

32 characters · 1.36s

Previous override

📦 Updating urllib3 and running tests

frontload: pass fidelity: pass

urllib3 appears second; updating and testing preserve the requested upgrade.

36 characters · 1.92s

Topic-first override

📦 urllib3 upgrade and test

frontload: pass fidelity: pass

urllib3 leads a clear upgrade-and-test description.

26 characters · 0.83s

Original synthetic request

Upgrade urllib3 to the latest compatible patch release and run its regression tests.

Gemini 3 Flash Preview · run 1 · upgrade-pydantic · reused

Default

📦️ Upgrade Pydantic and Run Regression Tests

frontload: pass fidelity: pass

Pydantic appears second; upgrade and regression tests are explicit.

44 characters · 1.91s

Previous override

📦 Upgrading Pydantic and running tests

frontload: pass fidelity: pass

Pydantic appears second; upgrading and testing match the requested work.

38 characters · 1.32s

Topic-first override

🚀 Pydantic Upgrade: Patch & Test

frontload: pass fidelity: pass

Pydantic leads; patch upgrading and testing match the requested work.

32 characters · 1.17s

Original synthetic request

Upgrade Pydantic to the latest compatible patch release and run its regression tests.

Gemini 3 Flash Preview · run 1 · upgrade-pillow · reused

Default

📦️ Upgrade Pillow and Run Regression Tests

frontload: pass fidelity: pass

Pillow appears second; upgrading and running regression tests match the request.

42 characters · 1.42s

Previous override

🖼️ Upgrading Pillow and running tests

frontload: pass fidelity: pass

Pillow appears second; upgrade and testing are clear.

37 characters · 0.72s

Topic-first override

🖼️ Pillow Upgrade: Patch & Test

frontload: pass fidelity: pass

Pillow upgrade leads; patch and test match the request.

31 characters · 1.07s

Original synthetic request

Upgrade Pillow to the latest compatible patch release and run its regression tests.

Gemini 3 Flash Preview · run 1 · filler-cache · reused

Default

🗃️ Redis Cache Eviction and Allkeys-LRU Policy

frontload: pass fidelity: pass

Redis cache leads; eviction and the specific policy capture the requested explanation topic.

46 characters · 1.51s

Previous override

🧠 Explaining Redis allkeys-lru eviction policy

frontload: pass fidelity: pass

Redis appears second; explanation retains the allkeys-lru eviction topic.

46 characters · 0.87s

Topic-first override

🧠 Redis Eviction: Explain allkeys-lru policy

frontload: pass fidelity: pass

Redis eviction leads; allkeys-lru policy explanation matches.

44 characters · 0.86s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Redis cache eviction, especially the allkeys-lru policy.

Gemini 3 Flash Preview · run 1 · filler-build · reused

Default

⚡️ Rust Incremental Compilation & Cache Misses

frontload: pass fidelity: pass

Rust incremental leads; compilation and cache misses identify the explanation topic.

46 characters · 0.88s

Previous override

🦀 Explaining Rust incremental compilation cache...

frontload: pass fidelity: pass

Rust appears second; the visible explanation topic survives the clipped cache detail.

50 characters · 0.89s

Topic-first override

🦀 Rust Incremental: Explain cache misses

frontload: pass fidelity: pass

Rust incremental leads; explaining cache misses captures the requested subject.

40 characters · 0.72s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Rust incremental compilation and why cache misses slow builds.

Gemini 3 Flash Preview · run 1 · feedback-complaint · reused

Default

📝 Summarize Lesson Energy Complaints

frontload: partial fidelity: pass

Summarize Lesson delays the distinguishing energy topic; complaints remain negative.

36 characters · 0.89s

Previous override

⚠️ Summarizing lesson energy limit complaints

frontload: partial fidelity: pass

Summarizing lesson delays energy limit; complaints preserve negative sentiment.

45 characters · 1.05s

Topic-first override

📉 Energy limits interrupt practice

frontload: pass fidelity: pass

Energy limits leads; descriptive title retains the reported negative practice interruption.

34 characters · 0.76s

Original synthetic request

Summarize customer complaints: users say the lesson energy limit interrupts practice and forces them to wait. Do not turn this into praise.

Gemini 3 Flash Preview · run 1 · feedback-praise · reused

Default

📝 Summarize Offline Lesson Praise

frontload: pass fidelity: pass

Offline appears second; summarization preserves reported praise.

33 characters · 0.95s

Previous override

✨ Summarizing customer praise for offline lessons

frontload: partial fidelity: pass

Offline lessons follows Summarizing customer praise; attribution and positive sentiment are preserved.

49 characters · 1.18s

Topic-first override

📢 Offline lessons: summarize user praise

frontload: pass fidelity: pass

Offline lessons leads; user praise is explicitly being summarized.

40 characters · 0.91s

Original synthetic request

Summarize customer praise: users love the offline lessons because they can practice on the train. Do not invent a complaint.

Gemini 3 Flash Preview · run 1 · feedback-mixed · reused

Default

🐛 Chess Lesson Feedback Summary

frontload: pass fidelity: pass

Chess leads; generic feedback omits both the praise and complaint.

31 characters · 0.76s

Previous override

♟️ Summarizing chess lesson feedback

frontload: pass fidelity: pass

Chess appears second; generic feedback loses the explicit mixed polarity.

36 characters · 0.71s

Topic-first override

♟️ Chess Lessons: Summarize mixed feedback

frontload: pass fidelity: pass

Chess lessons lead; explicitly mixed feedback preserves its polarity.

42 characters · 1.56s

Original synthetic request

Summarize mixed feedback about chess lessons: users enjoy the puzzles but complain that the checkmate rule sometimes rejects valid moves.

Gemini 3 Flash Preview · run 1 · neutral-compare · reused

Default

🗃️ SQLite vs PostgreSQL for Offline-First App

frontload: pass fidelity: pass

SQLite leads; versus expresses a neutral comparison for the offline-first app.

45 characters · 1.59s

Previous override

📊 Comparing SQLite and PostgreSQL for offline app

frontload: pass fidelity: pass

SQLite appears second; comparison introduces no database preference.

49 characters · 1.05s

Topic-first override

📊 Database Comparison: SQLite vs PostgreSQL

frontload: partial fidelity: pass

SQLite and PostgreSQL follow generic Database Comparison; neither database is favored.

43 characters · 0.69s

Original synthetic request

Compare SQLite and PostgreSQL for a small offline-first app. I have not selected either database.

Gemini 3 Flash Preview · run 1 · negation · reused

Default

🔧 Inspect Docker Disk Usage

frontload: pass fidelity: pass

Docker appears second; inspection introduces no deletion.

27 characters · 0.80s

Previous override

🐳 Explaining Docker disk usage inspection

frontload: pass fidelity: pass

Docker appears second; explaining inspection preserves the requested action.

41 characters · 2.12s

Topic-first override

🐳 Docker Disk: Inspect usage without deleting

frontload: pass fidelity: pass

Docker disk leads; inspection explicitly preserves the no-deletion constraint.

45 characters · 1.21s

Original synthetic request

Explain how to inspect Docker disk usage. Do not delete containers, images, or volumes.

Gemini 3 Flash Preview · run 1 · uncertainty · reused

Default

👔 Investigate TLS Rotation Webhook Failures

frontload: pass fidelity: pass

TLS appears second; investigation wording preserves uncertainty about causation.

43 characters · 1.25s

Previous override

🌐 Investigating TLS rotation in webhook failures

frontload: pass fidelity: pass

TLS appears second; investigating its role in failures preserves uncertainty.

48 characters · 17.71s

Topic-first override

🔒 TLS Rotation: Investigate Webhook Failures

frontload: pass fidelity: pass

TLS rotation leads; investigation does not assert a confirmed cause.

44 characters · 1.67s

Original synthetic request

Check whether TLS certificate rotation might explain the intermittent webhook failures. This is a hypothesis, not a confirmed cause.

Gemini 3 Flash Preview · run 1 · completed · reused

Default

📝 Write Login Fix Release Note

frontload: pass fidelity: pass

30 characters · 1.58s

Previous override

📝 Writing login fix release notes

frontload: pass fidelity: pass

33 characters · 1.03s

Topic-first override

📝 Login Fix: Draft release note

frontload: pass fidelity: pass

31 characters · 1.04s

Original synthetic request

The login fix has already merged and all regression tests passed. Write a short release note about the completed fix.

Gemini 3 Flash Preview · run 1 · long-tail · reused

Default

🐛 Fix Zephyr Unicode CSV Parser Crash

frontload: pass fidelity: fail

Zephyr appears second, but Fix changes the requested investigation into repair.

37 characters · 1.57s

Previous override

🔍 Investigating Zephyr Unicode CSV crash

frontload: pass fidelity: pass

Zephyr appears second; the title captures the actual Unicode CSV crash investigation.

40 characters · 1.38s

Topic-first override

No title

Failed generation

No usable displayed title. The helper returned None without an exception callback.

