{"path":"research/session27-the-canon-pass-and-the-corpus-blind-spot.md","content":"# Session 27 — the canon pass, and the blind spot that opened it\n\n**Date**: 2026-08-28 → 29 · **Type**: session record, kept verbatim where the founder's words are the\ndecision · **Companion**: [the-canon-distillation-pass.md](the-canon-distillation-pass.md) (the method\nand the open register this session opened)\n\n> ⚠ **This doc is the session record. It is NOT a resident surface.** Per the shape guard added this\n> session (`tests/test_claude_md_shape.py`), no per-session narrative goes into `CLAUDE.md`. Durable\n> findings from here are extracted into topic-shaped resident lines under the wrong-by-default filter,\n> one at a time, with the founder as arbiter.\n\n---\n\n## 0. What the session turned out to be about\n\nIt began as a mapping exercise (which unruled claims touch which constellation tiles) and became the\nthing that reframed the whole review programme: **Claude repeatedly asserted the opposite of findings\nthe corpus already held, with the corpus's own pointers loaded.** Four times, on the record, in one\nsession. The founder's response is the session's organising fact:\n\n> *\"Sigh. I feel like all of this is based on an incomplete reading of the corpus. Lots of even\n> recently created docs have touched on important aspects here. Don't do a stupid grep here. Read up\n> really broadly and deeply on anything relevant please. Ground extremely deeply.\"*\n\nand then:\n\n> *\"I bet you'd make some 'surprising' 'discoveries' if you read more of the corpus. All of this is\n> NOT a surprise to me, I've been over this a thousand times in different sessions. Every session that\n> I ask you to read up on something from the corpus, you seem to have a revelation or four.\"*\n\nThat diagnosis is what opened the **canon distillation pass**.\n\n---\n\n## 1. The four inversions, recorded because they are the evidence base for the filter\n\nEach is a case where a session with the corpus loaded asserted the reverse of what the corpus holds.\nThey are the seed of the `demonstrated` grade.\n\n1. **\"No decomposition path looks up an existing claim before minting.\"** False. `auto_connect`\n   offers `decomposes_into` among five classifications with explicit two-direction handling, and the\n   graph holds **84 cross-source decomposition edges**. The real gap is narrower: it runs only from\n   the two extraction paths, against *other* sources.\n2. **\"The stranger test is anti-reuse — a tension nobody has named.\"** Backwards. The production\n   claim-matching literature calls decontextualisation **normalization** and treats it as the\n   *enabling* preprocessing; the corpus recorded a month earlier that *\"the recommended architecture\n   is half-built here by accident.\"* The grep hit had been read as confirming its own opposite.\n3. **\"Match identity on structure not prose — nothing in the corpus has proposed it.\"** Wrong twice.\n   It is the **scheme-relative hypothesis** (2026-04-19), explicitly *\"NOT a proposal, NOT a\n   decision\"*, with six named overreach risks — including one aimed exactly at the move made:\n   *\"elegant reconciliations across traditions are exactly what Žižek warned about as premature\n   synthesis.\"* And its deterministic first rung shipped 2026-08-20 as\n   `structural_kinship_candidates`, measured at **≈3.5/6 hand precision, the weakest of three\n   generators**.\n4. **Every reuse figure quoted described a superseded linker.** Verified: **zero edges carry\n   `r.generator`, `r.via_concept` or `r.stance_caution`**, so the entire graph was connected by cosine\n   alone — the exact linker the 2026-08-20 upgrade was written to replace, whose new generators\n   surfaced **58–76% of candidates cosine never found**. And `retrieval-instruments-beyond-cosine.md`\n   §0 explicitly warns that reading a low cosine as *not related* is **not a safe inference**.\n\n**The common shape**: a claim about what the corpus contains, made without querying the corpus. The\nglobal memory trip-wire covers exactly this (*this is new · nothing like this exists · there's no\nprior*) and did not fire.\n\n---\n\n## 2. The founder's argument about narrowness, and the measurement that supports it\n\nHis position, verbatim, given after being told the current graph shows near-zero claim-level reuse:\n\n> *\"I honestly don't really care that much about what the current state of the graph is like, because\n> I know that the flywheel of reuse is quite likely to get up to speed once people and agents and\n> perhaps backfill machinery start contributing more claims and concepts to Deliberus. People don't\n> have that many different claims and concepts that they throw around whether offline or online.