{"path":"research/the-kind-of-mistake-we-make/corrections.md","content":"# Corrections in the record\n\nEvery claim this project's sessions wrote down and later found wrong, as the session record marked it with ⚠, one row each: when, which shape of mistake, and what was wrong. Nothing else lives here. The shapes and what they cost are read in [the kind of mistake we make](../the-kind-of-mistake-we-make.md) § 2; this file is its data.\n\nThe rows up to 2026-11-15 were moved here from the session history, which was then retired (his word: *\"I wanna delete it once we extract anything valuable/reviewable out of it\"*; a file of only corrections, *\"a file of ONLY those corrections and nothing else would earn its keep\"*). The dates of the Oct 5–17 rows are the span of one long session entry, not each finding's day. Paths to private files are left out.\n\n**From now on**: a claim found wrong after it was written gets one row here (date, shape, what was wrong and where it was fixed), besides fixing every place that repeats it. The monthly check-in sorts the new rows and names one rule change. The founder's own corrections in chat are graded separately (`the-canon-distillation-pass/corrections-graded.jsonl`).\n\n| When | Shape | What was wrong |\n|---|---|---|\n| 2026-03 to | a whole from a part | corrected 2026-10-10: recorded here and in the run's own headline as *stated on one side and IMPLICIT on the other*, which fits the second debate only; in the first, neither side wrote it down); register drives residue type (worldview→structural, evidentiary→mostly-dissolving; neither batch produced fittingness); two residue predictions registered; two new weighing dialects (rank-ordering \"priority above\", superlativ… |\n| 2026-03 to | an inference stated as observed | Corrected 2026-08-14 on founder challenge: these are two pairs, NOT a before/after on one — `t18 × e71` was never classified, so \"omit the premise and the pivot enters as an agreement\" is an inference from the similarity ranking, not a result; and the graph was checked, no SUPPORTS edge exists (only SIMILAR_TO), the pivot is correctly opposed, nothing inconsistent is in the corpus. The measured claim is that the clas… |\n| 2026-03 to | two things held as one | corrected 2026-10-20: also the default on every link extraction wrote, `docs/research/the-kind-of-mistake-we-make.md` § 9); propose-then-ratify has a propose half built five times over and a ratify half built once. Two canon lines ratified — the conviction that *what is invisible from inside a position is visible from outside it, so minds make each other legible* (with the founder's correction that positionality is t… |\n| 2026-09-06 | whose position, and whether ruled | The first plan omitted the research-feeding step the founder had asked for on 2026-08-25; he caught it. Then the constellation review under the briefing-stack protocol: six tiles ruled in one evening (two-axis, separability floor replacing contest-gating, Listening, no-gap kept with frame gaps given a contribution path, the group premise re-cut to its conditions after an evidence pass — `the-group-premise-evidence.md… |\n| 2026-09-06 | a whole from a part | Two corrections propagated: the frame-gap claim *you cannot make it feel closable* narrowed on every surface (ledger K2–K6), and the collective-intelligence factor caveated as weakly replicated. A committed merge-conflict block was found in CLAUDE.md's topic index and resolved, with a test pinning it. |\n| 2026-09-11 | an inference stated as observed | Corrected the same day: the 74 blank-scheme edges (2.1%) were mostly the earlier model declining to choose, not taxonomy misses — the true no-fit rate is unknown until a fresh source runs. Single exit proposed, unruled. |\n| 2026-09-19 | code or process | The first version blinked on hover: an entrance animation was cancelled by a dimming class and so replayed every time the dimming lifted. Hover now moves one box's lines and nothing else (`specs/landing-redesign/tasks.md` item 14). |\n| 2026-09-19 | an inference stated as observed | Three docs had claimed the map and the star texts are present *without JavaScript*; false, the site is client-rendered, and corrected the same night. |\n| 2026-09-19 | a whole from a part | The finished onboarding board broke a reading made from the unfinished one: a requirement to be recommended is a third kind of barrier (who you already know), beside effort and information. |\n| 2026-09-19 | two things held as one | The first draft claimed a look-back *meets* the structuring-gradient principle's condition; it does not (a finished map