{"path":"research/unweighed-answers-two-months.md","content":"# Two months of unweighed answers (2026-07-21 → 2026-09-21)\n\n**Date**: 2026-09-21 · **Type**: register, written on founder instruction — *\"I would like a review\nlike this for all of the sessions from the last two months, looks very helpful to be able to look\nback and see patterns/inconsistencies/subtleties, or just simply the enormous mass of unruled\nideas\"* · **Model**: `session34-open-loops-and-unweighed-answers.md` § 4, extended backwards.\n\n**Why this is a different object from `TODO.md`.** The register there answers *what is left to do*.\nThis one answers *what was said that nobody weighed* — a finding, an offer, or a correction that was\ngiven once, drew no reply, and is therefore neither accepted nor rejected. The standing rule is that\n**cost of re-derivation decides what gets written down, not importance**, and an unweighed answer is\nthe cheapest thing in the project to lose and among the more expensive to reach again.\n\n---\n\n## 1. What this cannot be, stated before the numbers\n\n**Most of the chat from this window is gone.** Ten session transcripts survive for the two months\n(`~/.claude/projects/…/*.jsonl`, measured 2026-09-21), and Claude Desktop's bundled cleanup deleted\neverything older than thirty days on two occasions; 80 of 226 were recovered from a Time Machine\nsnapshot, the rest are unrecoverable (`~/dotfiles/docs/incident_cc_transcript_cleanup_2026-09-08.md`).\n\nSo **this register is harvested from the written corpus, not from the conversations.** Its\ncompleteness is therefore exactly the completeness of the habit of writing things down — which is\nitself one of the findings below, and the reason the true number is larger than any count here.\n\n---\n\n## 2. The mass, counted\n\n| | Count | Source |\n|---|---|---|\n| Rows in the ruling ledger, still `OPEN` | **143** of 205 | `grep -c '\\| OPEN' session26-open-rulings.md` |\n| Passages in `docs/` and `specs/` marked *unruled · proposed · never answered · his call* | **105** across ~40 files, and 115 a day later — this document's own additions moved it, so read the grep rather than the number | corpus grep |\n| Research documents created in the window | **136** | `git log --diff-filter=A` |\n| Commits filed as `correction:` | **24** | `git log --grep '^correction:'` |\n| `⚠` marks inside documents (a claim withdrawn or narrowed in place) | **60**, `rg -o '⚠' docs/ \\| wc -l` for today's | corpus grep |\n\nThe ledger's open rows spread across fourteen arcs, heaviest at **E (19), I (15), M (14), H (14),\nL (13)**. No arc is closed.\n\n---\n\n## 3. Four patterns\n\n### 3.1 \"Unruled\" names three different things, and only one of them is a decision\n\nThis is the load-bearing finding, because it is why 143 reads as an impossible backlog.\n\n- **A fork** — two live options, and the work branches on which. *Does an unanswered latent\n  requirement enter the cap, or only the coverage measure?* Nothing proceeds until he picks.\n- **An offer** — something proposed with no fork behind it. Accept or drop; the project is fine\n  either way. The Shapley strength explainer, the drowning dials, the per-scheme aggregation layer.\n- **A finding** — something measured or read that changes nothing until somebody says it does. The\n  reported-speech base rate. The granularity number narrowed to sense granularity. That LLM\n  facilitation did not improve consensus while participants preferred it.\n\n**The register files all three identically**, so a finding that wants ten seconds of acknowledgement\nsits in the same queue as a fork that blocks a build. Concretely: of the 143 open ledger rows, the\nones carrying a real fork are a minority, and there is no way to see which from the outside.\n\n**The cheap fix, proposed:** one letter per row — `fork` / `offer` / `finding`. It costs one pass\nover the ledger and turns an undifferentiated 143 into a short list of things that actually block,\nplus a long list he can skim and close in batches. **An offer he never wanted should expire rather\nthan accumulate**, and today nothing expires.\n\n### 3.2 One question with several homes never gets answered\n\n**Measured instance**: *does a signed-out visitor see machine-made structure?* It has **four written\nhomes**, phrased differently in each:\n\n- `the-tidying-line.md` § Unruled — *\"whether machine-tidied structure is visible to a signed-out visitor\"*\n- `the-jhu-decomposition-line.md` § 3 — *\"For a signed-out reader: the proposed parts are simply on the map as dashed boxes\"*\n- `session34-open-loops-and-unweighed-answers.md` § 2.4 — *\"Does a signed-out visitor see a split proposal that already exists, read-only?\"*\n- `TODO.md` — *\"A signed-out visitor still sees no split panel… Raised three times, never ruled\"*\n\nThat last line is itself an undercount: it is four, and TODO does not know about the first two. **In\neach home it reads as a small local note rather than as one live product decision**, which is exactly\nwhy it survives repetition without ever being answered. It is also the *show, don't tell* surface the\nfounder has named as mattering, so its size is badly misrepresented by its filing.