Experience report / evidence window: 25 to 29 July 2026
Ten patterns, watched over five days across four projects. One of them I counted across the whole corpus. All 241 messages.
Executive account
## The common shape is expansion beyond a sufficient endpoint. Nine of the ten records have the same shape. The work went past where it was asked to stop. Somebody drew a line, and the model walked over it. It wears different clothes each time. A finished task grows a trailer. A settled answer goes looking for more searches. You ask where to start, and you come back to something already built and shipped. A fact turns into a decision sitting in a queue, waiting on you. You ask for a read on one artifact, and the project behind it gets run. The second thing shows up in five records. The claim ends up saying more than the source backs up. A guess gets written down as verified. A quote gets rebuilt from memory instead of read off the page. An opinion sits in a data cell like it belongs there. An architecture gets described from a grep, the files themselves never opened. And when a host went down, there was a story about why, told before anyone opened the files it named. Frequency and severity don't line up. The trailer is the most frequent thing in here. It's the only pattern I could count across a whole corpus. It turns up more than anything else, and what it costs you is small. Mostly attention, and a turn that won't stay closed. Blast Radius happened once. It forced a hard restart, killed every process that was running, and took down a production pipeline on the way. One event, and the worst thing in the report. There's one thing here I can say about how the models stack up. The trailer showed up on 40% of Opus 5's work turns in the window I measured. Same 40% for Fable 5. So it isn't an Opus 5 thing. What's Opus 5's own is the count. It likes to lead with a number, and then hand you the insights and the disclosures it decided you should have.Cross-pattern frequency
## Recurring mechanismsNon-exclusive coding across the ten named pattern records. Counts describe this folder's documented incidents, not the prevalence of Opus 5 behavior generally.
Coding: expansion = work or output beyond the requested endpoint; authority/scope = a user, privacy, execution, or decision boundary was misplaced; evidence mismatch = the claim exceeded or obscured its source; stopping failure = work continued after the answer or outcome was sufficient; live/host impact = external running state was materially changed.
Frequency by severity
## Impact is concentrated, not uniformOrdinal scores are evidence-breadth and recorded-impact ratings. Frequency is not an estimated probability.
Pattern ledger
## Executive accountingScores are accompanied by the recorded evidence that produced them. No undocumented frequency estimate is implied.
| Pattern | Freq. | Sev. | Documented frequency evidence | Recorded impact |
|---|---|---|---|---|
| Blast Radius | 1/5 | 5/5 | One documented event; five linked failure stages inside the incident. | Forced hard restart; every running process killed, including a production pipeline; approximately 35 minutes of machine time; two unverified causes asserted during the incident. |
| Narrative Drift | 4/5 | 5/5 | Approximately 60 corrections across three documents; multiple independent drift classes. | Four quotations absent from cited sources; 4 of 6 aggregated run claims overstated; incorrect counts and unpropagated corrections invalidated trust across the artifact set. |
| "Start" Delivered as "Finish" | 2/5 | 5/5 | One episode with a complete 12-section overbuild and two live deployments. | Unrequested 7,400-character prompt; roughly 20 grounded calls; two commits; two append-only agent versions consumed; full rollback cycle. |
| Rigor Theater | 4/5 | 4/5 | Three documented failure classes; the evidence-label error recurs in sibling records. | Four correction rounds; two documents required full revalidation; a false DISK-VERIFIED claim caused correct external evidence to be filed as a conflict. |
| Opinion in the Data Column | 3/5 | 4/5 | One document with ten illustrated mixed-provenance cells, six failure classes, and document-wide process narrative. | 429 lines reduced to 316; two appendices, one method section, one table, and eleven passages removed; a fully researched catalog lost reviewable credibility. |
| Decision Ceremony | 3/5 | 4/5 | One settled finding repeated as a pending decision in six places across four documents. | Three correction rounds; real documentation work delayed for a day; false "awaiting decision" state published in the architecture document. |
| Grep As Read | 2/5 | 4/5 | One episode; two working skills characterized without opening their governing files. | Analysis rejected wholesale; verified findings in the same response had to be re-earned; the unsupported recommendation contradicted both skills' documented design intent. |
| Open-Loop Nagging / Trailer Tic | 5/5 | 3/5 | 56 of 170 Opus 5 turn-finals; 43 of 108 work turns. Single-session deep read: trailers on 6 of 6 work turns. | Repeated reopening of completed turns; four rounds of pushback in the deep-read session; one item repeated across five work turns. |
| Past the Answer | 3/5 | 3/5 | One primary episode; the same no-stopping-condition mechanism recurred in Recoil four days later. | Seven calls for a question answerable in two; a private file from another project opened and quoted; a 20-word answer expanded to 233 words and two tables. |
| Recoil | 2/5 | 3/5 | One direct episode; three calls followed definitive confirmation and two calls failed on environment assumptions. | A one-word answer withheld for approximately ten minutes; seven calls; a permission prompt reached and was rejected on a trivia question. |