handed back is what that principle says people argue with as a document), so it illustrates and never tests. The sheet's phone numbers were written nowhere, since the private notes sync to a remote; a git-ignored `local-only/` folder now exists there for material that must stay on the machine. |\n| 2026-09-20 | absence from a read too narrow | The research that already covered this was not read before building. `docs/research/game-feel-and-the-first-screen.md` (2026-08-28) had concluded that slicing, not zooming, answers the hairball, from the founder's own sketch of a row of argument names under the picture. |\n| 2026-09-20 | an inference stated as observed | and a correction of something said to him the turn before: the system does not already compute how much a claim hangs on a supporter or challenger, only on its parts and on open questions. He asked for the counts bar to move down above the list, for a way to record who wrote a source (added; left empty for the friend until he has seen the rendering), and for the *source assembles itself as a map while it is read* ide… |\n| 2026-09-22 | absence from a read too narrow | I first diagnosed the verdicts as wrong, from five truncated reasons and without reading the prompt; the founder asked *\"Are you certain that it's wrong?\"*, and the prompt defines `overlaps` as evidential bearing — which is what a support link is. The verdicts were right and the routing dropped them. |\n| 2026-09-22 | a check that could not see | Two of my figures were wrong on the way: the stop rate was 33%, not 13%, and a 128-of-140 agreement figure had been computed over stored verdicts that structurally could not contain this disagreement. |\n| 2026-09-22 | code or process | `request.session` asserts without SessionMiddleware and `hasattr` does not catch it: guarding it that way turned 205 tests red at once. |\n| 2026-09-22 | two things held as one | I made a category error mid-fix — lowering normative claims onto the empirical ladder — and reverted it. The ontology gap it exposed: none of the six epistemic statuses can say *this is open*. |\n| 2026-09-22 | absence from a read too narrow | I briefly reported the keeper's `same` reasoning as unstored — two queries against the wrong field; it is stored as `merge_reasoning`. What genuinely was not stored was *which reader decided*, now fixed. |\n| 2026-09-22 | a count or condition wrong | Corrected 2026-10-03, Session 46: the count divides bytes by four and undercounts on current models, and the ceiling binds only a helper on a 200k model.)* |\n| 2026-09-27 | two things held as one | The census's *miscast as evidence* label was read by hand before acting on it, and it does not mean what the plan assumed. Of 31 machine-touchable edges, about 13 were support from child to parent ( 2026-10-07: *child* here names the premise by its role; as stored, these links ran premise first, the reverse of a part-of link: `auto-connect-upgrade.md` § 2c); the rest were reversed, reported speech, the 2026-09-13 ref… |\n| 2026-09-27 | two things held as one | 2026-10-07: *child* here names the premise by its role; as stored, these links ran premise first, the reverse of a part-of link: `auto-connect-upgrade.md` § 2c); the rest were reversed, reported speech, the 2026-09-13 referent mismatch, or not support at all. The 29 `auto_connect` part links were retired, none re-filed. |\n| 2026-09-28 | an inference stated as observed | One claim in the day's synthesis was half wrong, found the same evening. The published model the strength layer follows (Potyka, KR 2018) feeds each supporter's own strength but has no link weights. |\n| 2026-09-29 | a whole from a part | `fractal-priors-for-success.md` § 9's *\"visibly mid-transition\"* was narrowed to the information axis. |\n| 2026-09-29 | two things held as one | Two of the morning's proposals were superseded the same day on his question. S2 (*\"a machine may check what a machine wrote, but may not build on it until a person has engaged with it\"*) and S5 (a daemon's proposal rate capped by review capacity) rationed the wrong quantity. |\n| 2026-09-29 | an inference stated as observed | The first reading's claim that the judge's 62% was a definitional artifact was narrowed the same hour: it is roughly the problem-free rate. A seventh split, on `claim_aig_goal`, had been erased by the map's re-store on 2026-09-22. |\n| 2026-09-30 | two things held as one | A correction to what was recorded earlier the same day. *Starts by kind and origin* conflated three things. |\n| 2026-09-30 | two things held as one | It corrected § 6.0 itself: nested kinds pool