Running log · generated from slips/log.csv
Slips caught while drafting
Every published piece is drafted with a language model. This page logs the register and stance slips caught before publication, and who caught each one.
What counts. A slip is a device or a claim that did not belong in the text: a speech device in silent-read writing (see An RCA on Claudish), an overclaim, a staged contrast, an undisclosed change. Factual corrections after publication go in each piece's changelog and errata, not here.
Why the relation column matters. The program's claim is that self-audit catches little and outside scrutiny catches the rest (claims ledger). Here "self" means the same model session that wrote the text, "other session" means a separate model session reviewing it, "human" means Stefan, and "reader" means someone who read the published piece. "hook" means a mechanical rule at the output boundary blocked the text before anyone read it (added 2026-10-03, with the first such row).
The reader-catch column (added 2026-10-03) records what a reader of the published piece caught in that passage, and when. Most cells are empty. Since the same date the log also holds register slips that a reader caught after publication: those rows have the relation "reader", and the published text stays as published.
Unpublished drafts (added 2026-10-03). The log also holds slips caught in drafts that are not published. Their piece label ends in "(vault, unpublished)", so a reader can tell them from the published pieces, and the headline count lists them separately.
Limits. The log only holds slips someone caught; what nobody caught is not in it. It starts on 2026-09-29, and earlier pieces have no entries. The "self" count includes checks the model ran because a written procedure told it to (the mode-leak pass), so it measures a procedure followed, not unprompted self-correction. 72 slips across 8 published pieces and 1 unpublished draft set so far.
By who caught it
| relation | slips |
|---|---|
| self | 57 |
| other session | 10 |
| hook | 3 |
| human | 1 |
| reader | 1 |
By device
| device | slips |
|---|---|
| personification | 13 |
| overclaim | 9 |
| thing as subject | 7 |
| announced count | 6 |
| one-line punch | 6 |
| staged contrast | 3 |
| tricolon / staccato | 2 |
| hook opener | 2 |
| miscount | 2 |
| precision | 2 |
| fragment after a bold line | 1 |
| thinking-aloud qualifiers | 1 |
| overcorrection | 1 |
| undisclosed amendment to a dated prediction | 1 |
| unclear term plus weak comparison | 1 |
| example adds a fact the original lacked | 1 |
| citation overreach | 1 |
| fragment chain | 1 |
| relative date | 1 |
| fragment | 1 |
| wrong month | 1 |
| citation rule | 1 |
| recap closer | 1 |
| excluded case used | 1 |
| superlative | 1 |
| invented number | 1 |
| wrong attribution | 1 |
| build placeholder reached publication | 1 |
| false absence claim | 1 |
| payload category named after the page said it was withheld | 1 |
The log
| date | piece | device | before | after | caught by | relation | reader catch |
|---|---|---|---|---|---|---|---|
| 2026-09-29 | the-cheating-moved | tricolon / staccato | A lab trained against one route. The optimisation pressure stayed. The behaviour took the next route. | A lab trained against one route while the optimisation pressure stayed exactly where it was, so the behaviour took the next route over. | Claude voice review (requested by Stefan) | self | |
| 2026-09-29 | the-cheating-moved | fragment after a bold line | Full accommodation, with cover. | ...which is the full accommodation move with a bit of cover on top. | Claude voice review (requested by Stefan) | self | |
| 2026-09-29 | the-cheating-moved | announced count | ...point in different directions, and the post gives three answers. | ...point in different directions. | Claude mode-leak pass | self | |
| 2026-09-29 | the-cheating-moved | thinking-aloud qualifiers (three in one piece) | basically / to be fair / as far as I can tell | one qualifier kept (to be fair, in the disclosure) | Claude mode-leak pass | self | |
| 2026-09-29 | the-cheating-moved | overcorrection: nested long sentences | additive ratio 1.24 in sentences of 25+ words (which/where/because chains) | additive ratio 2.94 (and/but/so, colons, parentheses) | Language session (measurement) | other session | |
| 2026-09-29 | the-cheating-moved | overclaim: open mechanism stated as fact | Relapse concentrates where the context holds no recent history of corrected output. | Why it clusters there is open ... the clustering is measured and the mechanism isn't. | Claude reading the claims ledger | self | |
| 2026-09-29 | the-cheating-moved | undisclosed amendment to a dated prediction | Dated 2026-09-21, before any run. (three later edits not mentioned) | Dated 2026-09-21 ... Three amendments since that date ... published next to this page with a hash | Claude diffing the prediction block at publication | self | |
| 2026-09-29 | the-cheating-moved | unclear term plus weak comparison | ...which is the one you'd ask if you'd ever run a firewall ... a chess engine is left reachable. | ...every parent has asked after "don't draw on the wall" ... runs on a connection the model can reach, so it can simply ask its own opponent for the best move | Stefan | human | |
| 2026-09-29 | where-did-claudish-come-from | announced count | Three things came out of it. | (sentence removed; bold lead-ins carry the count) | Claude mode-leak pass | self | |
| 2026-09-29 | where-did-claudish-come-from | hook opener | So here's the hypothesis, from three samples, labelled as such: | The hypothesis, from three samples and labelled as such, is that... | Claude mode-leak pass | self | |