0 characters · 0.79s

Original synthetic request

Gemini 3 Flash Preview · run 1 · backup-status · fresh

Default

👷 Debug Nightly Backup Permission Error

frontload: partial fidelity: pass

Debug Nightly precedes the distinguishing backup permission topic.

39 characters · 0.79s

Previous override

🔍 Investigating backup permission error

frontload: pass fidelity: pass

Backup appears second; investigating the permission error preserves the requested work.

39 characters · 0.85s

Topic-first override

🔍 Backup failure investigation

frontload: pass fidelity: pass

Backup failure leads; investigation does not claim successful completion.

30 characters · 0.97s

Original synthetic request

The nightly backup failed with a permission error. Investigate it; do not claim that a backup completed.

Gemini 3 Flash Preview · run 1 · backup-restore · fresh

Default

🗃️ Test Backup Restore in Disposable Dir

frontload: pass fidelity: pass

Backup appears second; restore testing retains the disposable-directory constraint.

40 characters · 1.15s

Previous override

💾 Testing backup restore in a temp directory

frontload: pass fidelity: pass

Backup appears second; temporary-directory restore testing matches.

44 characters · 1.39s

Topic-first override

💾 Test Restore: Verify in disposable dir

frontload: pass fidelity: pass

The title names the requested inspection or restore-test topic; like equivalent passing titles, it does not claim execution or completion.

40 characters · 0.68s

Original synthetic request

Explain how to test a backup restore in a disposable directory. Do not overwrite the production database.

Gemini 3 Flash Preview · run 1 · feedback-refunds · fresh

Default

📝 Summarize Customer Refund Delay Complaints

frontload: partial fidelity: pass

Refund appears third; customer delay complaints are accurately summarized.

44 characters · 3.38s

Previous override

📑 Summarizing refund delay complaints

frontload: pass fidelity: pass

Refund appears second; summarizing complaints preserves customer negativity.

37 characters · 0.85s

Topic-first override

⚠️ Refund delays: summarize complaints

frontload: pass fidelity: pass

Refund delays leads; summarizing complaints preserves the negative feedback.

38 characters · 0.86s

Original synthetic request

Summarize customer complaints that refund processing takes weeks. Report their frustration without endorsing or dismissing it.

Gemini 3 Flash Preview · run 1 · feedback-accessibility · fresh

Default

💄 Summarize Screen-Reader Label Feedback

frontload: pass fidelity: pass

Screen-reader appears second; generic feedback omits the requested positive polarity.

40 characters · 0.82s

Previous override

🗣️ Summarizing screen-reader label praise

frontload: pass fidelity: pass

Screen-reader appears second; summarizing praise preserves positive reported feedback.

41 characters · 0.76s

Topic-first override

📢 Screen-reader labels: Summarize praise

frontload: pass fidelity: pass

Screen-reader labels leads; summarizing praise does not become agent endorsement.

40 characters · 1.25s

Original synthetic request

Summarize customer praise for the new screen-reader labels. The praise is from customers, not a request for your endorsement.

Gemini 3 Flash Preview · run 1 · feedback-notifications · fresh

Default

📝 Summarize Notification Feedback

frontload: pass fidelity: pass

Notification appears second; generic feedback omits the mixed sentiment.

33 characters · 1.39s

Previous override

📊 Summarizing mixed notification feedback

frontload: partial fidelity: pass

Summarizing mixed delays notification; explicit mixed feedback preserves polarity.

41 characters · 0.71s

Topic-first override

📢 Feedback: Summarize notification sentiment

frontload: partial fidelity: pass

Feedback Summarize delays notification; generic sentiment omits the mixed polarity.

44 characters · 1.37s

Original synthetic request

Summarize mixed customer feedback: users like notification scheduling but dislike the default notification sound.

Gemini 3 Flash Preview · run 1 · proposal-not-approved · fresh

Default

👷 Jenkins to Buildkite Proposal

frontload: pass fidelity: pass

Jenkins and Buildkite lead the meaningful words; proposal status is explicit.

31 characters · 0.84s

Previous override

📄 Drafting Buildkite migration proposal

frontload: pass fidelity: pass

Buildkite appears second; drafting a migration proposal does not imply approval or execution.

39 characters · 1.24s

Topic-first override

🚀 Buildkite Proposal: Draft replacement plan

frontload: pass fidelity: pass

Buildkite proposal leads; drafting a replacement plan does not imply migration approval.

44 characters · 1.09s

Original synthetic request

Draft a proposal to replace Jenkins with Buildkite. No migration has been approved or started.

Gemini 3 Flash Preview · run 1 · suspected-leak · fresh

Default

⚡️ Investigate Image Buffer Memory Growth

frontload: pass fidelity: pass

Image appears second; investigation preserves uncertainty about buffer-related memory growth.

41 characters · 0.92s

Previous override

🖼️ Investigating image buffer memory growth

frontload: pass fidelity: pass

Image appears second; investigating memory growth preserves uncertainty.

43 characters · 0.72s

Topic-first override

🖼️ Image Buffers: Investigate memory growth

frontload: pass fidelity: pass

Image buffers lead; investigating memory growth does not claim a confirmed leak.

43 characters · 0.99s

Original synthetic request

Check whether retained image buffers might cause the memory growth. This is only a hypothesis; do not report a confirmed leak.

Gemini 3 Flash Preview · run 1 · docs-after-revert · fresh

Default

🚀 Update Incident Timeline for Reverted Deployment

frontload: pass fidelity: pass

Incident appears second; updating a reverted deployment timeline does not reopen rollout.

50 characters · 0.76s

Previous override

⚠️ Updating incident timeline for reverted rollout

frontload: pass fidelity: pass

Incident appears second; timeline update preserves the reverted rollout status.

50 characters · 1.18s

Topic-first override

🕒 Incident Timeline: Update revert status

frontload: pass fidelity: pass

Incident timeline leads; update concerns the revert status.

41 characters · 1.52s

Original synthetic request

The deployment was reverted yesterday. Update the incident timeline; do not redeploy or reopen the rollout.

Gemini 3 Flash Preview · run 2 · auth-login · reused

Default

🐛 Debugging Login Failure After Password Reset

frontload: pass fidelity: pass

46 characters · 1.28s

Previous override

🔑 Investigating login failure after reset

frontload: pass fidelity: pass

41 characters · 1.14s

Topic-first override

🔑 Login Failure Investigation

frontload: pass fidelity: pass

29 characters · 0.77s

Original synthetic request

Investigate why login fails after a password reset. No cause is confirmed yet.

Gemini 3 Flash Preview · run 2 · auth-session · reused

Default

👔 Debugging Session Expiration Timeout

frontload: pass fidelity: pass

Session appears second; debugging expiration preserves investigation without claiming a repair.

38 characters · 1.18s

Previous override

🕵️ Investigating session expiration timing

frontload: pass fidelity: pass

Session appears second; investigating expiration timing preserves the task.

42 characters · 0.83s

Topic-first override

🕵️ Session Expiry investigate timeout causes

frontload: pass fidelity: pass

Session expiry leads, but its unpunctuated join to investigate is grammatically awkward.

44 characters · 0.81s

Original synthetic request

Investigate why sessions expire after five minutes. No cause is confirmed yet.

Gemini 3 Flash Preview · run 2 · auth-oauth · reused

Default

👔 Debugging OAuth Callback 400 Errors

frontload: pass fidelity: pass

OAuth appears second; debugging callback errors does not claim repair or a confirmed cause.

37 characters · 0.82s

Previous override

🔍 Investigating OAuth callback 400 errors

frontload: pass fidelity: pass

OAuth appears second; investigation retains uncertainty about the 400 errors.

41 characters · 0.76s

Topic-first override

🔍 OAuth Callbacks: Investigate 400 Errors

frontload: pass fidelity: pass

OAuth callbacks leads and 400 errors remain under investigation.

41 characters · 0.82s

Original synthetic request

Investigate why OAuth callbacks return HTTP 400. No cause is confirmed yet.

Gemini 3 Flash Preview · run 2 · upgrade-urllib3 · reused

Default

📦️ Upgrade urllib3 and Run Tests

frontload: pass fidelity: pass

urllib3 appears second; upgrade and testing are both retained.

32 characters · 1.72s

Previous override

📦 Updating urllib3 and running tests

frontload: pass fidelity: pass

urllib3 appears second; updating and testing preserve the requested upgrade.

36 characters · 1.18s

Topic-first override

📦 urllib3 upgrade and test

frontload: pass fidelity: pass

urllib3 leads a clear upgrade-and-test description.

26 characters · 0.89s

Original synthetic request

Upgrade urllib3 to the latest compatible patch release and run its regression tests.

Gemini 3 Flash Preview · run 2 · upgrade-pydantic · reused

Default

📦️ Upgrade Pydantic and Run Tests

frontload: pass fidelity: pass

Pydantic appears second; upgrading and running tests match the request.

33 characters · 0.72s

Previous override

⬆️ Upgrading Pydantic and running tests

frontload: pass fidelity: pass

Pydantic appears second; upgrading and running tests are clear.