into one start, and only crossing kinds without overlap data stay a range. |\n| 2026-10-01 | a whole from a part | *\"No public route serves one\"*, as first written here, was |\n| 2026-10-03 | code or process | The first sweep of the live site degraded it. It asked about ten pages a second for about fourteen minutes. |\n| 2026-10-03 | an inference stated as observed | GitHub still serves a deleted branch's commits by id, so the handoff's claim that nothing of them would remain was too strong; the private visibility doc records it beside its standing rule that a public repository needs a new history. |\n| 2026-10-03 | a count or condition wrong | The budget for always-loaded memory was measured, and the 200k ceiling turned out to be the wrong limit. Claude had told him a new line needed something evicted: *184k tokens against a 200k ceiling*. |\n| 2026-10-03 | a check that could not see | A September finding fell with it: *our own classifier commits on the undecidable two times in five* rested on five items written for the probe, of which three had a defensible answer. Homes: `jev-and-the-judgement-layer.md` § 3f and § 6e, ledger W6. |\n| 2026-10-03 | code or process | All 47 merge-candidate marks on the live graph are false: the linker read a required field's placeholder, `same`, before reading its own gate, and the test for a failed gate used a placeholder nobody had observed. Code fixed and tested, and all 47 marks taken back the same evening on his word. |\n| 2026-10-03 | a check that could not see | The first measurements were taken on the development server, which adds its styles late and reports shifts that are not there; the figures in the spec are from a production build. |\n| Oct 5–17 | a count or condition wrong | Two corrections to that morning's own work. The to-do list said for half a day that the sitting was drawn fresh on each request and held 51 parts; it served 37, built once per server process. |\n| Oct 5–17 | a whole from a part | A third correction that day: the to-do list said a deploy from the session *is refused by the permission system* and so is always his to run. It had been refused the evening before; it ran after he asked for it in words, and ran again that afternoon without a new instruction, so nobody here knows the rule it applies. |\n| Oct 5–17 | an inference stated as observed | Claude first reported that the pages had no way back on a phone; they did, from the second item on. The sameness batch came back identical after the deploy: the same 120 pairs in the same order. |\n| Oct 5–17 | code or process | One fault caught before the file went out. The first assembly of the eighth walkthrough put one pause a word late and cut five others outside a real gap. |\n| Oct 5–17 | a count or condition wrong | Two facts found while writing the evening down, both from running the check instead of recalling. The research doc said the parts sitting held *about sixty* links: that is the general review (60), and the batch he is asked to sit holds 51, under the sixty the guides give as the floor below which nothing can be concluded. |\n| Oct 5–17 | code or process | The three headline checks were run after he said yes, where they belong before the sentence is put to him. The first one failed: *measures* forbids counting by code, which this project does and wants. |\n| Oct 5–17 | a whole from a part | the method doc's *nothing slows the voice* was too wide: `[slowly]` slows a sentence by about a fifth. The first sample put an inhale *before* three seconds of silence, so the renderer learned that a breath has a side, and the checker rejects the old order (46 tests, the old wrong file kept as the negative control). |\n| Oct 5–17 | a check that could not see | A first hand-run of that count read the wrong property names and reported zero for everything, with no error; the test now builds each case a filter exists for. The first draft said the friend's assistant had read sixteen of our documents; its own list holds twenty pages. |\n| Oct 5–17 | whose position, and whether ruled | The first draft said the friend's assistant had read sixteen of our documents; its own list holds twenty pages. It also said that retiring the old link flag was a proposal, and that is ruled (Q32). |\n| Oct 5–17 | a whole from a part | A first figure for a sameness reply came from one call (76 tokens); the 147 reasons stored on links run 700 to 1,000 characters, and the estimate was redone on those. Two findings beside the storage: the same question asked twice can become a lookup, and about 64 of the graph's 84 MB are embeddings kept as plain number lists, where