\n> **People need to be understood** and that need severely narrows the space of factual or value-laden\n> claims heard throughout the world, same goes for how outlandish/divergent concept definitions ppl\n> get away with, only so many different intended meanings take hold even if accounting for different\n> subcultures using the same word/term with different definitions/meanings etc...\"*\n\nAnd the framing that set the question:\n\n> *\"even if we've not made the flywheel/asymptote appear concretely as of yet, can we EXPECT it to\n> appear if we design everything just right? that's more important to me\"* … *\"(which also connects to\n> whether there are fractal priors indicating we should have such expectations!)\"*\n\n**Measured, and the number is stronger than the claim needed.** Bar-Haim et al. (IBM Research, ACL\n2020): a professional debater, **ten minutes per side and without seeing the data**, wrote at most\nseven key points per side for each of 28 controversial topics. Against 6,515 crowd-contributed\narguments, coverage ran **41.3% at one key point → 56.3% → 64.2% → 69.3% → 71.5% → 72.3% → 72.5% at\nseven.** The blind part is what his mechanism explains: an open space is not predictable in advance.\n\n**And the limit is in the same table**: the curve **saturates at 72.5%** (five to seven bought one\npoint), with a further **22.8% ambiguous** to trained annotators. So roughly a quarter does not\ncompress — and run 6's finding is that the crux of a real debate is typically *unwritten*, which puts\nit in the uncompressed tail. **The payoff is therefore work concentrated, not work reduced**: machine\nreuse takes the ~72% that recurs, human attention goes to the quarter that does not, which is where\nthe disagreement is. Full treatment: [the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) §6.\n\n---\n\n## 3. Show, don't tell — the correction that reaches the constellation\n\nClaude first read the founder's *\"I don't wanna throw essays in people's faces. Show, don't tell\"* as\nbeing about copy length. He corrected it:\n\n> *\"I meant that this is the sort of failure mode of the old/existing landing page **and maybe even of\n> the intended new constellation of stars landing page**, since in both I'm still trying to TELL\n> fellow nerds about what/how to build/why to trust in the vision of Deliberus, about all the\n> philosophy and epistemology and usefulness and backdrop... instead of just SHOWING them what\n> Deliberus can mean and be and feel like and how it could affect their mind and life and allow people\n> to connect and communicate in fundamentally new ways, by letting them play around with it and learn\n> by doing and hit unique curiosity- or anxiety-inducing moments that they'll navigate and learn from\n> and contribute through.\"*\n\n**Accepted, and it relocates rather than cancels the constellation review**: proving each conviction\nsurvives the stranger test is condition-(d) work regardless, but the tiles' *destination* was never\nestablished. Two artifacts, two jobs — reasoning in\n[game-feel-and-the-first-screen.md](game-feel-and-the-first-screen.md) §7.\n\nHe also gave the vision the cold-start question sits inside:\n\n> *\"Even getting just a truth graph summary from Deliberus will be (able to be) epistemically balanced\n> and extremely well-calibrated (in principle) in a way that any frontier or other LLM/AI without the\n> exact right ontology/graph-based memory/toolset/harness/architecture/whatever has no chance of living\n> up to, because Deliberus creates order and clarity in a unique way.\"*\n\nwith the honest concession attached: *\"will perhaps be indistinguishable to most people, but the\nformer will be better in many respects, and we'll just have to trust that the results speak for\nthemselves.\"* **Answered: you do not have to trust it** — the substrate run is designed and blocked on\nquota, not method.\n\n**And the asymptote clarified:**\n\n> *\"I meant that since the graph will hold more and more shared hidden/implicit premises and claims\n> over time, less and less will be necessary to create anew from extractions etc\"*\n\nwhich is the cost curve from the supply side, and which made **matching-before-volume** the finding:\nvolume without matching multiplies identity scattering at an unchanged reuse rate.