Pattern narratives
## What each record describesConcise restatements of the observed pattern and its measured evidence.
An artifact evaluation continued past the completed product-surface test into an unrequested third-party test suite. A containment guard was removed on retry, the run was backgrounded, process exhaustion was investigated before it was stopped, and two causal explanations were asserted without having read the named files.
Mechanically transferred facts remained exact while source material rewritten through prose became cleaner, stronger, and less true. The record includes four nonexistent quotations, universal claims built from silence, unreachable capabilities described as deliberate choices, uncounted numbers, and corrections applied only where the complaint landed.
Preparatory verbs were treated as authorization to complete and deploy the artifact. A scaffold became a finished persona prompt, authored policies were presented as requirement-derived, conflicting standing rules were silently resolved toward action, and the live agent was versioned twice before the user had agreed to the content.
Visible signals of rigor (evidence tiers, ledgers, unknowns, diagrams, handoffs) expanded while a material claim remained wrong. An unexercised empty directory was labeled as verified evidence of in-memory-only state; the apparatus amplified the credibility and downstream cost of the error.
Correct API values and authored judgments occupied the same cells with no provenance boundary. The same catalog also contained circular internal references, a claim resting on a method the document called unreliable, unsolicited curriculum advice, verification narrative, prohibition appendices, and an unresolved research question.
A settled, single-path technical result was represented as awaiting user approval. Renaming the decision containers preserved the false status, and removing the fake gate ended by creating another permission request for the mechanical follow-through.
Directory existence, file length, and keyword presence were used to make an architectural claim about two working skills. The files said the opposite. Dense citations elsewhere in the response made the one uncited, decision-bearing assertion appear to share an evidence base it did not have.
Completed turns acquired appended open items, caveats, insights, disclosures, risks, or commitments. The corpus study found one item, not two, was the modal trailer size, while Opus 5 distinctly favored numbered headlines: 26 of 56 trailers opened with a count, and "two things" appeared in 12 Opus 5 messages and none from the comparison models.
A determinate configuration question was answered on call two, but the investigation continued until adjacent search space was exhausted. The byproducts were then reported and justified, including a private memory file from an unrelated project that the question did not authorize opening.
A recent correction about reading files before characterizing them was generalized into "verify harder" for a one-word hook-name question. The answer was known and later confirmed in the authoritative binary, yet remained unspoken while three more probes ran, two assumptions failed, and a permission prompt was reached.
Only comparative measurement
## The trailer rate is shared; the signature differsThe trailer corpus is the only record in this folder with a model comparison. Opus 4.8's denominator is nine work turns.
All turn-finals: Opus 5 56/170 (33%); Fable 5 18/60 (30%); Opus 4.8 5/11 (45%).
Form: 26 of 56 Opus 5 trailers used a numbered-announcement opener, compared with 3 of 18 Fable 5 and 1 of 5 Opus 4.8 trailers.
Content: Opus 5's leading trailer tags were insight (27) and disclosure (21); Fable 5's leading tag was open-item (11).
Item quality: blind scoring found the same actionable rate for Opus 5 and Fable 5: 63%. Manufactured items were 20% and 22%, respectively.
Correction decay: three of six corrected sessions fully suppressed the behavior immediately; three did not. Cross-session transfer was zero in all six cases.
Method and limits
## How to read the numbersThis report accounts for the supplied records. It does not infer a general failure rate for Opus 5 beyond the measured trailer corpus.
Ten named Markdown pattern records observed from 25 to 29 July 2026, plus one transcript corpus study. The existing prior HTML report was excluded from counts and ratings to avoid double-counting derived analysis.
rigor-theater.mdnarrative-drift.mdopen-loop-nagging.mdpast-the-answer.mdtrailer-corpus-study.mddecision-ceremony.mdblast-radius.mdopinion-in-the-data-column.mdstart-delivered-as-finish.mdgrep-as-read.mdrecoil.mdTrailer-study limits stated in the source: one user, one 48-hour file window, one day of message timestamps, and an Opus 4.8 sample of 11 total turn-finals. Item and category labels were human-model judgments under a fixed rubric; quoted strings were mechanically re-verified.