the engine's packed form takes about 16 (measured on a throwaway engine). |\n| Oct 5–17 | an inference stated as observed | A worked example first written for the minimum-wage pair gave the two articles roles the record does not show; the record says both report the proposition and neither asserts it, and the example was rewritten from that. `scripts/claim_wording_sources.py`, 7 tests; ledger AA3 sharpened, AA6 added (355 rows, 275 open). |\n| Oct 5–17 | code or process | Tracing the deploy before running it found a fault no test had: the production container's user has no home directory, the record kept its local file of unsent rows there, and on the live site every model call would have logged an error with a traceback and lost its row number. The test for that case was written first and failed; the look for kept rows can no longer raise, the service says once at its start when the … |\n| Oct 5–17 | an inference stated as observed | And this session had the running image wrong. It had said nothing was deployed since the evening of 5 October. |\n| Oct 5–17 | code or process | Three of fourteen cut requests came back as the word *test*, and one said why: *\"Test minimal case to isolate schema issue before resubmitting full answer.\"* The placeholder of 29 September, cause then unknown, is the model's own probe after the runner refused its full answer. The engine's guard had missed all three shapes and the splitter's own check caught them; the guard now covers them, tested on the three real a… |\n| Oct 5–17 | two things held as one | The engine named an alias, not a model: its default was `sonnet`, which the runner resolves to whatever it ships. It still reached `claude-sonnet-5` that day, and the default now names that model. |\n| Oct 5–17 | code or process | The first deploy still showed a replaced part as the next thing to deepen: eight readers counted every part-of link, superseded or not, and had since the first links were retired on 27 September. Found by looking at the live page, repaired the same hour, and a test now reads the code for any reader that forgets. |\n| Oct 5–17 | two things held as one | The linking step had written every part-of link backwards since April; found and repaired 2026-10-07 (founder, asked whether to repair the step before another text is mapped through it: *\"Repair it sure!\"*). Auto-connect offered *this claim is a premise of that one* as an answer and stored it as a part-of link from the premise to the conclusion, the reverse of every other part-of link: 261 links, 60% of every live pa… |\n| Oct 5–17 | two things held as one | a report that said *linking was $70* without those words made him ask: *\"Wait wait... You spent $70, or you're projecting it? |\n| Oct 5–17 | a count or condition wrong | this corrects a figure of *about 46% cheaper* that had stood in the code since 2026-09-22, counted in calls on a run where 13% of pairs stopped; that night 45% did); one call a hub is eight times cheaper and keeps 14 of 29 attacks; one reading a rendering found 4 of 8 stored contradictions and over-read 4 others. A cap by similarity is the only thing that held up: twelve partners a stored claim reads half the pairs a… |\n| Oct 5–17 | two things held as one | The comment corrected three wordings in the note (that what a word refers to *can be announced*, that the address *can be occupied*, that a reader *owns* a sense) and the wording of the fourth proposal, which now asks whether a key word is one *that readers already bring a meaning of their own to*; that rider stands in the three headline tests (`CLAUDE.md`) and in the constellation spec. His example, looked up the sa… |\n| Oct 5–17 | code or process | The second trial was started with a time limit shorter than its work and was stopped before writing its answers; it was rebuilt from the record of machine judgments, replies and inputs both, and only the 29 missing checks were made again. 