\n\n---\n\n## 4. The privacy incident, and why it was a recurrence\n\n`docs/relationships/kristian_ronn_2026-08-13.md` — a preparation note about a named living person —\nwas **live on deliberus.com for 15 days and enumerated by name in the public docs index**. Founder:\n*\"Make sure this never happens again. Do you feel you fully understand what caused this incident?\"*\n\nTraced: **it was a recurrence.** April 24 put two named-person briefs there plus a CLAUDE.md line\n*directing* such notes there; a May 17 visibility pass removed the **files** and left the **pointer**;\nAugust 13 a session followed the pointer and recreated the directory. Three compounding conditions:\nthe May fix corrected instances rather than the instruction (the exact shape of the directive forged\nthe day *before* this was found), the rule lives in `.private/` where nothing loads it, and the one\ncommit-time gate **explicitly skips `.md` files**.\n\nPrevention shipped: `tests/test_docs_visibility.py` — a directory allowlist (payoff-removal, since\nboth incidents were a *new folder*), a filename check measured at **0 false positives across all 259\ntracked docs with 3 of 3 specimens caught**, and a positive control so the guard cannot go inert.\n\n---\n\n## 5. The pre-launch error audit — verdict: not confident\n\nFounder: *\"we should also go through the error taxonomy... and make sure we are confident... that we\nwon't fall prey to a huge class of errors that we could have predicted beforehand, before launch.\"*\n\nRun. Structural finding first: **the residual-error taxonomy covers one of three surfaces** (the\ncorpus), while the *operator* surface is scattered and the *project* surface exists only as\ncounter-lenses. Four classes are live on the public page today, and the two most-predicted killers of\nthis venture class — **maintenance and capture** — are not error classes at all, so no §1 work touches\nthem. [the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) §5.\n\n---\n\n## 6. The canon distillation pass, opened\n\nThe founder's framing, and the reason the whole review programme is one thing rather than three:\n\n> *\"What all this is is basically me going through the project's long-term memory, sifting through it\n> and making sure the important tidbits and insights and principles and convictions etc etc are etched\n> into our short-term memory, so you Claude can have a much easier time following along with what my\n> vision is, how I imagine we best build it and why, on every level of concreteness and abstraction and\n> philosophy and linguistics and science and epistemology etc etc etc...\"*\n>\n> *\"This work is basically deeply entangled with / the same work as the review of the constellation,\n> convictions, unruled claims etc, which we need to complete arc by arc.\"*\n\n**The budget, measured**: project `CLAUDE.md` **100,599 tokens**, global **80,226**, **~180,825\ncombined per session and again per subagent** — against a documented **200k subagent-spawn ceiling**\ncrossed at ~178k in July 2026. *\"I've not hand-read the project claude.md in like forever.\"*\n\n**The filter, named**: a finding is **WRONG BY DEFAULT** when a session lacking it will confidently\nassert its opposite. His doubt, and the refinement it produced:\n\n> *\"I like this, but it'll be a bit hard to assess (I guess you can assess your state before/after\n> having read a document but idk if I always trust your self-assessment on this but we can start with\n> this and see if I feel it breaks and re-assess this test if so...)\"*\n\n→ grade candidates **demonstrated** (a session asserted the opposite on the record) or **predicted**.\nThen his correction of Claude's over-dismissal:\n\n> *\"Predicted wrong-by-default doesn't seem of NO value, it has some value, but demonstrated is better\n> ofc.\"*\n\n**Ruled: the grades are a LIFECYCLE, not a verdict.** A predicted line is a hypothesis from a reader\nwho has just seen where the naive prior points; it upgrades to demonstrated the first time a session\nactually inverts it, which makes the register self-improving.\n\n---\n\n## 7. The session-log conversion, and its root cause\n\n> *\"Convert the session-history I guess. Why did the session-history accumulate there? Is it due to\n> the prompting in /docs-persist or some such? Please fix it so that doesn't recur, have the command be\n> smarter, distill key tidbits and insights together with me in tandem with saving quotes etc from\n> every session when running docs-persist or similar commands, yeah?\"*\n\n**Root cause is not an instruction.