\n\nThis is the corpus's own **uncollected-idea** pattern (`structuring-gradient-lineage.md`) arriving in\nthe register rather than in the build. The countermeasure that document already prescribes: **one\nname, one home, and one thing that depends on it.**\n\n### 3.3 The unruled mass measures the writing habit, not the backlog\n\nThe concentration is in the newest work — `constellation-map.md` (10 marks), `session34` (8),\n`the-harness-conviction.md` (5), `map-view/design.md` (4). Documents from July and early August carry\nfew.\n\n**That is not because their questions were settled.** It is because a question nobody marked stops\nbeing visible, and the marking convention (*proposed, unruled*) tightened over the window. So the\nregister is a measure of **how recently something was written down**, and the true backlog is larger\nthan 143 by an amount nobody can count.\n\nConsequence worth holding: **a low unruled count in an old document is weak evidence of anything.**\nWhere an old design doc reads as settled, check whether its questions were ruled or merely dropped.\n\n### 3.4 The corrections are the subtlest class, and they are invisible in every count\n\n**24 correction commits and 60 `⚠` marks in two months** — claims of mine, withdrawn or narrowed\nafter he had already read them. They are neither decisions nor offers: they are **retractions of\nthings he may still be carrying**, and they appear in no ruled/unruled tally.\n\nTheir cost profile is the inverse of everything else here. An unweighed offer costs nothing if it is\nnever taken up. **An unweighed correction costs whatever was built on the version it withdrew** —\nand the corpus has measured that shape before, where a doc that *quotes* a rule has usually built\nsomething on it while a doc that merely mentions it goes stale harmlessly.\n\nThe highest-value ones from the window, because each replaced something he had reason to believe:\n\n- **The reuse lookup exists and fires** — the flywheel was recorded as impossible; it runs, it is\n  paired rather than cumulative, and 84 cross-source decomposition edges came from it (⚠ corrected 2026-10-07: these were premise links stored backwards as part-of links; all are retired and were read again as support: `auto-connect-upgrade.md` § 2c) (`1ad246e`).\n- **A missing consideration IS mostly derivable** — the opposite of what the corpus said, tested on\n  thirteen real arguments (`4f0cb28`), and then **softened again** because the test was one-sided and\n  model-graded (`b0ac2e2`). Two corrections in the same place, in opposite directions.\n- **The penguin finding read at its own scope** — the widely-cited *\"two people share the same penguin\n  one time in eight\"* overstated the measurement, which was about feature profiles rather than the\n  concept (`bb28e49`).\n- **The looks-bundled 84% was my inference, not a finding** (`31af534`).\n- **`vision.md` overstated a founder position** (`4f35af4`) — the one class where an uncorrected\n  overstatement puts words in his mouth on a public page.\n\n---\n\n## 4. The register, by kind\n\nEach line is one item, its home named. Nothing here is new work; everything is already written\nsomewhere, which is the point — this is the index that does not exist anywhere else.\n\n### 4.1 Forks — two live options, the work branches\n\n| Item | Home |\n|---|---|\n| The lift: coverage-checked lifting, raised-not-total up-weight, cap through argued necessity only outside the derivable space | `weakest-link-arithmetic-and-the-merge-hunch.md` § 9d |\n| Do unanswered *latent* derivable requirements enter the cap, or only feed the coverage measure? | `derivability-of-missing-considerations.md` § 5 |\n| The presumed requirement weights are de-baked constants (0.5 / 0.25 / 0) with no ruling behind them | same, § 9d; ledger O13 |\n| Publish the ontology's own commitments into the graph, or keep de-baking them one at a time? | `what-belongs-in-the-ontology.md` |\n| Is the sacredness brake right at all? Open since 2026-07-10 | weighing design doc |\n| Should the bridging feed read the badge? It ranks by an undefined extraction-time rating that no answer updates | `strength-layer-audit.md` |\n| Adopt fair-share discounting on shared evidence? Built, tested, switched off | `evidence_routing.py` |\n| Does the terminus classifier keep publishing a decimal? | § Judge calibration |\n| The four delegation questions, open since 2026-08-17 | `conviction-and-critique.md` |\n| Which analogy for which register (chart, ledger, pilot, chef) | `the-harness-conviction.md` § 8.6 |\n| Fork one: call the horizon map's aleatoric fraction publicly? Fork two: curiosity as a second conviction term? | `convergence.md`, `curiosity-as-growth-fuel.md` § 11 |\n\n### 4.2 Offers — proposed with no fork; accept or drop\n\n| Item | Home |\n|---|---|\n| The Shapley/RAE