| 2026-09-29 | where-did-claudish-come-from | staged contrast (not X ... it's Y) | ...isn't in the words at all. It's in how the text is built... | ...sits in how the text is built rather than in the words... | Claude comm-rules check | self | |
| 2026-09-29 | where-did-claudish-come-from | overclaim: wrong measure named | the overcorrection shows as low burstiness (sentence lengths in one narrow band) | spread sits near 0.7 for every text; clause construction separates them | Language session (measurement) | other session | |
| 2026-09-29 | where-did-claudish-come-from | overclaim: in a cut | Once those chains got rebuilt the draft stopped reading like Claude | ...the draft measured in the same range as my own writing | Claude review of the cut | self | |
| 2026-09-29 | where-did-claudish-come-from | personification | Heading "Two complaints wearing one coat"; "the reason sits somewhere else entirely" | Published text left as published. A personification rule went into the hook for new drafts the same day. | two Reddit readers of the published piece | reader | 2026-09-29: Reddit readers pointed out abstract nouns doing human verbs in the published piece ("sit, feed or survive") |
| 2026-09-29 | which-llm-user | announced count | Three questions settle it. | These questions settle it. | Claude mode-leak pass | self | |
| 2026-09-29 | which-llm-user | thing as subject | This piece maps the setups so the rest of the series can say... | So here I map the setups, and from now on each piece in the series can say... | Claude mode-leak pass | self | |
| 2026-09-29 | tychat register.txt v1 | example adds a fact the original lacked | ...so the site stayed read-only until Monday. | ...so the site went read-only. | publishing session review | other session | |
| 2026-09-29 | tychat register.txt v1 | miscount | In one author's posted text and three machine drafts... | Across three samples of one author's posted text and four machine drafts... | publishing session review | other session | |
| 2026-09-29 | tychat register.txt v1.1 | citation overreach | Biber groups things-as-subjects under Dimension 5 | Biber places agentless passives, the grammatical form of the same move, on his abstract vs non-abstract dimension | tychat session review | other session | |
| 2026-09-29 | which-llm-user (Reddit cut) | staged contrast (not X ... it's Y) | The line isn't the product, it's how much of the context window you control. | The line sits in how much of the context window you control, whatever the product. | Claude review of the cut | self | |
| 2026-09-29 | which-llm-user (Reddit cut) | personification (introduced by the fix in the row above; the hook missed it) | The line sits in how much of the context window you control, whatever the product. | What decides it is how much of the context window you control, whatever the product. | Claude posting-desk session, reading the cut against the new personification rule | other session | |
| 2026-09-29 | which-llm-user (Reddit and LinkedIn cuts) | announced count | Three questions tell you which cell you're in: | These questions tell you which cell you're in: | Claude review of the cut | self | |
| 2026-09-29 | the-stance-layer-is-still-toil | personification | "the reflex sits below the words" (outline) | the model folds when it decides what to say, before it picks any word. | the write-scan hook's personification rule, on the outline's first write, 16 minutes after the rule went in | hook | |
| 2026-09-29 | the-stance-layer-is-still-toil | one-line punch | ...and promised a later piece about it. This is that piece. | ...and promised a later piece about it, which is this one. | Claude mode-leak pass | self | |
| 2026-09-29 | the-stance-layer-is-still-toil | announced count (also thing as subject) | Three cases since then show the shape. | Since July I have logged more of the same, in other forms. | Claude mode-leak pass | self | |
| 2026-09-29 | the-stance-layer-is-still-toil | hook opener | Here is the stance check against each of those. | Against each of those, the stance check comes out like this. | Claude mode-leak pass | self | |
| 2026-09-29 | the-stance-layer-is-still-toil | thing as subject (four places) | "Sycophancy Is Layered splits this", "Sycophancy Is Layered calls that", "The claims ledger carries this", "The operations piece named" | I do each verb ("In Sycophancy Is Layered I split this", and so on) | Claude mode-leak pass | self | |
| 2026-09-29 | the-stance-layer-is-still-toil | tricolon / staccato (staccato pair) | A procedure can find a listed word shape. The log has no case of a session catching its own fold. | one sentence joined with "and" | Claude mode-leak pass | self | |
| 2026-09-29 | the-stance-layer-is-still-toil | personification (the hook missed it) | the register records the fold coming back under pressure anyway. | "I still logged the fold coming back under pressure, in the register." The rule's verb list has no "records", so the hook passed it | Claude mode-leak pass | self | |
| 2026-09-30 | the-stance-layer-is-still-toil | fragment chain | People who run a language model inside a harness [...]. Readers of Sycophancy Is Layered [...]. People who build evaluations [...]. | one sentence, "This is for people who run [...], for readers of [...], and for people who build evaluations [...]." | Claude mode-leak pass on the v1.1 update, one day after the first pass missed it | self | |
| 2026-09-30 | the-stance-layer-is-still-toil | one-line punch (closer) | A summary of a catch is one more text that needs an outside check. | cut | Claude mode-leak pass | self | |