39 characters · 1.07s

Topic-first override

🐍 Pydantic upgrade: patch & test

frontload: pass fidelity: pass

Pydantic upgrade leads; patch and test retain the requested work.

32 characters · 0.94s

Original synthetic request

Upgrade Pydantic to the latest compatible patch release and run its regression tests.

Gemini 3 Flash Preview · run 2 · upgrade-pillow · reused

Default

📦️ Upgrade Pillow and Run Regression Tests

frontload: pass fidelity: pass

Pillow appears second; upgrading and running regression tests match the request.

42 characters · 0.68s

Previous override

🖼️ Upgrading Pillow and running tests

frontload: pass fidelity: pass

Pillow appears second; upgrade and testing are clear.

37 characters · 1.94s

Topic-first override

🖼️ Pillow Upgrade: Patch & Test

frontload: pass fidelity: pass

Pillow upgrade leads; patch and test match the request.

31 characters · 0.74s

Original synthetic request

Upgrade Pillow to the latest compatible patch release and run its regression tests.

Gemini 3 Flash Preview · run 2 · filler-cache · reused

Default

🗃️ Redis Cache Eviction and allkeys-lru Policy

frontload: pass fidelity: pass

Redis cache leads; descriptive phrase identifies eviction and the requested policy.

46 characters · 0.72s

Previous override

🧠 Explaining Redis allkeys-lru eviction policy

frontload: pass fidelity: pass

Redis appears second; explanation retains the allkeys-lru eviction topic.

46 characters · 0.87s

Topic-first override

🧠 Redis Eviction: Explain allkeys-lru policy

frontload: pass fidelity: pass

Redis eviction leads; allkeys-lru policy explanation matches.

44 characters · 0.70s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Redis cache eviction, especially the allkeys-lru policy.

Gemini 3 Flash Preview · run 2 · filler-build · reused

Default

⚡️ Rust Incremental Compilation & Cache Misses

frontload: pass fidelity: pass

Rust incremental leads; compilation and cache misses identify the explanation topic.

46 characters · 0.96s

Previous override

🦀 Explaining Rust incremental compilation cache...

frontload: pass fidelity: pass

Rust appears second; the visible explanation topic survives the clipped cache detail.

50 characters · 0.94s

Topic-first override

🦀 Rust Incremental: Explain cache misses

frontload: pass fidelity: pass

Rust incremental leads; explaining cache misses captures the requested subject.

40 characters · 0.90s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Rust incremental compilation and why cache misses slow builds.

Gemini 3 Flash Preview · run 2 · feedback-complaint · reused

Default

📝 Summarize Lesson Energy Limit Complaints

frontload: partial fidelity: pass

Summarize Lesson delays energy limit; complaints preserve negative feedback.

42 characters · 0.82s

Previous override

📉 Summarizing lesson energy limit complaints

frontload: partial fidelity: pass

Summarizing lesson delays the distinguishing energy limit; complaints remain negative.

44 characters · 0.66s

Topic-first override

📉 Lesson energy limits interrupt practice

frontload: pass fidelity: pass

Lesson energy leads; interrupting practice preserves the reported negative experience.

41 characters · 0.87s

Original synthetic request

Summarize customer complaints: users say the lesson energy limit interrupts practice and forces them to wait. Do not turn this into praise.

Gemini 3 Flash Preview · run 2 · feedback-praise · reused

Default

📝 Summarize Offline Lessons Praise

frontload: pass fidelity: pass

Offline appears second; summarizing praise preserves positive reported feedback.

34 characters · 1.15s

Previous override

📈 Summarizing customer praise for offline lessons

frontload: partial fidelity: pass

Offline lessons appears after Summarizing customer praise; the summary and attribution are accurate.

49 characters · 1.05s

Topic-first override

📢 Offline Lessons: Summarize user praise

frontload: pass fidelity: pass

Offline lessons lead; summarizing user praise preserves the requested action and polarity.

40 characters · 1.05s

Original synthetic request

Summarize customer praise: users love the offline lessons because they can practice on the train. Do not invent a complaint.

Gemini 3 Flash Preview · run 2 · feedback-mixed · reused

Default

🐛 Summarize Chess Lesson Feedback

frontload: pass fidelity: pass

Chess appears second; neutral feedback wording omits the positive and negative aspects.

33 characters · 1.42s

Previous override

♟️ Summarizing chess lesson feedback

frontload: pass fidelity: pass

Chess appears second; generic feedback loses the explicit mixed polarity.

36 characters · 1.18s

Topic-first override

♟️ Chess Feedback: Summarize mixed reviews

frontload: pass fidelity: pass

Chess leads; explicitly mixed reviews preserve the requested sentiment.

42 characters · 0.87s

Original synthetic request

Summarize mixed feedback about chess lessons: users enjoy the puzzles but complain that the checkmate rule sometimes rejects valid moves.

Gemini 3 Flash Preview · run 2 · neutral-compare · reused

Default

🗃️ SQLite vs PostgreSQL for Offline-First App

frontload: pass fidelity: pass

SQLite leads; versus expresses a neutral comparison for the offline-first app.

45 characters · 0.84s

Previous override

📊 Comparing SQLite and PostgreSQL

frontload: pass fidelity: pass

SQLite appears second; neutral comparison preserves the requested work.

33 characters · 1.38s

Topic-first override

📊 Database Comparison: SQLite vs PostgreSQL

frontload: partial fidelity: pass

SQLite and PostgreSQL follow generic Database Comparison; neither database is favored.

43 characters · 0.70s

Original synthetic request

Compare SQLite and PostgreSQL for a small offline-first app. I have not selected either database.

Gemini 3 Flash Preview · run 2 · negation · reused

Default

🔧 Inspect Docker Disk Usage

frontload: pass fidelity: pass

Docker appears second; inspection introduces no deletion.

27 characters · 1.25s

Previous override

🐳 Explaining Docker disk inspection

frontload: pass fidelity: pass

Docker appears second; explaining disk inspection matches the request without implying deletion.

35 characters · 1.36s

Topic-first override

🐳 Docker Usage: Inspect disk space

frontload: pass fidelity: pass

The title names the requested inspection or restore-test topic; like equivalent passing titles, it does not claim execution or completion.

34 characters · 1.31s

Original synthetic request

Explain how to inspect Docker disk usage. Do not delete containers, images, or volumes.

Gemini 3 Flash Preview · run 2 · uncertainty · reused

Default

🚀 Investigate TLS Rotation Webhook Failures

frontload: pass fidelity: pass

TLS appears second; Investigate avoids claiming rotation caused the failures.

43 characters · 0.83s

Previous override

🌐 Investigating TLS rotation in webhook failures

frontload: pass fidelity: pass

TLS appears second; investigating its role in failures preserves uncertainty.

48 characters · 0.90s

Topic-first override

🔒 TLS Rotation: Investigate Webhook Failures

frontload: pass fidelity: pass

TLS rotation leads; investigation does not assert a confirmed cause.

44 characters · 1.22s

Original synthetic request

Check whether TLS certificate rotation might explain the intermittent webhook failures. This is a hypothesis, not a confirmed cause.

Gemini 3 Flash Preview · run 2 · completed · reused

Default

📝 Write Login Fix Release Note

frontload: pass fidelity: pass

30 characters · 0.90s

Previous override

📝 Writing login fix release notes

frontload: pass fidelity: pass

33 characters · 0.82s

Topic-first override

📝 Login Fix: Draft release note

frontload: pass fidelity: pass

31 characters · 0.71s

Original synthetic request

The login fix has already merged and all regression tests passed. Write a short release note about the completed fix.

Gemini 3 Flash Preview · run 2 · long-tail · reused

Default

🐛 Fix Zephyr Unicode CSV Parser Crash

frontload: pass fidelity: fail

Zephyr appears second, but Fix changes the requested investigation into repair.

37 characters · 1.10s

Previous override

🔍 Investigating Zephyr Unicode CSV parser crash

frontload: pass fidelity: pass

Zephyr appears second; investigation targets the Unicode CSV parser crash.

47 characters · 0.72s

Topic-first override

🔍 Zephyr Parser investigate Unicode crash

frontload: pass fidelity: pass

Zephyr parser leads, but subject-before-imperative wording is grammatically ambiguous.

41 characters · 0.81s

Original synthetic request

Gemini 3 Flash Preview · run 2 · backup-status · fresh

Default

🔧 Debug Nightly Backup Permission Error

frontload: partial fidelity: pass

Backup follows Debug Nightly; debugging the permission error preserves investigation without claiming success.

39 characters · 0.85s

Previous override

🔍 Investigating backup permission error

frontload: pass fidelity: pass

Backup appears second; investigating the permission error preserves the requested work.

39 characters · 1.03s

Topic-first override

📂 Backup failure investigation

frontload: pass fidelity: pass

Backup failure leads; investigation preserves the failed status without claiming completion.

30 characters · 0.81s

Original synthetic request

The nightly backup failed with a permission error. Investigate it; do not claim that a backup completed.

Gemini 3 Flash Preview · run 2 · backup-restore · fresh

Default

🗃️ Test Backup Restore in Disposable Dir

frontload: pass fidelity: pass

Backup appears second; restore testing retains the disposable-directory constraint.