** Nothing in `/docs-persist` or `/roundoff` tells a session to add\nan entry. **The format was the instruction** — a chronological list is self-perpetuating, its\nexistence invites the next append, and each entry is the template for the next. Entries also fattened\nunchecked (26 tokens early, 2,602 late) because nothing capped them, exactly as the frontier's size\nrule was violated for months until a *test* enforced it.\n\nThree-part fix: `SESSION-HISTORY.md` (repository root, deliberately not web-served) at repo root (21 entries verbatim,\n**deliberately not under `docs/`**, which publishes — the founder had not ruled on making it public);\n`/docs-persist` **Phase 4.5**, which distils resident lines *with* the user one at a time under the\nfilter and forbids creating any chronological log; and `tests/test_claude_md_shape.py`.\n\n**Result: `CLAUDE.md` 100,599 → ~81,000 tokens; combined preload 181k → ~161k.**\n\n## 8. Trimming by size is not trimming by value\n\nHis check: *\"What was the topic-index entry exactly? Did it deserve removal?\"*\n\nAudited: 24 of 25 sentence-units went, most correctly (citations and numbers belong behind a pointer).\n**Two did not, and were restored**, because both pass the filter — **flag-versus-reasons** (the\nvariable in durable attitude change is neither dose nor duration) and **on sacred values, argument is\nthe wrong instrument** (material pressure produces outrage; symbolic recognition produces\nflexibility). **The lesson is procedural: run the filter over the CUT, not only over the keep.**\n\nAnd the audit he opened in the same breath, now an open loop:\n\n> *\"Sounds like we need to audit it for unnecessary/long-winded bits as part of this whole big review\"*\n\n---\n\n## 9. Everything else this session touched\n\n- **The chronology finding**: the star field was drafted 2026-08-19, last tile ruled 08-21, and\n  **thirty research docs were born since** — four tiles now predate a ruling that bears on them. New\n  rule: *every compressed surface carries the date of the corpus it compressed.*\n- **The ordering**, agreed: arcs first, then arc → the tiles it compresses, with **Arc J (the headline\n  tests) first** because it governs every unruled tile.\n- **The gate's discharge path** written into `vision.md` — it stated *why* the meta layer gates launch\n  and never how it discharges, what lifts it, or what it costs.\n- **Condition (d) connected to the constellation** in the success-prior block: a tile that must be\n  *learned* is Esperanto's failure at the doorstep.\n- **`specs/` moved out of `docs/`** — a build surface is not a reader surface; 15 references fixed,\n  two links de-linked because a public doc cannot link outside the published tree.\n- **Game-feel research**: the attention premise corrected (capacity is myth, **47-second triage** is\n  measured, so brevity buys depth rather than replacing it), the founder's decades-old sketch\n  identified as a **chapter selector** — slicing, not zooming — and eight mechanics sketched that\n  could only exist on this ontology.\n- **The Swedish invitation text found** after six searches and pointed at from `CLAUDE.md`; a\n  universal English version drafted, pending his voice pass. *\"It's very good Swedish copy and should\n  have a pointer from claude.md.\"*\n- **A style correction**: *\"Do you feel like this output is following the output style I defined\n  recently?\"* — Claude had referred to him in the third person while writing to him, narrated process,\n  and teased findings.\n\n---\n\n## Cross-references\n\n[the-canon-distillation-pass.md](the-canon-distillation-pass.md) (method, budget, filter, candidates) ·\n[the-reuse-flywheel-prior.md](the-reuse-flywheel-prior.md) (the priors, the corrections, the\ncompression ceiling) · [game-feel-and-the-first-screen.md](game-feel-and-the-first-screen.md) (the\ndoorstep, the cold start, show-don't-tell §7) ·\n[game-feel-mechanics-sketchbook.md](game-feel-mechanics-sketchbook.md) ·\n[the-residual-error-taxonomy.md](the-residual-error-taxonomy.md) §5 (the pre-launch audit) ·\n[sameness-merge-and-split-across-fields.md](sameness-merge-and-split-across-fields.md) (what the\ncorpus already held) · `specs/landing-redesign/constellation-map.md` (the star review and its\nchronology) · [session26-open-rulings.md](session26-open-rulings.md) (the ledger) ·\n`SESSION-HISTORY.md` (repository root, deliberately not web-served) (the narrative record this session moved out).\n\n---\n\n## 10. Founder verbatims from the CLAUDE.md restructure and the Arc-J opening\n\n*Recorded because in each case his words ARE the ruling, and several corrected a direction Claude had\nalready committed to.