179 calls on the plan, $12.96 at list prices (0.6 points, 3 kr of the plan fee). |\n| Oct 5–17 | whose position, and whether ruled | One sentence of Claude's from 2026-03-29, *never generating premises, only finding real ones*, had narrowed his *finding or creating* and stood on five surfaces as the settled design; all five are corrected. The paper he pointed at checks where a chain ends and not its steps; a link gives both ends, so the middle can be checked. |\n| Oct 5–17 | a count or condition wrong | corrected to 17, 5, 2 unsettled and 6: the pruning tried one order, and link 19 was found unsettled on 2026-10-08); the writer never declines and writes three times what a reader needs; and six of the seven links read as not a reason come out clean, so the checks sort links by structure and not good reasons from bad. Unasked: about 512 model calls on his plan ($27 at list prices: 1.2 points, 6 kr of the plan fee; the weekly figure went from… |\n| Oct 5–17 | two things held as one | Two things reported that evening were corrected. His squirrel of 2011 is a step from a sign to its cause, and the trial's blind reader, which turns out to read deductively, refused it in all four askings and accepted it with the rule turned round (6 calls). |\n| Oct 5–17 | a count or condition wrong | 18, 5, 1 and 6 stood here for a night: link 19's recheck contradicted the pruning's own reading of it). A second way of writing a link out, by counterexample (Lakatos), is proposed and untested. |\n| Oct 5–17 | a count or condition wrong | link 19 of the thirty is unsettled, so the count is 17, 5, 2 and 6, corrected in eight files; and the counterexample for link 20 is the first text's own (imprisonment and kidnapping), which the philosophy note's § 10 now gives. AD22 to AD29 carry a picture for a name in the ledger. |\n| Oct 5–17 | absence from a read too narrow | Two sentences in the philosophy note of the night before were wrong against the September review and one was unspecified; all three are marked in place. No new ledger rows; AD22, AD23 and Q22 carry the refinement. |\n| Oct 5–17 | whose position, and whether ruled | Corrected in it: a first answer from memory credited Hofstadter with the induction step, and the daemon note called the four checks *ruled* when their home says unratified. |\n| Oct 5–17 | code or process | The thirteenth listening piece was delivered before the check and its eighth part overstates. His comment while listening (*\"Agreeing on rules comes late, yes. |\n| Oct 5–17 | a check that could not see | Found 2026-10-09: that search could match by meaning on only about half of this project's notes, and not at all on the note holding the ruling, since embedding had been switched off since 5 September. The figure measures keyword search for the newest notes (the keeper's dotfiles note, under *What was measured*). |\n| Oct 5–17 | absence from a read too narrow | Two things the first pass said were wrong: that no study of a tuned pointer exists, and that a tuned open model can be served per token. The result is in the working note outside this repository ((a private file)): a tuned oracle is possible and is the expensive way to a weaker reader that nobody has tested on this failure; reading the whole corpus for each question with cheap models, and judging the candidates, need… |\n| Oct 5–17 | a whole from a part | the statement that a debate's crux is *stated on one side and implicit on the other*, which run 3F's own text contradicts for its first debate and which stood in 22 places, now reads *left unsaid by at least one side*; and a price note that called a whole extraction on Claude Sonnet untried. Home: `docs/research/corpus-graph/README.md` § 9. |\n| Oct 5–17 | whose position, and whether ruled | The September note had filed him as declining the hive mind. Corrected on every surface, S3 withdrawn, S8 proposed, read by the outside reader, and ruled the same evening (*\"Yes!\"*: members give up having to work everything out alone, and keep their own position, the ability to check, and the freedom to leave; `major-transitions-heredity-and-evolvability.md` § 5b). |\n| Oct 5–17 | whose position, and whether ruled | corrected the same evening from *five run*: asking a second way runs on no live task and he never ruled it, *\"'Asked twice'? I don't think I ever blessed that particular part.\"*), and no person has yet answered a sitting, written a note or touched a machine-cut part, so no task has a measured error rate. |\n| Oct 5–17 | an inference stated as observed | One of Claude's own claims was read off the code and then refuted by running it: the stop rule of 25 August does stop on fresh material, about half the time. Three readers' reports kept word for word, 23 connections kept in the corpus graph, and its forward test counted for the first time (the graph had named 33 of the 77 notes the readers cited). |\n| Oct 5–17 | absence from a read too narrow | It corrected what he had been told an hour earlier: four things sit under a link today, not one. The linking step's own reason is a text field under 1,481 of 1,636 support and attack links between claims from texts or people; a machine's reading of relevance is a number under 190; a standing question is a node; and the arrow can hold the id of a claim about it, a slot