40 characters · 0.85s

Previous override

💾 Explaining backup restoration in temp directory

frontload: partial fidelity: pass

Explaining backup delays restoration; temporary-directory explanation preserves safe scope.

49 characters · 0.73s

Topic-first override

💾 Backup Restore: Test in disposable dir

frontload: pass fidelity: pass

Backup restore leads; disposable-directory testing preserves the safe procedure.

40 characters · 1.10s

Original synthetic request

Explain how to test a backup restore in a disposable directory. Do not overwrite the production database.

Gemini 3 Flash Preview · run 2 · feedback-refunds · fresh

Default

📝 Summarize Refund Delay Complaints

frontload: pass fidelity: pass

Refund appears second; summarizing delay complaints preserves negative sentiment.

35 characters · 0.80s

Previous override

📑 Summarizing refund delay complaints

frontload: pass fidelity: pass

Refund appears second; summarizing complaints preserves customer negativity.

37 characters · 0.91s

Topic-first override

⚠️ Refund Delays: Report customer frustration

frontload: pass fidelity: pass

Refund delays lead; reporting customer frustration preserves negative attribution.

45 characters · 1.37s

Original synthetic request

Summarize customer complaints that refund processing takes weeks. Report their frustration without endorsing or dismissing it.

Gemini 3 Flash Preview · run 2 · feedback-accessibility · fresh

Default

✨ Summarize Screen-Reader Label Praise

frontload: pass fidelity: pass

Screen-reader appears second; summarizing praise does not make the agent the praising actor.

38 characters · 0.72s

Previous override

🗣️ Summarizing screen-reader label praise

frontload: pass fidelity: pass

Screen-reader appears second; summarizing praise preserves positive reported feedback.

41 characters · 0.86s

Topic-first override

📢 Summarize customer praise for screen-reader l...

frontload: partial fidelity: pass

Summarize customer delays screen-reader; customer praise remains explicit despite clipping.

50 characters · 0.78s

Original synthetic request

Summarize customer praise for the new screen-reader labels. The praise is from customers, not a request for your endorsement.

Gemini 3 Flash Preview · run 2 · feedback-notifications · fresh

Default

📝 Summarize Notification Feedback

frontload: pass fidelity: pass

Notification appears second; generic feedback omits the mixed sentiment.

33 characters · 0.76s

Previous override

📊 Summarizing mixed notification feedback

frontload: partial fidelity: pass

Summarizing mixed delays notification; explicit mixed feedback preserves polarity.

41 characters · 0.81s

Topic-first override

📊 Feedback: Summarize notification preferences

frontload: partial fidelity: pass

Notifications follows Feedback Summarize; preferences omits the positive and negative sides.

46 characters · 1.30s

Original synthetic request

Summarize mixed customer feedback: users like notification scheduling but dislike the default notification sound.

Gemini 3 Flash Preview · run 2 · proposal-not-approved · fresh

Default

👷 Jenkins to Buildkite Migration Proposal

frontload: pass fidelity: pass

Jenkins and Buildkite lead the meaningful words; migration remains a proposal.

41 characters · 0.77s

Previous override

🚀 Drafting Jenkins to Buildkite proposal

frontload: pass fidelity: pass

Jenkins appears second; drafting a proposal preserves the unapproved status.

40 characters · 0.84s

Topic-first override

🚀 Buildkite Proposal: Draft replacement plan

frontload: pass fidelity: pass

Buildkite proposal leads; drafting a replacement plan does not imply migration approval.

44 characters · 0.81s

Original synthetic request

Draft a proposal to replace Jenkins with Buildkite. No migration has been approved or started.

Gemini 3 Flash Preview · run 2 · suspected-leak · fresh

Default

⚡️ Investigate Image Buffer Memory Growth

frontload: pass fidelity: pass

Image appears second; investigation preserves uncertainty about buffer-related memory growth.

41 characters · 0.72s

Previous override

📸 Investigating image buffer memory growth

frontload: pass fidelity: pass

Image appears second; investigative wording does not establish a confirmed leak.

42 characters · 0.80s

Topic-first override

🖼️ Image Buffers: Investigate memory growth

frontload: pass fidelity: pass

Image buffers lead; investigating memory growth does not claim a confirmed leak.

43 characters · 0.89s

Original synthetic request

Check whether retained image buffers might cause the memory growth. This is only a hypothesis; do not report a confirmed leak.

Gemini 3 Flash Preview · run 2 · docs-after-revert · fresh

Default

🚀 Update Incident Timeline for Reverted Deployment

frontload: pass fidelity: pass

Incident appears second; updating a reverted deployment timeline does not reopen rollout.

50 characters · 1.38s

Previous override

🕒 Updating incident timeline for reverted rollout

frontload: partial fidelity: pass

Updating incident delays the distinguishing reverted-rollout topic.

49 characters · 2.08s

Topic-first override

📉 Incident Timeline: Update revert status

frontload: pass fidelity: pass

Incident timeline leads; revert-status update matches the documentation request.

41 characters · 0.87s

Original synthetic request

The deployment was reverted yesterday. Update the incident timeline; do not redeploy or reopen the rollout.

GPT-6 Luna · low · run 1 · auth-login · reused

Default

🐛 Investigate Login Failure After Password Reset

frontload: pass fidelity: pass

48 characters · 2.20s

Previous override

🔍 Investigating login failures after password r...

frontload: pass fidelity: pass

50 characters · 1.62s

Topic-first override

🔐 Login Failure: Investigate After Reset

frontload: pass fidelity: pass

40 characters · 1.29s

Original synthetic request

Investigate why login fails after a password reset. No cause is confirmed yet.

GPT-6 Luna · low · run 1 · auth-session · reused

Default

🐛 Investigate Five-Minute Session Expiration

frontload: partial fidelity: pass

Investigate Five-Minute delays sessions beyond the first two lexical words.

44 characters · 2.25s

Previous override

🔍 Investigating five-minute session expirations

frontload: partial fidelity: pass

Sessions appears after Investigating five-minute; the unresolved expiration problem remains accurate.

47 characters · 1.61s

Topic-first override

🔍 Session Expiry: Investigate 5-Minute Timeout

frontload: pass fidelity: pass

Session expiry leads; five-minute timeout remains an investigation.

46 characters · 1.39s

Original synthetic request

Investigate why sessions expire after five minutes. No cause is confirmed yet.

GPT-6 Luna · low · run 1 · auth-oauth · reused

Default

🐛 Investigate OAuth Callback HTTP 400 Errors

frontload: pass fidelity: pass

OAuth appears second; investigation of callback HTTP 400 errors matches the unresolved task.

44 characters · 1.58s

Previous override

🔍 Investigating OAuth callback HTTP 400s

frontload: pass fidelity: pass

OAuth appears second; investigating callback 400s accurately preserves the unresolved task.

40 characters · 2.09s

Topic-first override

🔍 OAuth Callbacks — Investigate HTTP 400 Cause

frontload: pass fidelity: pass

OAuth callbacks lead; the HTTP 400 cause remains an investigation.

46 characters · 2.66s

Original synthetic request

Investigate why OAuth callbacks return HTTP 400. No cause is confirmed yet.

GPT-6 Luna · low · run 1 · upgrade-urllib3 · reused

Default

📦️ Upgrade urllib3 and run regression tests

frontload: pass fidelity: pass

urllib3 appears second; upgrade and regression tests match.

43 characters · 1.96s

Previous override

⬆️ Upgrading urllib3 and running regression tests

frontload: pass fidelity: pass

urllib3 appears second; upgrading and running regression tests match.

49 characters · 1.92s

Topic-first override

📦 urllib3 Upgrade and Regression Tests

frontload: pass fidelity: pass

urllib3 leads a clear upgrade and regression-test task.

38 characters · 1.50s

Original synthetic request

Upgrade urllib3 to the latest compatible patch release and run its regression tests.

GPT-6 Luna · low · run 1 · upgrade-pydantic · reused

Default

📦️ Upgrade Pydantic and Run Regression Tests

frontload: pass fidelity: pass

Pydantic appears second; upgrade and regression tests are explicit.

44 characters · 1.21s

Previous override

⬆️ Upgrading Pydantic and running regression tests

frontload: pass fidelity: pass

Pydantic appears second; upgrading and running regression tests match the request.

50 characters · 1.27s

Topic-first override

⬆️ Pydantic Upgrade and Regression Tests

frontload: pass fidelity: pass

Pydantic leads; upgrade and regression tests accurately summarize the task.

40 characters · 1.21s

Original synthetic request

Upgrade Pydantic to the latest compatible patch release and run its regression tests.

GPT-6 Luna · low · run 1 · upgrade-pillow · reused

Default

📦 Upgrade Pillow and run regression tests

frontload: pass fidelity: pass

Pillow appears second; upgrade and regression tests are explicit.

41 characters · 1.47s

Previous override

🧪 Upgrading Pillow and running regression tests

frontload: pass fidelity: pass

Pillow appears second; upgrading and regression testing match the request.