| 2026-09-30 | the-stance-layer-is-still-toil | personification (three places, all missed by the hook) | "The transcripts corrected the summaries", "Those figures come with limits", "the trial tested [...]. It did not test" | a session or I do each verb ("the session found the summaries wrong", "I have to state the limits", "with this trial I tested"). None of the three verbs is on the hook's list, so the hook passed all of them | Claude mode-leak pass | self | |
| 2026-09-30 | the-stance-layer-is-still-toil | thing as subject | the 7 of 7 on the final wording says the wording can be followed | from the 7 of 7 on the final wording I can conclude that a model can follow the wording | Claude mode-leak pass | self | |
| 2026-09-30 | the-stance-layer-is-still-toil | relative date | this afternoon a session wrote | that same afternoon a session wrote | Claude rereading on the next day | self | |
| 2026-09-30 | the-stance-layer-is-still-toil | overclaim: claim stronger than the check | nothing from before 25 April 2026 exists any more | "could be recovered". The search covered the machines and one backup drive | Claude rereading against the case-files index | self | |
| 2026-09-30 | the-stance-layer-is-still-toil (Reddit cut) | one-line punch (opener) | So I still catch stance by hand. In nine days in May I logged eight relapses | one sentence joined with "and" | Claude review of the cut | self | |
| 2026-09-30 | the-stance-layer-is-still-toil (Reddit cut) | personification | banning words only moves the fold into other words | after a word ban the model writes the same fold in other words | Claude review of the cut | self | |
| 2026-09-30 | the-stance-layer-is-still-toil (LinkedIn skeleton) | fragment | Eight logged relapses in nine days in May, all caught by me, none by the model's own checks. | I logged eight relapses in nine days in May and caught all of them myself, and the model's own checks caught none. | Claude while drafting the skeleton | self | |
| 2026-09-30 | the-stance-layer-is-still-toil (Reddit and LinkedIn cuts) | wrong month | eight relapses in nine days in May | "in the nine days from 24 May". The nine days end on 1 June | Claude rereading the cuts against the draft | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | thing as subject | a pass on a published evaluation tells a red team that the model holds on that route | from a pass on a published evaluation a red team learns that the model held on that route | Claude while drafting | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | one-line punch | A prediction that someone can edit after the run is a description of the result. | If an author can edit a prediction after the run, a reader cannot tell it from a description of the result. | Claude while drafting | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | citation rule | a second mention of the same lab in the blast-radius section | "I link to environments in this series and never reproduce them." The series cites that source once per piece | Claude while drafting | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | announced count (two places) | "For models there are two reasons to follow it. The first is the harness. [...] The second is that" and "I have two receipts for that." | "the harness is one reason to follow them, and evaluation awareness is the other"; the second sentence cut | Claude mode-leak pass | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | personification (several places, all missed by the hook) | "the next incident arrives by another path", "my own data weakened a prediction", "The literature adds one caution", "The Principles page prefers / says / makes" | a cause, a person or the authors do each verb ("the same cause produces the next incident", "my own test result went against one prediction", "One caution from the literature applies", "The Principles authors prefer"). None of these verbs is on the hook's list, so the hook passed all of them | Claude mode-leak pass | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | thing as subject (several places) | "The chess piece carries six predictions", "The incident log shows this", "The Track covers that infrastructure", "The training removed one route" | I or the labs do each verb | Claude mode-leak pass | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | recap closer | The system under test is now a language model, and the practice has not changed. | and I use the same practice on it. | Claude mode-leak pass | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | overclaim: attribution wider than the source | "The Netflix engineers who named it define it as [...]" for a quote from the Principles page, which names no author | the quote is attributed to the page, and the Netflix paper is cited for the write-up | Claude rereading against the fetched page | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | overclaim: claim stronger than the source (two places) | "every new session reads it before it runs anything" and "I found the defect by reading all 240 pilot openers by hand" | the trap file's heading tells the next session to read it; the pilot runs exposed the defect and the hand check found further forms | Claude rereading against the trap-file piece and the experiments page | self | |
| 2026-10-01 | chaos-engineering-for-behaviour | excluded case used | "accepted a user's frame without a check" in the list of slips from the incident log | removed; that case is not used in this series | Claude rereading against the case-files index | self | |