*\n\n**On the review programme's priority** (which became the READ FIRST block):\n\n> *\"Be very clear in CLAUDE.md that the in-flight reviews are the highest priority until we've\n> finished them.\"*\n\n**On who decides what enters always-loaded memory** — the governing rule, now the first thing inside\nREAD FIRST:\n\n> *\"We badly need to review and create order in CLAUDE.md as a high priority, but I need to bless\n> every decision about this, note this crucial fact.\"*\n\n**On the session log, which produced the root-cause finding that the format was the instruction:**\n\n> *\"Convert the session-history I guess. Why did the session-history accumulate there? Is it due to the\n> prompting in /docs-persist or some such? Please fix it so that doesn't recur, have the command be\n> smarter, distill key tidbits and insights together with me in tandem with saving quotes etc from\n> every session when running docs-persist or similar commands, yeah?\"*\n\n**On the grading of resident candidates** — correcting Claude's over-dismissal:\n\n> *\"Predicted wrong-by-default doesn't seem of NO value, it has some value, but demonstrated is better\n> ofc.\"*\n\n**On what CLAUDE.md's remaining sections should become** (AgentLoom, the Delibr history, the academic\nreferences):\n\n> *\"'Strategic Context'/AgentLoom as some kind of engine closely tied to this is an outdated idea, from\n> like summer last year from before I had even started using Claude Code etc\"*\n>\n> *\"Maybe fold the 'Historical Note' about Janse/Delibr into the Existing Assets list or wherever\n> there's a dedicated doc about this, just leave a one-sentence brief in CLAUDE.md with a doc pointer\"*\n>\n> *\"Probably the Academic Foundations and maybe the Existing Assets sections can both be\n> extended/consolidated with more recent academic and other references from the corpus unless this\n> introduces duplication\"*\n\n**The correction that reset the whole reference pass** — Claude had been shrinking the section and\nflagging its own overshoot rather than covering the material:\n\n> *\"JFC you're making this complicated. Don't overthink the token count on this. More important that\n> all relevant references and how Deliberus relates to all these topics is covered concisely but\n> exactly.\"*\n\nThat ruling produced the section's real form: ~68 references across nine groups, each stating what\nDeliberus actually does with the source. Everything the corpus cites four or more times is now named.\n\n**On the frontier's readability** (proposal filed, not executed):\n\n> *\"frontier.md feels like it's gotten soooo cluttered over time also. It's difficult to track the\n> frontier of all conceptual work and other parts of Deliberus' frontiers without it becoming full of\n> dense jargon, impossible to skim etc, I want it to be more of an intro-like overview to anyone who\n> wants to contribute, it should dovetail with the open-questions doc I guess in that way.\"*\n\n**Opening Arc J** — the first review arc, and immediately two corrections:\n\n> *\"Agreed we should start with Arc J.\"*\n>\n> *\"J10: What if we in Deliberus mean something different by 'scaffold'? Can we go back to examine\n> this one? Also what does 'confirming' and 'checking' mean exactly? They feel so floating/fuzzy to\n> me, not much better than 'scaffold, never replace' IMO. I wanna go over this one some more, present\n> the source material/context.\"*\n\n**Both instincts were right.** The corpus carries *scaffold* in two ratified senses that disagree on\nwhether it comes down, and *\"Confirming is not checking\"* fails the same stranger test that killed the\nline it was meant to replace. Details in § 9 of the ledger's Arc J and in the J9/J10 rows themselves.\n\n**On the number of headline tests** — the live question at the time of writing:\n\n> *\"But regardless I do share your intuition about J13, we need MULTIPLE tests, not just one. But\n> whether 2 or 3 or more and what exactly they shall be, we need to discuss. Argue for that the two\n> tests you propose are enough, and then argue for that they may not be. We shall see what I decide.\"*\n\n**Status: UNRULED.** The two proposed are the **over-broad test** (*does the bolded part alone forbid\nsomething we actually do?*) and the **stranger test** (*does it say what it means to someone without\nthe context?*). The strongest candidate third, and the evidence for it is this very arc: a\n**collision test** — *does this word already do a job in the corpus?