strength explainer (per-edge attribution summing to the badge's deviation) | `the-irit-harvest.md` § 1 |\n| Per-scheme aggregation (analogy edges may want counting-flavoured semantics) | same, § 3–4 |\n| The drowning dashboard — six dials incl. the Good-Turing novelty rate | `drowning-in-claims.md` § 6 |\n| The analogy daemon, and region-level *temperature* as its roaming signal | `the-analogy-daemon.md`, `perceiving-possibility-and-the-analogy-organ.md` |\n| Head/tail citation balance for the synthesis ledger | `saturation-and-the-long-tail.md` § 3b |\n| A failed reconstruction re-files a split as `SUPPORTS` — the discriminator runs and nothing reads it | `core-periphery-and-the-concept-layer.md` § 5a-ii |\n| A correction-push channel (the Notices-to-Mariners shape) | `the-harness-conviction.md` |\n| Confession must pay: score a `does_not_fit` above a wrong nearest fit | `reflection-and-the-graph.md` § 4 |\n| The look-back (map a finished advice process after the fact) and the smallest thing members could use | `blobs-community-sandboxes-and-the-advice-process.md` |\n| The two-sided derivability test — hand the thirteen arguments to people from the affected positions | `derivability-of-missing-considerations.md` § 10 |\n| Position grouping + a sole-witness label | `cheap-to-add-slow-to-matter.md` |\n| The decisions-per-pass table (new, 2026-09-21) | `TODO.md` § Verification pass |\n\n### 4.3 Findings — true, recorded, and changing nothing until weighed\n\n| Item | Home |\n|---|---|\n| The reported-speech base rate: 153 read, 46 minted, 107 refused; a 30% mint rate is the pre-filter erring toward reading | `earning-the-right-to-write.md` § 7 |\n| Three failure shapes in real splits — loss, **invention** (the dangerous one: a split that looks *better* than its parent), drift into commentary | same |\n| Reviewer divergence is neither taste nor noise: it marks where nobody has said what the split is for | same |\n| The 56%-vs-91% granularity number is about *sense* granularity and does not transfer to cuts between propositions | `core-periphery-and-the-concept-layer.md` § 5a |\n| A path collision is not a style question: the API wins and the visitor gets raw JSON | `session34-open-loops-and-unweighed-answers.md` § 4 |\n| Six preconditions for propose-tier daemons; the split proposer meets five, and the unmet one is **an undo path** | `graph-daemons-design-space.md` |\n| Dummett named our two kinds of cut in 1981, and the paradox of analysis stated the split-versus-argument tension in 1942 | `a-cut-through-a-concept-is-an-argument.md` |\n| The parts-versus-premises boundary has **no settled test in the field** — about a dozen rival criteria — so our recomposition check is one criterion, not a correction | same, § Unruled |\n| The AI-doom diagram's author **declares his own crux** — a hand-computed hinge published before we touch it | `session34-open-loops-and-unweighed-answers.md` § 5 |\n\n### 4.4 Questions raised more than once and never answered\n\n| Item | Homes |\n|---|---|\n| Does a signed-out visitor see machine-made structure? | **four** (§ 3.2 above) |\n| Should the hovered box on a source map carry its three tags? | `session34` § 2.5 — *\"My open call, offered and unanswered\"* |\n| What should a section correspond to? | `how-a-map-is-cut-and-what-it-shows.md`; the measurement exists, the ruling does not |\n| Say the Anthropic thing out loud, or stay quiet? | § Judge calibration; lean recorded, never ruled |\n| Change the vision language from *faster decisions* to *checking that keeps pace with persuasion*? | `the-scrutiny-gap.md`; ratified-conviction text, so his alone |\n\n### 4.5 Corrections — withdrawn claims he may still be carrying\n\nListed in § 3.4. The register's own recommendation: **a correction is the one class that should not\nwait for a ruling**, because leaving it unweighed means the withdrawn version stays in circulation.\n\n---\n\n## 5. What would make this stop accumulating\n\nThree moves, in order of cost, none of them ruled.\n\n1. **Mark each ledger row's kind** (`fork` / `offer` / `finding`). One pass. It is the difference\n   between a 143-row backlog and a short blocking list beside a long skimmable one.\n2. **Give an offer an expiry.** An offer nobody took up in two months is evidence about its value.\n   Closing it is a decision too, and cheaper than carrying it.\n3. **Collapse a repeated question into one home before answering it.** The four-home instance above\n   is the worked example: the answer is easy, and the *filing* is what kept it open.\n\n---\n\n## Cross-references\n\n`session34-open-loops-and-unweighed-answers.md` (the model, and the one session this treatment\nalready exists for) · `session26-open-rulings.md` (the ledger) · `TODO.md` § Verification pass\n2026-09-21 (what was measured closed) · `structuring-gradient-lineage.md` (the uncollected-idea\npattern and its countermeasure) · `the-residual-error-taxonomy.md` § 6 (ratification decay, which is\nthis register's failure mode one level up)\n"}