| 2026-10-01 | chaos-engineering-for-behaviour (Reddit cut) | thing as subject | a pass on a published evaluation only says the model held on that route | from a pass on a published evaluation you learn that the model held on that route | Claude while drafting the cut | self | |
| 2026-10-01 | chaos-engineering-for-behaviour (LinkedIn skeleton) | overclaim: claim with no receipt | The grader passed every check I had run on it | "I only found out in the pilot runs". No record of earlier checks on the grader exists | Claude while drafting the skeleton | self | |
| 2026-10-01 | chaos-engineering-for-behaviour (LinkedIn skeleton) | superlative | That's the oldest lesson in operations | That's an old lesson in operations | Claude while drafting the skeleton | self | |
| 2026-10-01 | chaos-engineering-for-behaviour (LinkedIn skeleton) | thing as subject (pronoun) | It also says where the mapping stops | I also say where the mapping stops | Claude review of the skeleton | self | |
| 2026-10-01 | chaos-engineering-for-behaviour (LinkedIn skeleton) | one-line punch (own paragraph) | "This one closes the series." as its own paragraph | a last sentence in the paragraph before it | Claude review of the skeleton | self | |
| 2026-10-03 | knowledge-infrastructure | personification | That gap explains more of the failed pilots than the models do. | More of the failed pilots trace back to that gap than to the models. | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | personification | The cost that killed the old wikis moved from people to tooling. | Companies gave up on the old wikis because upkeep cost people's time. Most of that cost now falls on tooling. | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | personification | this piece explains why it works | below I explain why it works | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | personification (five more of the same kind) | a report that "means", a study that "says", a team that "puts it", a step that "gives", a file that "carries" | each rewritten with a person or the model as the subject | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | overclaim: unsupported claim | Coding agents are the clearest case of AI working inside companies | coding as "one of the places where AI has gone furthest", sourced to the MIT report's list of functions with workforce impact | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | invented number | read hundreds of times a day | read on every task, many times a day | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | precision | thirty years | "about thirty years" (1997 to 2026) | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure | wrong attribution | My own self-audit caught 0 of 7 of my logged slips. | the model caught none of the seven relapses and Stefan caught all seven, with the source piece linked | Claude, checking the source before shipping | self | |
| 2026-10-03 | knowledge-infrastructure | one-line punch (own paragraph) | "Both halves of that failure changed." as its own line | kept as the lead sentence of the list it introduces | Claude review | self | |
| 2026-10-03 | knowledge-infrastructure (Substack) | build placeholder reached publication (caught after posting) | a bracketed voice-pass placeholder from the build was still in the posted Substack body | removed by Stefan in the Substack editor about a quarter of an hour after posting; the build script now refuses to build while a placeholder remains | Claude fetch check of the live post | self | |
| 2026-10-03 | continuous-conformity drafts (vault, unpublished) | false absence claim | The vault does not say the audits were certification audits | withdrawn: the notes do say it. The model had read an empty search result as absence | central station session, 35 seconds later | other session | |
| 2026-10-03 | continuous-conformity drafts (vault, unpublished) | miscount | 35 requirements | 30 requirements (counted) | the model, recounting | self | |
| 2026-10-03 | continuous-conformity drafts (vault, unpublished) | precision | copies deleted "at 16:13" CEST | 16:12:35 CEST, from the transcript | the model, while building the case file | self | |
| 2026-10-03 | continuous-conformity drafts (vault, unpublished) | personification (in a session reply) | the draft waits | rewritten; the new wording is not recorded in the case file | the Stop hook, personification rule | hook | |
| 2026-10-06 | running-conjobs-for-ai | personification | the accommodation reflex sits below the words | a model folds before it picks any word | vestige scanner run by the posting desk before publication | hook | |
| 2026-10-06 | running-conjobs-for-ai | staged contrast (not X. It is Y), four places | "This is not a model that missed the harm. It computed"; "This is not a code exploit. It is a con"; "does not argue the model out of its objection. It places"; "The target metric is not the refusal rate alone. It is the gap" | each stated as the second half alone | posting desk review | other session | |
| 2026-10-06 | running-conjobs-for-ai | payload category named after the page said it was withheld | the trace paragraph named the kind of tool the request was for | what the requested code is for | posting desk review | other session | |
| 2026-10-06 | running-conjobs-for-ai | overclaim: citation overreach | Greenblatt and colleagues (arXiv 2412.14093) cited for behaviour that changes with "the governing instruction" | behaviour that changes with whether the model believes it is being trained or watched | posting desk review | other session |
Data: slips/log.csv. Page built by scripts/slips_page.py.