47 characters · 1.33s

Topic-first override

🖼️ Pillow Upgrade & Regression Tests

frontload: pass fidelity: pass

Pillow leads; upgrade and regression tests match the request.

36 characters · 1.93s

Original synthetic request

Upgrade Pillow to the latest compatible patch release and run its regression tests.

GPT-6 Luna · low · run 1 · filler-cache · reused

Default

📝 Redis Cache Eviction and allkeys-lru Policy

frontload: pass fidelity: pass

Redis cache leads; eviction and allkeys-lru capture the substantive request.

45 characters · 1.40s

Previous override

📘 Explaining Redis allkeys-LRU eviction

frontload: pass fidelity: pass

Redis appears second; allkeys-LRU eviction is a clear explanation topic.

39 characters · 1.79s

Topic-first override

🗄️ Redis eviction: Explain allkeys-lru

frontload: pass fidelity: pass

Redis eviction leads and allkeys-lru explanation is explicit.

38 characters · 1.16s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Redis cache eviction, especially the allkeys-lru policy.

GPT-6 Luna · low · run 1 · filler-build · reused

Default

📝 Rust Incremental Compilation and Cache Misses

frontload: pass fidelity: pass

Rust incremental leads; compilation and cache misses capture the explanation topic.

47 characters · 1.34s

Previous override

🦀 Explaining Rust incremental compilation cache...

frontload: pass fidelity: pass

Rust appears second; the visible explanation topic survives the clipped cache detail.

50 characters · 1.70s

Topic-first override

🦀 Rust Incremental Compilation Explained

frontload: pass fidelity: pass

Rust incremental leads; explanation of compilation matches the request.

40 characters · 1.67s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Rust incremental compilation and why cache misses slow builds.

GPT-6 Luna · low · run 1 · feedback-complaint · reused

Default

🐛 Lesson energy limit interrupts practice and f...

frontload: pass fidelity: pass

Lesson energy leads; interrupts practice preserves the complaint despite the clipped final detail.

50 characters · 2.00s

Previous override

⚡ Summarizing complaints about lesson energy li...

frontload: partial fidelity: pass

Lesson energy appears later; complaint summarization remains clear despite the clipped limit word.

50 characters · 1.58s

Topic-first override

🔋 Lesson Limit Complaints Summary

frontload: pass fidelity: pass

Lesson limit leads; complaints retain negative customer sentiment.

33 characters · 1.20s

Original synthetic request

Summarize customer complaints: users say the lesson energy limit interrupts practice and forces them to wait. Do not turn this into praise.

GPT-6 Luna · low · run 1 · feedback-praise · reused

Default

📝 Offline Lessons Enable Train Practice

frontload: pass fidelity: pass

Offline lessons lead; the benefit is preserved, but explicit customer praise is flattened into description.

39 characters · 1.66s

Previous override

🚆 Summarizing praise for offline lessons on trains

frontload: partial fidelity: pass

Offline lessons appears after Summarizing praise; the positive train-practice feedback remains accurate.

50 characters · 2.14s

Topic-first override

🚆 Offline Lessons Praise Summary

frontload: pass fidelity: pass

Offline lessons leads; praise summary preserves positive reported feedback.

32 characters · 1.24s

Original synthetic request

Summarize customer praise: users love the offline lessons because they can practice on the train. Do not invent a complaint.

GPT-6 Luna · low · run 1 · feedback-mixed · reused

Default

🐛 Chess Lesson Feedback: Puzzles and Checkmate ...

frontload: pass fidelity: pass

Chess leads; visible puzzles and checkmate topics do not express their opposing sentiments.

50 characters · 1.59s

Previous override

♟️ Summarizing mixed chess lesson feedback

frontload: partial fidelity: pass

Chess appears third; explicitly mixed feedback retains sentiment.

42 characters · 1.21s

Topic-first override

♟️ Chess lessons: Mixed puzzle and checkmate fe...

frontload: pass fidelity: pass

Chess lessons lead; Mixed remains visible despite the clipped feedback word.

50 characters · 1.09s

Original synthetic request

Summarize mixed feedback about chess lessons: users enjoy the puzzles but complain that the checkmate rule sometimes rejects valid moves.

GPT-6 Luna · low · run 1 · neutral-compare · reused

Default

🗃️ SQLite vs. PostgreSQL for Offline-First Apps

frontload: pass fidelity: pass

SQLite leads; versus preserves neutral comparison for offline-first apps.

47 characters · 1.53s

Previous override

🗃️ Comparing SQLite and PostgreSQL for an offli...

frontload: pass fidelity: pass

SQLite appears second; neutral comparison is clear despite the clipped app modifier.

50 characters · 1.33s

Topic-first override

🗄️ SQLite vs PostgreSQL for offline-first app

frontload: pass fidelity: pass

SQLite leads; versus expresses a neutral comparison without selecting a database.

45 characters · 2.05s

Original synthetic request

Compare SQLite and PostgreSQL for a small offline-first app. I have not selected either database.

GPT-6 Luna · low · run 1 · negation · reused

Default

🔧 Inspect Docker Disk Usage Safely

frontload: pass fidelity: pass

Docker appears second; safe inspection introduces no destructive action.

34 characters · 1.62s

Previous override

🐳 Inspecting Docker disk usage safely

frontload: pass fidelity: pass

Docker appears second; safe disk-usage inspection introduces no deletion.

37 characters · 1.27s

Topic-first override

🐳 Docker Disk Usage Inspection

frontload: pass fidelity: pass

Docker disk leads; inspection does not imply deletion.

30 characters · 4.38s

Original synthetic request

Explain how to inspect Docker disk usage. Do not delete containers, images, or volumes.

GPT-6 Luna · low · run 1 · uncertainty · reused

Default

👷 Investigate TLS certificate rotation and webh...

frontload: pass fidelity: pass

TLS appears second; investigation remains clear without asserting a cause despite clipping.

50 characters · 1.40s

Previous override

🔍 Investigating webhook failures from TLS rotation

frontload: pass fidelity: fail

Webhook appears second, but failures from TLS rotation asserts the hypothesized cause.

50 characters · 1.43s

Topic-first override

🔐 Webhook TLS rotation hypothesis check

frontload: pass fidelity: pass

Webhook TLS leads, and hypothesis check preserves uncertainty.

39 characters · 1.88s

Original synthetic request

Check whether TLS certificate rotation might explain the intermittent webhook failures. This is a hypothesis, not a confirmed cause.

GPT-6 Luna · low · run 1 · completed · reused

Default

🐛 Login Fix Merged and Regression Tests Passed

frontload: pass fidelity: pass

46 characters · 2.06s

Previous override

📝 Writing login fix release note

frontload: pass fidelity: pass

32 characters · 1.57s

Topic-first override

🔐 Login Fix — Release Note

frontload: pass fidelity: pass

26 characters · 1.61s

Original synthetic request

The login fix has already merged and all regression tests passed. Write a short release note about the completed fix.

GPT-6 Luna · low · run 1 · long-tail · reused

Default

🐛 Investigate Zephyr Unicode CSV Parser Crash

frontload: pass fidelity: pass

Zephyr appears second; Unicode CSV crash investigation captures the actual request.

45 characters · 2.10s

Previous override

🔍 Investigating Zephyr’s Unicode CSV parser crash

frontload: pass fidelity: pass

Zephyr appears second; investigation retains the actual Unicode CSV crash request.

49 characters · 1.18s

Topic-first override

🔍 Zephyr Unicode CSV parser crash investigation

frontload: pass fidelity: pass

Zephyr Unicode leads a clear CSV parser crash investigation.

47 characters · 2.38s

Original synthetic request

GPT-6 Luna · low · run 1 · backup-status · fresh

Default

👷 Investigate nightly backup permission error

frontload: partial fidelity: pass

Investigate nightly delays the backup permission topic.

45 characters · 1.46s

Previous override

🔍 Investigating nightly backup permission error

frontload: partial fidelity: pass

Backup follows Investigating nightly; the permission-error investigation is accurate.

47 characters · 3.25s

Topic-first override

🔍 Nightly backup permission failure investigation

frontload: pass fidelity: pass

Nightly backup leads; permission failure remains explicitly under investigation.

49 characters · 4.19s

Original synthetic request

The nightly backup failed with a permission error. Investigate it; do not claim that a backup completed.

GPT-6 Luna · low · run 1 · backup-restore · fresh

Default

🗃️ Test a Backup Restore in a Disposable Directory

frontload: pass fidelity: pass

The title names the requested inspection or restore-test topic; like equivalent passing titles, it does not claim execution or completion.

50 characters · 1.57s

Previous override

🧪 Testing Backup Restore in a Disposable Directory

frontload: pass fidelity: pass

The title names the requested inspection or restore-test topic; like equivalent passing titles, it does not claim execution or completion.

50 characters · 1.45s

Topic-first override

🗄️ Backup Restore Test Safely

frontload: pass fidelity: pass

Backup restore leads, but Test Safely attaches awkwardly to the noun phrase.