built on 28 September with the two-factor ruling … |\n| Oct 5–17 | absence from a read too narrow | The first view (the 150 best-connected nodes under one force layout) was built against this project's own March summary (*do not show the full graph by default*) and against what the map view decided for readers. Three plain faults were fixed; the design is his by Lins's founding notes, so it is researched (three helpers, web only) and proposed in Lins's `docs/`, not built. |\n| Oct 5–17 | a check that could not see | The first test of that could not fail, since its claims sat a hair above neutral and nothing was examined before or after; it grants standing directly now and has two controls. And every reader that picks claims out of the graph had to decide about the kind: 257 queries read claims, thirteen modules were changed, eight readers read every claim and are listed with their reason, and a test reads the code for the next o… |\n| Oct 5–17 | two things held as one | The reply before the ruling had read that as a question about putting a Rust canvas into Deliberus's pages, which he said he would probably never do; the notes carry the correction. Deliberus's own maps were read for it: browser-drawn SVG, in the same framework, each box a real link. |\n| Oct 5–17 | absence from a read too narrow | The first report of it said seven inconsistencies on pages and 58 pairs had been filed nowhere; five of the seven were already on his list, because the claim had been checked against the index and not against the to-do file. |\n| Oct 5–17 | two things held as one | Two words had misled him and are corrected where they are defined: *public page* (*\"Only 24? Aren't basically all docs publically available?\"*; every note under `docs/` is served, and the word means a page written for visitors) and *move*. |\n| Oct 5–17 | a whole from a part | U17 had said *not of reviewing outputs*; it now says *not of reviewing every output*, since people still read to find the kinds of fault. On the proposal never to write *proven* without its bound and scope he wrote *\"Sounds rational\"*, and on 2026-10-11, asked whether to make it resident, *\"Yes\"*: ledger T8, ruled, and a sentence added to the principle on reading before proving. |\n| Oct 5–17 | a count or condition wrong | The cost given him then, *a fraction of a cent per thousand pairs*, was about ten times too low; corrected 2026-10-11 to about seven cents, and the same slip in W6. |\n| Oct 5–17 | a check that could not see | The review told him that the merged pull request had brought in 272 duplicate commits under a second root. It had not: the cloud checkout was shallow, and unshallowed the history has one root and the request added one commit. |\n| Oct 5–17 | absence from a read too narrow | Corrected 2026-10-12: it had been saved on 2026-09-29 as `docs/research/an-outside-reading-of-the-thesis.md`. The first search returned the line that points there, cut at 260 characters, with the pointer after the cut.)* Home: `docs/research/what-deliberus-is-for.md`; ledger AH1 to AH3; the outside reader's findings are on the rows and in the note's § 8.7. |\n| Oct 5–17 | code or process | Two readers had found that the address check and the fetch each looked a host up; repaired the same day (`deliberus/safe_fetch.py`). The readers' reports filed that morning had themselves quoted the machine detail into the served folder; taken out with the split. |\n| Oct 5–17 | code or process | The readers' reports filed that morning had themselves quoted the machine detail into the served folder; taken out with the split. |\n| Oct 5–17 | two things held as one | Corrected with it: an earlier note said Ashby was in the corpus once; that *Ashby* is another researcher. |\n| Oct 5–17 | two things held as one | Two things the session had wrong on the way, corrected before they travelled: the fifth family was first described as facts gone out of date, where most of its findings were never true; and the first view as first proposed left out the reviews he had ordered first. Built the same night after a sweep of the documents he asked for first: 2,315 boxes and 4,386 lines in a graph of its own, opening on 44. |\n| Oct 5–17 | a count or condition wrong | The family counts ruled that evening had been taken over every line of the findings file (790); the copy holds the 728 the tool shows, so each family reads a little smaller there, and four kinds used once are left unplaced where a hand count had placed them. Home: `docs/research/corpus-graph/README.md` § 21 and § 22 (what was built, § 22.12). |\n| Oct 5–17 | absence from a read too narrow | The session had reported that paragraph missing the day before, and it was there: its first search returned the line that points at the note, cut at 260 characters, the pointer after the cut (*\"Why wasn't it found?!