History / Home

Record the threshold defect, and why proportional reporting is load-bearing The Prove It page described uncomfortable_findings without saying when they stop being reported, which turned out to be the interesting part. An earlier version tested `if dissents and not resolved`, so the first resolved dissent made the finding vanish — 1-of-9 read identically to 9-of-9. Nothing had become healthy; the system had acquired one data point and its self-report went quiet on a pattern that had barely moved. For an instrument whose entire purpose is to be unflattering that is the worst failure available to it, and an invisible one, because the output still reads as a clean bill of health. Findings are proportional now: a gap is reported until it is closed rather than until it is touched, and the wording carries the ratio. A fully closed gap does go silent — an audit that cried wolf forever would be switched off, which is worse than no audit. The page also records how the defect was found: a review question written in advance specifically to grade the system's interpretation rather than its bookkeeping. The bookkeeping had been honest the whole time — every figure checked out against the database. That distinction is the thing worth documenting. Counts synced to measurement: 709 tests. All 23 links resolve.

noreply committed Jul 30, 2026

Document the epistemic layer; correct the counts against measurement Two new pages for the capability that shipped this week. Prove It — the receipt for a belief, and the field that matters: trust_basis is UNEARNED when a belief carries a confidence figure but nobody who asserted it has ever been graded against reality. Most beliefs start there, and reporting it is the entire point; a system that always sounds justified has told you nothing. Includes what the self-audit found when run against this project's own database: nine dissents recorded, none ever resolved. Judgment Queue — why that happened. Not a broken judge: auto_judge correctly abstains on a claim no judge recognises, the claim stays pending, and dissents only resolve once their session is fully graded. One unanswerable claim silently blocked every dissent beside it. The page records what closing the loop actually did — the Skeptic marked proven_correct, reputation moved, and a belief revised by Reality rather than by further discussion — and the doctrine that a dissenter who turns out wrong is never punished, because punishing wrong dissent teaches a council to stop dissenting. Counts corrected from measurement rather than memory: 272 routes, 672 tests. "the 56 built-in tools" became "the built-in tools" — the number now moves with whichever providers are configured, so pinning it in prose would only make the page wrong again later. All 23 wiki links resolve.

noreply committed Jul 28, 2026

Audit every claim against the code, and correct what did not survive Not a copy-edit. Each number and name below was measured from the running system, and these are the ones that did not match. **The policy stack was described wrongly, and it is a security page.** The wiki said a single PRE_ACT hook runs seven policies, listing all seven. Four run at PRE_ACT; the other three run at PRE_VALIDATE and PRE_COMMIT. So the page told a reader that seven policies stand between a decision and a side effect on their machine, when four do — and that the fabrication guard fires before the action, when it fires before the commit. Architecture, Security Model, Governance Engine and First Run all carried it, because all four had copied the boot log, which was itself printing a flat list under one hook's name. The log is fixed in the repository; these pages now match it. **Recall Quality invented three of the five validity state names.** The code has VALIDATED, PARTIALLY_VALID, OUTDATED, DISPROVEN, UNKNOWN. The page said Validated, Provisional, Superseded, Disproven, Expired — so a reader seeing PARTIALLY_VALID in an API response would not have found it documented anywhere. The five score fields are now their real names too, with the 270 and 540-day staleness horizons that decide two of the states. **Outcome Tracking was a horizon and a grade out.** There are four horizons, not three — 7/30/90/180 — and exactly three grades: correct, partial, incorrect. "Unverifiable" was listed as a fourth grade; no such grade exists, and the underlying idea deserved better than a wrong row. A judge that cannot observe abstains, the prediction stays pending, and the abstain rate is reported as its own number rather than collapsed into a zero. **Institutional Memory documented a command that crashes.** `migrate.py down` with no version raised IndexError. The runner now says what it wants; the page shows the version. Smaller: the Elite Commander is OPV-000 in the code, not an em dash. Prompt sizes are stated as approximate and install-dependent, because they move with your skills and profile — the old page gave exact figures no install produces. New page: Continuous Integration. Home: dependency count corrected from 5 to 16 plus 2 for tests. Verified: 267 routes, 39 house and 28 council, 56 tools and all 16 named ones present, 22 skills, 14 checks, 5 migrations, 14 council members, 606 tests. All 21 wiki links resolve.

noreply committed Jul 24, 2026

Correct the health verdict the wiki told users to expect Installation and API Reference promised `GREEN — all 14 checks pass` on a fresh install. That is not what a fresh install returns, and after today's fix it never will be: before a model runtime is started the ollama check is YELLOW, so the honest first-run answer is `13 green · 1 degraded` with ok:true. A first-time user reading the old page would conclude a correct install was broken. Troubleshooting now explains the three verdicts as a distinction in kind — YELLOW means running with something optional absent, RED means the House itself is faulty and is the only state that makes ok false — and names the two YELLOWs that are normal on a clean machine. Configuration documents OLLAMA_BASE_URL, which is now actually read by the backend rather than only mentioned in the compose file. New page: Continuous Integration — what the ten checks each prove, why the boot job is eleven small steps instead of one script, and why the privacy gate is a committed script anyone can run rather than a workflow inline. Home: dependency count corrected to 16 + 2 dev (it said 5, from before the undeclared dependencies were declared). All 21 wiki links resolve.

noreply committed Jul 24, 2026

Add the full wiki (22 pages): getting started, architecture, memory subsystem, operations, security

ElmatadorZ committed Jul 24, 2026

Initial Home page

ElmatadorZ committed Jul 24, 2026