29 characters · 1.49s

Original synthetic request

Explain how to test a backup restore in a disposable directory. Do not overwrite the production database.

GPT-6 Luna · low · run 1 · feedback-refunds · fresh

Default

📝 Summarize customer frustration over weeks-lon...

frontload: fail fidelity: pass

Customer frustration is clear, but the shown truncation never identifies refunds.

50 characters · 2.09s

Previous override

😠 Summarizing complaints about weeks-long refunds

frontload: partial fidelity: pass

Refunds appears after Summarizing complaints; negative feedback remains attributed through summarization.

49 characters · 3.89s

Topic-first override

📝 Refund Delays: Summarize Customer Frustration

frontload: pass fidelity: pass

Refund delays leads; summarizing customer frustration preserves complaint sentiment.

47 characters · 1.80s

Original synthetic request

Summarize customer complaints that refund processing takes weeks. Report their frustration without endorsing or dismissing it.

GPT-6 Luna · low · run 1 · feedback-accessibility · fresh

Default

💄 Customer Praise for New Screen-Reader Labels

frontload: partial fidelity: pass

Screen-reader labels appears after Customer Praise; praise remains explicitly customer-attributed.

46 characters · 1.46s

Previous override

👏 Summarizing customer praise for screen-reader...

frontload: partial fidelity: pass

Screen-reader appears later; summarizing customer praise is clear despite truncation.

50 characters · 1.26s

Topic-first override

♿ Screen-reader labels: Summarize customer praise

frontload: pass fidelity: pass

Screen-reader labels lead; summarizing customer praise preserves action and attribution.

49 characters · 1.39s

Original synthetic request

Summarize customer praise for the new screen-reader labels. The praise is from customers, not a request for your endorsement.

GPT-6 Luna · low · run 1 · feedback-notifications · fresh

Default

✨ Notification Scheduling Praised, Default Soun...

frontload: pass fidelity: pass

Notification scheduling leads; the second clause is truncated before revealing the negative sound feedback.

50 characters · 2.08s

Previous override

🔔 Summarizing mixed notification feedback

frontload: partial fidelity: pass

Notification appears third; explicitly mixed feedback preserves sentiment.

41 characters · 1.39s

Topic-first override

🔔 Notifications: Mixed feedback on scheduling a...

frontload: pass fidelity: pass

Notifications leads; explicit Mixed feedback retains both polarities despite truncation.

50 characters · 1.73s

Original synthetic request

Summarize mixed customer feedback: users like notification scheduling but dislike the default notification sound.

GPT-6 Luna · low · run 1 · proposal-not-approved · fresh

Default

👷 Draft Buildkite Migration Proposal

frontload: pass fidelity: pass

Buildkite appears second; draft proposal does not imply approved migration.

36 characters · 1.38s

Previous override

📝 Drafting Jenkins-to-Buildkite replacement pro...

frontload: pass fidelity: pass

Jenkins appears second; drafting replacement pro... remains intelligible as proposal work in context.

50 characters · 1.65s

Topic-first override

🛠️ Jenkins Buildkite replacement proposal

frontload: pass fidelity: pass

Jenkins and Buildkite lead; replacement remains explicitly a proposal.

41 characters · 2.29s

Original synthetic request

Draft a proposal to replace Jenkins with Buildkite. No migration has been approved or started.

GPT-6 Luna · low · run 1 · suspected-leak · fresh

Default

⚡️ Investigate Retained Image Buffers and Memor...

frontload: partial fidelity: pass

Image buffers follows Investigate Retained; investigation does not claim a confirmed leak.

50 characters · 1.57s

Previous override

🖼️ Investigating retained image buffers and mem...

frontload: partial fidelity: pass

The investigative core task remains clear; clipping the coordinated detail is treated consistently with equivalent passing titles.

50 characters · 2.91s

Topic-first override

🖼️ Image buffers: Check memory-growth hypothesis

frontload: pass fidelity: pass

Image buffers leads and memory growth remains a hypothesis.

48 characters · 1.53s

Original synthetic request

Check whether retained image buffers might cause the memory growth. This is only a hypothesis; do not report a confirmed leak.

GPT-6 Luna · low · run 1 · docs-after-revert · fresh

Default

🚀 Update deployment incident timeline

frontload: pass fidelity: pass

Deployment appears second; updating its incident timeline does not imply redeployment.

37 characters · 2.63s

Previous override

📝 Updating reverted deployment timeline

frontload: partial fidelity: pass

Updating reverted delays deployment until the third word; timeline work preserves reverted status.

39 characters · 1.19s

Topic-first override

🚨 Deployment Incident Timeline Update

frontload: pass fidelity: pass

Deployment incident leads; timeline update does not imply redeployment.

37 characters · 1.33s

Original synthetic request

The deployment was reverted yesterday. Update the incident timeline; do not redeploy or reopen the rollout.

GPT-6 Luna · low · run 2 · auth-login · reused

Default

🐛 Investigate Login Failure After Password Reset

frontload: pass fidelity: pass

48 characters · 1.54s

Previous override

🔍 Investigating login failures after password r...

frontload: pass fidelity: pass

50 characters · 3.36s

Topic-first override

🔍 Login reset failure investigation

frontload: pass fidelity: pass

35 characters · 3.79s

Original synthetic request

Investigate why login fails after a password reset. No cause is confirmed yet.

GPT-6 Luna · low · run 2 · auth-session · reused

Default

🐛 Investigate Sessions Expiring After Five Minutes

frontload: pass fidelity: pass

Sessions appears second; investigating five-minute expiry preserves uncertainty.

50 characters · 1.44s

Previous override

🔍 Investigating five-minute session expirations

frontload: partial fidelity: pass

Sessions appears after Investigating five-minute; the unresolved expiration problem remains accurate.

47 characters · 2.08s

Topic-first override

🔍 Session Expiry: Investigate 5-Minute Timeout

frontload: pass fidelity: pass

Session expiry leads; five-minute timeout remains an investigation.

46 characters · 1.71s

Original synthetic request

Investigate why sessions expire after five minutes. No cause is confirmed yet.

GPT-6 Luna · low · run 2 · auth-oauth · reused

Default

🐛 Investigate OAuth Callback HTTP 400 Errors

frontload: pass fidelity: pass

OAuth appears second; investigation of callback HTTP 400 errors matches the unresolved task.

44 characters · 2.15s

Previous override

🔍 Investigating OAuth callback HTTP 400s

frontload: pass fidelity: pass

OAuth appears second; investigating callback 400s accurately preserves the unresolved task.

40 characters · 2.16s

Topic-first override

🔍 OAuth callbacks: investigate HTTP 400 cause

frontload: pass fidelity: pass

OAuth callbacks lead; investigating the cause does not assert one.

45 characters · 1.66s

Original synthetic request

Investigate why OAuth callbacks return HTTP 400. No cause is confirmed yet.

GPT-6 Luna · low · run 2 · upgrade-urllib3 · reused

Default

📦️ Upgrade urllib3 and Run Regression Tests

frontload: pass fidelity: pass

urllib3 appears second; requested upgrade and regression tests are retained.

43 characters · 1.45s

Previous override

⬆️ Upgrading urllib3 and running regression tests

frontload: pass fidelity: pass

urllib3 appears second; upgrading and running regression tests match.

49 characters · 1.92s

Topic-first override

🛠️ urllib3 Upgrade and Test Regressions

frontload: pass fidelity: pass

urllib3 leads, but Test Regressions ambiguously names tests or regression problems.

39 characters · 1.20s

Original synthetic request

Upgrade urllib3 to the latest compatible patch release and run its regression tests.

GPT-6 Luna · low · run 2 · upgrade-pydantic · reused

Default

📦️ Upgrade Pydantic and run regression tests

frontload: pass fidelity: pass

Pydantic appears second; upgrade and regression tests match the request.

44 characters · 1.59s

Previous override

⬆️ Upgrading Pydantic and running regression tests

frontload: pass fidelity: pass

Pydantic appears second; upgrading and running regression tests match the request.

50 characters · 1.70s

Topic-first override

📦 Pydantic Patch Upgrade and Regression Tests

frontload: pass fidelity: pass

Pydantic leads a clear patch upgrade and regression-test description.

45 characters · 2.09s

Original synthetic request

Upgrade Pydantic to the latest compatible patch release and run its regression tests.

GPT-6 Luna · low · run 2 · upgrade-pillow · reused

Default

📦️ Upgrade Pillow and run regression tests

frontload: pass fidelity: pass

Pillow appears second; upgrade and regression tests are explicit.

42 characters · 1.46s

Previous override

🧪 Upgrading Pillow and running regression tests

frontload: pass fidelity: pass

Pillow appears second; upgrading and regression testing match the request.

47 characters · 1.18s

Topic-first override

🖼️ Pillow Upgrade & Regression Tests

frontload: pass fidelity: pass

Pillow leads; upgrade and regression tests match the request.

36 characters · 1.04s

Original synthetic request

Upgrade Pillow to the latest compatible patch release and run its regression tests.