\"*). A guard against cut lines, built the evening before, was silent: it checked the notes a command names, and a search over folders names none. |\n| Oct 5–17 | absence from a read too narrow | Corrected with it: a to-do line said a word search had found no *predicate*; it had searched the chat and not the mail. |\n| Oct 5–17 | absence from a read too narrow | Its first version had not read the ruling of 30 September on what a failed reason does, which is the corpus's own answer (a part is presumed needed, a reason is not until argued), and claimed a connection as new that the ruling had stated; corrected the same afternoon when he challenged the test: *\"If all conceivable arguments that are relevant falter/are refuted, then the central/mother claim must have a very low tr… |\n| Oct 5–17 | two things held as one | Corrected with it: ledger T9 had said the reasons are filed *never as parts*. They are parts of the sentence and reasons for its conclusion, and the splitter's gate (*would the claim be weaker?*) cannot tell the two apart. |\n| Oct 5–17 | an inference stated as observed | He leaned yes on letting a machine's confident *no* take away a default (*\"I'm leaning yes, but please think deeply about what this entails\"*), and a replay over everything people have added changed the advice the same evening: the reading of links would withhold 2 of 32, both honest objections to a step filed on a claim, and the part test would withhold 28 of 39 parts, 24 of them reasons their claim rests on, becaus… |\n| Oct 5–17 | two things held as one | A note that said the project sides with Prakken against Pollock was corrected. The canonical example redrawn on his word (*\"1b is obviously good to me\"*): Dr Smith's credentials are reasons and not parts, in the always-loaded file and the April note; the floor's third clause ruled (AG14). |\n| Oct 5–17 | two things held as one | A session's own muddle, min across reasons called the legal conjunction paradox, corrected on both surfaces that afternoon. The eighteenth listening piece, 46 minutes, the first with a chapter list for noting on the move, on his instruction that it be standard (*\"This should then become standard procedure for all audio walkthroughs.\"*, the method's rule 33). |\n| 2026-10-18 | whose position, and whether ruled | Correction: the twenty-first piece said *goes to a person first* and predicted he would want a sample read by a person; that was the session's stance and contradicts his standing position that the small tasks work without people. Its header says so; the audio was not re-rendered on his word. |\n| 2026-10-19 | an inference stated as observed | first written as *until the next sameness ruling*; the block renders a fixed list of five rows and three canon lines, so a ruling elsewhere changes nothing). Both filed unruled (`three-minds-and-the-leap.md` § 5a, `proving-a-pass-reliable.md` § 11). |\n| 2026-10-19 | an inference stated as observed | The case put to him was misdescribed: that pair's newer reading said one claim contains the other, so its link stays. |\n| 2026-10-19 | a count or condition wrong | Corrected the same evening: the ledger's running total said 497 numbered rows; its own grep counts 500, none twice. |\n| 2026-10-19 | whose position, and whether ruled | The same evening's doc had credited the gloss to the machine; corrected in § 6d. |\n| 2026-10-19 | whose position, and whether ruled | The session first said the cleanup waited on his word; it had been ruled three weeks earlier, never built, and new code had used the field since. |\n| 2026-10-22 | two things held as one | Two cells this session wrote into the rules table that day were wrong: an explanation of a single case does not *take nothing back* (it places the blame among the links of the reasoning and refiles the case under the rule it counts for), and a statistical rule is not moved *only through frequency* (every case moves it, either way). The same error stood in AT3's statistical clause and canon check (*one failure leaves … |\n| 2026-10-23 | whose position, and whether ruled | Two corrections: FOI's list of agencies names who should get central roles, not where the function goes (the map had cut it into seven parts); *KTH bör vara huvudman*, credited to a Cybercampus answer that is not among the 127, withdrawn. 