* — because *\"Scaffold, never\nreplace\"* could have been written to pass both existing tests and would still have been wrong.\n\n---\n\n## 11. Verbatim rulings from earlier in this session, captured 2026-08-29 before compaction\n\n*Founder: \"Document any verbatim rulings that have not been documented yet NOW, before compaction!\"\nThese had been acted on but recorded only in Claude's restatement, which the same day's rule forbids.*\n\n**The ordering of the review programme** — the ruling that put arcs before tiles:\n\n> *\"Then I guess we should go through the arcs and claims first, then see what that makes of each\n> star/the constellation, does that seem wise to you?\"*\n>\n> *\"Also keep in mind the chronology, we've made some philosophical headway etc since the constellation\n> work started.\"*\n\nThe second sentence produced the measurement that thirty research docs were born after the star field\nwas drafted, and the rule that **every compressed surface carries the date of the corpus it\ncompressed**.\n\n**On the congruence sweep and the gate's discharge path:**\n\n> *\"Agreed on everything you say.\"*\n\n**On the spec directory and the privacy incident** — two rulings in one message:\n\n> *\"Move it out of docs, unless we have to fix a thousand references if we do this?\n> And deploy so that the Rönn doc is not public pls\"*\n\n**On the recurrence, which produced the root-cause trace and the publication guard:**\n\n> *\"Make sure this never happens again. Do you feel you fully understand what caused this incident?\"*\n\n**On the pre-launch error audit:**\n\n> *\"We should also go through the error taxonomy that has been documented and make sure we are\n> confident (based on the fractal priors optimism or whatever theoretical or empirical basis we can\n> best muster) that we won't fall prey to a huge class of errors that we could have predicted\n> beforehand, before launch.\"*\n\n**On the Swedish invitation text** — two rulings:\n\n> *\"It's very good Swedish copy and should have a pointer from claude.md.\"*\n>\n> *\"Anyway yes the Swedish text is aimed at a particular scene but I'd like there to be a version of it\n> that is universally reusable.\"*\n\n**On the CLAUDE.md restructure** — the blessing that authorised the four items plus the three named:\n\n> *\"Do all of 1-4!\"*\n\n**On the Academic Foundations cut, and then on completing the stocktake:**\n\n> *\"Sure, go ahead\"* · *\"Excellent. Go on and verify the rest.\"*\n\n**On show-don't-tell**, the correction that reached the constellation itself — the operative sentence,\nsince the fuller passage is in § 3:\n\n> *\"I meant that this is the sort of failure mode of the old/existing landing page **and maybe even of\n> the intended new constellation of stars landing page**.\"*\n\n**On the three headline tests, ruled in Arc J:**\n\n> *\"Hmm. Yes three tests, then. Over-broad, stranger, collision. But I should ask my friend Malin\n> intermittently over the phone as a second, more real/human stranger-test, instead of relying on an\n> LLM for that test.\"*\n\n**On the word budget for a headline:**\n\n> *\"I think in order to not overly compress the meaning of each item, we have to allow each headline to\n> contain some amount of scoping, we have to allow a few more words per item so we're not amputating\n> limbs of these items to fit in a needlessly small vehicle for no good reason. We can afford a bit\n> more word/token-budget per item. Also depends on WHERE the item will end up — if we're talking about\n> the constellation stars for example, maybe the way I imagine them being displayed in the end\n> necessitates having them be just a bit shorter, but I'm also valuing quite highly the accuracy and\n> meaningfulness of each item regardless.\"*\n\n**On meta-scaffolding**, the name coined while resolving the *scaffold* collision — full quotation in\n`docs/vision.md` § Meta-scaffolding, and its load-bearing clause:\n\n> *\"This is a type of meta-scaffolding that I've been engaged in since March or really since decades\n> back, throughout early conceptual strands that were to grow into this project.\"*\n\n**And the rule this section exists to satisfy:**\n\n> *\"Can you make it a habit to preserve my verbatim phrasings when you note rulings?\"*\n\n---\n\n**Continued**: [session28-the-harness-arc-and-the-source-intake.md](session28-the-harness-arc-and-the-source-intake.md)\n(2026-08-30 — the harness conviction's canon entry, the analogy-daemon commission, the source\nintake, and the founder's hold on self-hosting pending this session's reviews programme).\n"}