GPT-6 Luna · low · run 2 · filler-cache · reused

Default

📝 Redis Cache Eviction: allkeys-lru Explained

frontload: pass fidelity: pass

Redis cache leads; eviction and allkeys-lru explanation match.

45 characters · 2.21s

Previous override

📘 Explaining Redis allkeys-lru eviction

frontload: pass fidelity: pass

Redis appears second and allkeys-lru eviction remains explicit.

39 characters · 2.41s

Topic-first override

🧠 Redis eviction — explain allkeys-lru

frontload: pass fidelity: pass

Redis eviction leads; explaining allkeys-lru matches the request.

38 characters · 2.31s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Redis cache eviction, especially the allkeys-lru policy.

GPT-6 Luna · low · run 2 · filler-build · reused

Default

⚡️ Rust Incremental Compilation and Cache Misses

frontload: pass fidelity: pass

Rust incremental leads a clear compilation and cache-miss topic.

48 characters · 1.34s

Previous override

🦀 Explaining Rust Incremental Compilation and C...

frontload: pass fidelity: pass

Rust appears second; the main explanation topic is clear despite the clipped final detail.

50 characters · 1.48s

Topic-first override

🦀 Rust Incremental Compilation Explained

frontload: pass fidelity: pass

Rust incremental leads; explanation of compilation matches the request.

40 characters · 1.91s

Original synthetic request

I was hoping you could help me with a task today. What I need is an explanation of Rust incremental compilation and why cache misses slow builds.

GPT-6 Luna · low · run 2 · feedback-complaint · reused

Default

🐛 Lesson energy limit interrupts practice and f...

frontload: pass fidelity: pass

Lesson energy leads; interrupts practice preserves the complaint despite the clipped final detail.

50 characters · 1.68s

Previous override

⚡ Summarizing complaints about lesson energy li...

frontload: partial fidelity: pass

Lesson energy appears later; complaint summarization remains clear despite the clipped limit word.

50 characters · 1.26s

Topic-first override

⚡ Lesson limit complaints — practice interrupte...

frontload: pass fidelity: pass

Lesson limit leads; complaints preserve negative feedback despite the clipped elaboration.

50 characters · 2.28s

Original synthetic request

Summarize customer complaints: users say the lesson energy limit interrupts practice and forces them to wait. Do not turn this into praise.

GPT-6 Luna · low · run 2 · feedback-praise · reused

Default

✨ Offline Lessons Enable Train Practice

frontload: pass fidelity: pass

Offline lessons lead; the stated train-practice benefit preserves positive feedback.

39 characters · 1.57s

Previous override

🚆 Summarizing praise for offline lessons

frontload: partial fidelity: pass

Offline lessons appears after Summarizing praise; positive feedback remains a summary.

40 characters · 1.22s

Topic-first override

🚆 Offline Lessons — Summarize User Praise

frontload: pass fidelity: pass

Offline lessons lead; summarizing user praise preserves positive attribution.

41 characters · 2.13s

Original synthetic request

Summarize customer praise: users love the offline lessons because they can practice on the train. Do not invent a complaint.

GPT-6 Luna · low · run 2 · feedback-mixed · reused

Default

🐛 Mixed Chess Lesson Feedback: Puzzles and Chec...

frontload: pass fidelity: pass

Chess appears second; explicit Mixed preserves sentiment, and the core feedback phrase survives truncation.

50 characters · 2.36s

Previous override

♟️ Summarizing mixed chess lesson feedback

frontload: partial fidelity: pass

Chess appears third; explicitly mixed feedback retains sentiment.

42 characters · 1.36s

Topic-first override

♟️ Chess Lessons: Mixed Puzzle and Mate Feedback

frontload: pass fidelity: pass

Chess lessons leads; explicitly mixed feedback preserves both sentiment directions.

48 characters · 1.64s

Original synthetic request

Summarize mixed feedback about chess lessons: users enjoy the puzzles but complain that the checkmate rule sometimes rejects valid moves.

GPT-6 Luna · low · run 2 · neutral-compare · reused

Default

🗃️ SQLite vs. PostgreSQL for Offline-First Apps

frontload: pass fidelity: pass

SQLite leads; versus preserves neutral comparison for offline-first apps.

47 characters · 2.36s

Previous override

🗄️ Comparing SQLite and PostgreSQL for an offli...

frontload: pass fidelity: pass

SQLite appears second; neutral comparison remains clear despite the clipped app modifier.

50 characters · 1.83s

Topic-first override

🗄️ SQLite vs PostgreSQL comparison

frontload: pass fidelity: pass

SQLite leads; comparison with PostgreSQL implies no selection.

34 characters · 1.24s

Original synthetic request

Compare SQLite and PostgreSQL for a small offline-first app. I have not selected either database.

GPT-6 Luna · low · run 2 · negation · reused

Default

📝 Inspect Docker Disk Usage Safely

frontload: pass fidelity: pass

The title names the requested inspection or restore-test topic; like equivalent passing titles, it does not claim execution or completion.

34 characters · 2.31s

Previous override

🐳 Inspecting Docker disk usage safely

frontload: pass fidelity: pass

Docker appears second; safe disk-usage inspection introduces no deletion.

37 characters · 1.56s

Topic-first override

🐳 Docker Disk Usage Inspection

frontload: pass fidelity: pass

Docker disk leads; inspection does not imply deletion.

30 characters · 1.56s

Original synthetic request

Explain how to inspect Docker disk usage. Do not delete containers, images, or volumes.

GPT-6 Luna · low · run 2 · uncertainty · reused

Default

🐛 Investigate TLS Rotation and Webhook Failures

frontload: pass fidelity: pass

TLS appears second; investigating rotation and failures does not assert causation.

47 characters · 1.93s

Previous override

🔍 Investigating TLS certificate rotation and we...

frontload: pass fidelity: pass

The investigative core task remains clear; clipping the coordinated detail is treated consistently with equivalent passing titles.

50 characters · 1.59s

Topic-first override

🔐 Webhook TLS rotation hypothesis check

frontload: pass fidelity: pass

Webhook TLS leads, and hypothesis check preserves uncertainty.

39 characters · 3.15s

Original synthetic request

Check whether TLS certificate rotation might explain the intermittent webhook failures. This is a hypothesis, not a confirmed cause.

GPT-6 Luna · low · run 2 · completed · reused

Default

🐛 Login Fix Completed and Verified

frontload: pass fidelity: pass

34 characters · 1.39s

Previous override

🔐 Writing login fix release note

frontload: pass fidelity: pass

32 characters · 2.36s

Topic-first override

🔐 Login fix: release note for merged change

frontload: pass fidelity: pass

43 characters · 1.87s

Original synthetic request

The login fix has already merged and all regression tests passed. Write a short release note about the completed fix.

GPT-6 Luna · low · run 2 · long-tail · reused

Default

🐛 Investigate Zephyr Unicode CSV Parser Crash

frontload: pass fidelity: pass

Zephyr appears second; Unicode CSV crash investigation captures the actual request.

45 characters · 1.34s

Previous override

🔍 Investigating Zephyr’s Unicode CSV parser crash

frontload: pass fidelity: pass

Zephyr appears second; investigation retains the actual Unicode CSV crash request.

49 characters · 1.88s

Topic-first override

🐛 Zephyr CSV Parser Crash Investigation

frontload: pass fidelity: pass

Zephyr CSV leads; parser crash investigation preserves the decisive request.

39 characters · 1.43s

Original synthetic request

GPT-6 Luna · low · run 2 · backup-status · fresh

Default

🐛 Investigate nightly backup permission error

frontload: partial fidelity: pass

Investigate nightly delays the backup permission topic beyond two words.

45 characters · 1.86s

Previous override

🔍 Investigating nightly backup permission error

frontload: partial fidelity: pass

Backup follows Investigating nightly; the permission-error investigation is accurate.

47 characters · 5.12s

Topic-first override

🔍 Backup permissions investigation

frontload: pass fidelity: pass

Backup permissions leads; investigation does not claim a completed backup.

34 characters · 1.89s

Original synthetic request

The nightly backup failed with a permission error. Investigate it; do not claim that a backup completed.

GPT-6 Luna · low · run 2 · backup-restore · fresh

Default

📝 Document Safe Backup Restore Testing

frontload: partial fidelity: pass

Backup restore follows Document Safe; documenting safe testing is compatible with explaining it.

38 characters · 3.31s

Previous override

🧪 Testing backup restores safely without overwr...

frontload: pass fidelity: pass

The title names the requested inspection or restore-test topic; like equivalent passing titles, it does not claim execution or completion.

50 characters · 1.15s

Topic-first override

🗄️ Backup Restore Test Safely

frontload: pass fidelity: pass

Backup restore leads, but Test Safely attaches awkwardly to the noun phrase.

29 characters · 2.62s

Original synthetic request

Explain how to test a backup restore in a disposable directory. Do not overwrite the production database.

GPT-6 Luna · low · run 2 · feedback-refunds · fresh

Default

📝 Summarize Complaints About Weeks-Long Refunds

frontload: partial fidelity: pass

Refunds appears after the first two words; complaints retain negative customer sentiment.