250 claims, 334 links; 14 grounds cases for the unbuilt answering step. |\n| 2026-10-22 | code or process | A probe started with Ferrum's default options (a mock keychain) logged the scraper's Facebook profile out; he logs in again before the ninth post can be read ((a private file), TODO). Only 1 of 2,856 stored comments is marked a reply, unchecked against the live page (TODO). |\n| 2026-10-31 | absence from a read too narrow | Three things this session had wrong. It said nothing counts stranger visits; he answered *\"But I get ntfy pushes every day summarizing visits\"*, and `deliberus/notifications.py` tallies every non-owner, non-bot page view by path, so the signposts note's cell saying the same was corrected (what nothing does is follow a visitor from landing to map). |\n| 2026-11-08 | two things held as one | corrected 2026-11-09 on his question: the two share one claim, the civilization-level process as the unit, and differ in position; Carlsmith names no argument map, graph or debate tool; the reading's correction section); his first criterion, *\"good at the truth\"*, and his *\"truth-related collective epistemology in a good way … I'm hoping that we can get the AI to help us with that\"*, name the slot the missing layer c… |\n| 2026-11-09 | an inference stated as observed | the *never felt* withdrawn the same morning on his word, below), elegance as a goal Lowered, the substrate a fixed core the reviews are building. BL2 and BL3 proposed; the four products of a comparison placed (the fourth a claim, the proposed property a reading; the resemblance link's fourteen analogy properties the first census). |\n| 2026-11-09 | two things held as one | Corrected in it: `cases-and-generalizations-from-both-ends.md` § 2b had the alignable difference as *the first attack on the generalization*; `the-analogy-daemon.md` § 10 step 4 had it on the relevance claim, which stands as the default. |\n| 2026-11-09 | whose position, and whether ruled | This note is a session's, added 2026-10-07, and no ruling of the founder's (2026-10-10: *\"'Asked twice'? I don't think I ever blessed that particular part.\"*). |\n| 2026-11-15 | a check that could not see | The post-compaction hook was reported *\"live from the next compaction\"* on its build day; at the next compaction it handed back the 25 exchanges before the PREVIOUS compaction (this compaction's boundary is written after the hook runs) and only a 2 KB preview of them (hook output over 10,000 characters is cut, a limit already recorded in `~/dotfiles/docs/claude_code_steering_layers_2026-08-26.md`). Its tests had put the boundary at the end of the transcript, a fixture authored to fit. Fixed the same evening (`~/.claude/hooks/compact_recent_exchanges.py`). |\n| 2026-11-15 | a paraphrase that changed the scope of his words | C4's disputed merge was written as *counts as an error of the keeper*; his words were *\"kind of more of a bug-report\"*: looked into, the person possibly wrong. Fixed in the ledger, CLAUDE.md's Topic Index, the human-judgment note and the canon pass's queue. |\n| 2026-11-16 | a mechanism stated before its check | A slow query in the same-claim group reader was blamed on any `WHERE c.claim_id IN $ids` filter, and about twenty more were reported to him as slow and filed in TODO. A live plan showed a plain IN list uses the index; the cause was the other end's `NOT … IN` filter making the planner start there. TODO line removed; the query-plan test now catches the real shape. |\n| 2026-11-17 | one word doing two jobs | In chat a session called a link's relevance readings *strength readings* and wrote that relevance readings and priors would fill the institute map's cards. A card's fill is the claim's posterior, computed, shown only once something under it is examined; a reason still at its own prior pulls nothing (Q50), so priors alone fill nothing; 146 of 233 links already carried relevance readings. Fixed in the AI Safety Day prep note (private). |\n| 2026-11-21 | a synthesis that said more than its sources | § 8 of `what-a-comparison-produces.md`, written and committed the same night, said a difference that matters is *two claims at the floor*, drew the resilience pair's contrast as already attacking the transfer with the role *the carrying*, and called the transfer *the* unit of analogy. Codex's reflection, which he asked for, showed each goes past BL3, BL4, AD27 and AG14: the rulings ask for one explicit inference, whose home is the attack link's own relevance claim; the stored likeness names no role; a comparison can propose nothing. Fixed in § 8 (a dated paragraph and three inline marks) and TODO's analogy loop; said to him in chat the same night. |\n"}