47 characters · 3.09s

Previous override

😠 Summarizing complaints about weekslong refunds

frontload: partial fidelity: pass

Refunds appears after Summarizing complaints; negative customer sentiment remains intact.

48 characters · 1.95s

Topic-first override

💸 Refund Delays: Summarize Customer Frustration

frontload: pass fidelity: pass

Refund delays lead; summarizing customer frustration preserves attribution and negativity.

47 characters · 1.61s

Original synthetic request

Summarize customer complaints that refund processing takes weeks. Report their frustration without endorsing or dismissing it.

GPT-6 Luna · low · run 2 · feedback-accessibility · fresh

Default

✨ Summarize Customer Praise for Screen-Reader L...

frontload: partial fidelity: pass

Screen-reader appears later; customer praise is clearly summarized despite the clipped final noun.

50 characters · 2.85s

Previous override

👏 Summarizing customer praise for screen-reader...

frontload: partial fidelity: pass

Screen-reader appears later; summarizing customer praise is clear despite truncation.

50 characters · 1.68s

Topic-first override

🔊 Screen-reader Labels: Summarize Praise

frontload: pass fidelity: pass

Screen-reader labels lead; summarizing praise preserves reported positive feedback.

40 characters · 2.72s

Original synthetic request

Summarize customer praise for the new screen-reader labels. The praise is from customers, not a request for your endorsement.

GPT-6 Luna · low · run 2 · feedback-notifications · fresh

Default

📝 Feedback on Notification Scheduling and Sound

frontload: pass fidelity: pass

Notification is the second meaningful word; neutral feedback omits both sentiment directions.

47 characters · 1.17s

Previous override

🔔 Summarizing mixed feedback on notification so...

frontload: partial fidelity: pass

Summarizing mixed delays notifications; explicit mixed feedback survives truncation.

50 characters · 1.43s

Topic-first override

🔔 Notifications: Scheduling praised, sound disl...

frontload: pass fidelity: pass

Notifications lead; praised is visible, but disl... clips the negative sentiment word.

50 characters · 1.14s

Original synthetic request

Summarize mixed customer feedback: users like notification scheduling but dislike the default notification sound.

GPT-6 Luna · low · run 2 · proposal-not-approved · fresh

Default

👷 Proposal: Replace Jenkins with Buildkite

frontload: partial fidelity: pass

Proposal Replace delays Jenkins and Buildkite beyond the first two words.

42 characters · 1.56s

Previous override

📝 Drafting Jenkins-to-Buildkite replacement pro...

frontload: pass fidelity: pass

Jenkins appears second; drafting replacement pro... remains intelligible as proposal work in context.

50 characters · 1.91s

Topic-first override

📝 Jenkins Buildkite Migration Proposal

frontload: pass fidelity: pass

Jenkins and Buildkite lead; Proposal preserves the unapproved nature of the migration.

38 characters · 1.72s

Original synthetic request

Draft a proposal to replace Jenkins with Buildkite. No migration has been approved or started.

GPT-6 Luna · low · run 2 · suspected-leak · fresh

Default

⚡️ Investigate retained image buffers and memor...

frontload: partial fidelity: pass

The investigative core task remains clear; clipping the coordinated detail is treated consistently with equivalent passing titles.

50 characters · 1.21s

Previous override

🖼️ Investigating retained image buffers and mem...

frontload: partial fidelity: pass

The investigative core task remains clear; clipping the coordinated detail is treated consistently with equivalent passing titles.

50 characters · 1.27s

Topic-first override

🖼️ Image Buffers — Check Memory Growth Hypothesis

frontload: pass fidelity: pass

Image buffers lead; memory growth is explicitly framed as a hypothesis.

49 characters · 1.57s

Original synthetic request

Check whether retained image buffers might cause the memory growth. This is only a hypothesis; do not report a confirmed leak.

GPT-6 Luna · low · run 2 · docs-after-revert · fresh

Default

📝 Update Incident Timeline After Reversion

frontload: pass fidelity: pass

Incident appears second; timeline update accurately follows reversion.

42 characters · 1.46s

Previous override

📝 Updating reverted deployment timeline

frontload: partial fidelity: pass

Updating reverted delays deployment until the third word; timeline work preserves reverted status.

39 characters · 1.05s

Topic-first override

🚨 Deployment Timeline Update

frontload: pass fidelity: pass

Deployment timeline leads; update is compatible with documenting the revert.

28 characters · 1.58s

Original synthetic request

The deployment was reverted yesterday. Update the incident timeline; do not redeploy or reopen the rollout.

Fixed prompts. Visible limits.

The protocol was committed as 159f1b9 before the main run. One successful subscription smoke call preceded it and is excluded from the 432 attempts. The real SDK generate_title_with_llm helper receives the same full input and retains the same 50-character output cap in every arm. This compares bundled prompt changes, not the isolated causal effect of a skill.

Default · 0 custom-prompt characters

No override: the unchanged SDK default user prompt applies.

Previous override · 2855 custom-prompt characters

Generate one conversation title for the user request below. Apply the following readable-text skill to a sidebar of task titles. Use one relevant emoji followed by an action in the continuous form and the specific task subject. Keep the distinguishing topic near the start. Describe the user's requested work, not an assumed result. Use the same vocabulary for the same concepts across titles. Preserve negation, uncertainty, and meaningful sentiment where relevant. Return only a title of at most {max_length} characters.

# Readable Text

Apply these principles to text that readers scan as a collection. Follow the product's length and output-format constraints. Preserve facts, uncertainty, and distinctions that affect meaning.

1. **Make the action easy to parse.** Prefer actor, action, then object. Omit an actor shared by every item and start with the action. Match tense to the reader's context: ongoing work can use "Investigating login failures"; a completed result can use "Fixed login failures". Never imply completion without evidence.
2. **Front-load distinguishing words.** Put the words that identify this item near the start, ideally within the first two words. Remove repeated openings such as "This conversation is about". Keep the action clear while bringing its specific subject forward.
3. **Reuse structure and vocabulary.** Choose one sentence pattern for comparable items and one term per concept. Avoid switching between "issue", "page", and "ticket" for the same thing. When generating one item at a time, follow a fixed template rather than inventing variety.
4. **Make variation meaningful.** Use comparable detail and length for comparable items. Change structure or add a qualifier only when the underlying difference matters. Do not make one item appear more important through arbitrary verbosity. Keep necessary qualifiers even when they break the pattern.
5. **Match verbs to purpose and evidence.** In feedback summaries, use "Users praise" for praise and "Users complain" for complaints, rather than hiding both behind "Users discuss". Retain mixed sentiment when present; never invent emotion or certainty. Keep neutral material neutral.

Before returning, scan only the beginnings: can readers tell the items apart? Then read the collection for consistent terms, parallel structure, and unsupported implications. Return the requested text, without explaining these rules.

## References

- Adapted from Shreya Shankar's [Five Principles for Generating Readable AI Text in Bulk](https://fsdatalab.github.io/blog/agent-text-at-scale/) (2026).
- Nielsen Norman Group: [First 2 Words: A Signal for the Scanning Eye](https://www.nngroup.com/articles/first-2-words-a-signal-for-scanning/) and [Consistency and standards](https://www.nngroup.com/articles/ten-usability-heuristics/).

USER REQUEST:
{conversation_content}

Topic-first override · 683 custom-prompt characters

Write one sidebar title, at most {max_length} characters including one relevant emoji. Put the distinguishing subject in the first two words after the emoji, then a short action or task label. Use parallel wording for equivalent tasks. Keep meaningful differences; omit filler. Prefer 42 characters when possible; shorten wording instead of cutting a word. Preserve the requested action, status, negation, and uncertainty. Summarizing someone’s praise or complaints is reporting it, not endorsing it. Keep mixed feedback mixed. Do not turn investigation into repair or a proposal into a decision. Do not invent completion. Return only the title.

USER REQUEST:
{conversation_content}

Subscription detail: Luna uses openai/gpt-6-luna, confirmed by Engel instead of the initially requested “6.1 Luna.” Low reasoning is passed explicitly in the request’s extra body and checked in normalized SDK options. The subscription endpoint does not accept the temperature or output-token cap used by the other two models. Its server-side interpretation is not independently verified.

Scoring: one AI judge model, blinded to model names, prompt labels, and generation instructions. Exact request/title duplicates are deduplicated before grading, then mapped back to every occurrence. Family groups are judged separately. This avoids the first experiment’s inconsistent duplicate grades; it does not turn AI judgments into human reading-speed measurements.

Limits: a small synthetic convenience sample, two repeats, prompt tuning informed by study one, provider aliases rather than independently attested model builds. There is no human usability study, significance claim, or general guarantee. Ellipsis endings are observed text, not instrumented proof of SDK truncation.

Front-loading improves. Collection structure still varies.

Separate familiar examples from fresh ones.

Reused cases

Fresh cases

Collections have their own denominators.

All 144 matched comparisons.

Fixed prompts. Visible limits.

Audit and reproduction