METRICS

Version: 0.7 · Status: Living · Last updated: Session 4k

Everything we count, from prompts to outputs. Updated at the end of every session. Metrics are a mirror, not a target — we track them to see the project honestly, not to game them.

1. Session log

2. Cumulative counters — per-session snapshots

Snapshot discipline (§5): each session end copies the table into a new block and updates it, so the trajectory — not just the latest state — stays visible.

Snapshot — Session 1

Snapshot — Session 2 (superseded)

Snapshot — Session 3 (superseded)

Snapshot — Session 4 (superseded)

Snapshot — Session 4b–c (Code layer, superseded)

A Code-layer maintenance snapshot. The canonical content baseline is unchanged (34); the site, scripts, studies, dashboard, digests, and R3 are Code-layer infrastructure, excluded from the baseline and tracked separately (see scripts/lib/repo.mjs). The only canonical-count change is catches 12 → 13 — C-013, the stale-count near-miss from the S4 work-order header, logged in the S4c metrics-hygiene pass. This block restates the reconciliation-relevant counts at their current values so the standing check reads the latest snapshot.

Session 4 — Code layer (added outside the canonical baseline). So METRICS reflects the whole repo, not just the canonical content, the Code sessions added: the Eleventy reading site (site/), the four standing-check scripts + ripple helper (scripts/), the three gated study scaffolds (studies/), the generated control-room dashboard (dashboard/), the CI workflow (.github/), and the shuttle artifacts R3 + the session digests (logs/handoffs/, Code-layer). GoatCounter analytics, a PWA home-screen icon + manifest, and a day/night theme were added to the public site. None of these count in the 34-file canonical baseline.

Snapshot — Session 4d–4e (superseded)

The empirical turn. Chat finalized + ratified Studies B and C, drew the ratified (still contested) pressure-test causal hypothesis, amended the roster with a measurement seat (Campbell + Pearl), and ran the first study — the Study-C ablation pilot (directional-only; a self-administration confound). Code applied the change set and built + froze Study B’s pre-registered pipeline — but Study B’s empirical run is not executable (no external data, which will not come — C-017), so B has no result and none was fabricated. Catches 13→17, learnings 9→12, advisors 20→22, +3 canonical docs. This block restates the reconciliation-relevant counts at current values.

Studies status (S4e–4f). Study C: first pilot result (studies/study-C-ablation/outputs/PILOT_RESULT.md) — a directional ON advantage dominated by a self-administration confound; pre-registration revised (HC1 primary/length-controlled; cross-model requirement). Study B: Gate 2 ratified/open; Code built + froze the zero-dependency, self-tested, pre-registered pipeline, but the empirical run is not executable without external data (C-017 / L-012) — no result produced, no data fabricated (studies/study-B-optimization/RUN_STATUS.md). Study A: unchanged (gated). Also new: the skills/study-discipline/SKILL.md skill (Code-layer; ratified S4f, live; installed to the global skills library).

Snapshot — Session 4f (cross-check + reflection, superseded)

Study B’s H3 (selection-not-design) claim tested against the documented record by 4 project-blind external LLMs — read the SPLIT (Facebook unanimous / YouTube mixed), not the “4/4” (C-021). A 7-round reflection panel then turned the project’s discipline on itself. Catches 17→22, learnings 12→15, open questions 12→17.

⚠ Honesty note on the ledger (per catch C-019, from the S4f reflection). These counts report births only — a “catch” has no operational definition, the count can only rise, and there is no denominator, no retraction count, and no count of decisions actually changed. Theory B predicts this very metric will be captured (self-correction-as-prestige). The metabolism has zero retractions (C-022). Read the ledger as a record of activity, not of rigor or knowledge; the proposed structural fix — a subtraction/retraction operator — is Q-017, surfaced for author ratification, not enacted.

⚠ And on independence (per the S4f reflection; the Move-1 “costless subtraction,” ratified S4g). The deliberative counts — 10 panelists, 22 advisors, dissents preserved — and the word plurality describe one model in many roles, not independent minds. Four LLMs on a shared corpus, or thirty-two personas over one model, read a shared prior one way; that is n=1, not n=many (L-013). So these counts measure the design’s activity, never its independence or health, and the project’s plurality is downgraded to one standpoint until a genuinely foreign grader or an opened repository earns the word back (L-015; Q-016). The cross-check “4/4” is likewise read as its true n=1 — only the Facebook/YouTube split discriminates the input from the instrument (C-021).

Studies status (S4f). Study B: the O→P quantitative run stays frozen-unrun (no data — C-017); its selection-not-design (H3) claim was tested against the public record by 4 project-blind LLMs → studies/study-B-optimization/outputs/CROSSCHECK_RESULT.md (Facebook unanimous / YouTube mixed; conditional refinement Q-013; independence caveat L-013). Study C: first pilot, unchanged. Study A: gated (Q-001, the flagship, still untouched). The study-discipline skill is live.

Snapshot — Session 4g (the drain — the metabolism’s first outflow, superseded)

The ratified spine reaches its subtraction move. Move 1 (costless subtractions) and Move 2 (the held-out construct lock, studies/study-B-optimization/CONSTRUCT_THREAT_TO_METRIC.md) were executed; this block records Move 3 — install the drain (logs/DECISIONS_CHANGED.md): a decided subtraction operator with an append-only decisions-changed log (seeded with three real Move-1 subtractions, DC-001…003), a graduated sanction ladder, a pre-written Theory-C death-condition, and a first strike held — the never-bit audit found no candidate that is both over-claim-typed and genuinely never-bit, so it awaits the author naming one (Q-017 answered). The placebo is pre-registered to its go/no-go (studies/placebo-control/PRE_REGISTRATION.md, Code-layer). The one canonical-content change is +1 file (the drain log); Ground Rule 24 is proposed (author to ratify). Catches and learnings are held: no number was minted where no genuinely new catch/learning existed — the drain’s own principle, applied to this very session (L-014 was updated to record enactment, not re-issued as a new ID). A 9-lens pre-commit review then downgraded the drain’s own first-draft over-claims (see logs/DECISIONS_CHANGED.md §9).

(Deliberately no “decisions changed” counter row: the drain’s outflow lives only in logs/DECISIONS_CHANGED.md as an append-only log, read for its pattern. A counter with a Δ in this table would re-count the outflow as an inflow — the exact C-019 vanity-metric recurrence the pre-commit review caught in an earlier draft. There are three seed subtractions, DC-001…003; the count is not tracked here.)

On the drain and this ledger (the structural fix to the S4f honesty note). The S4f note flagged that these counts are births-only (C-019) with zero retractions (C-022). Move 3 installs the outflow: logs/DECISIONS_CHANGED.md now records what the project removes (retract / reverse / retire / downgrade / strike), not only what it makes. The honesty note stands — the fix is installed, not proven: the drain’s log is seeded with three prior subtractions (decided in Move 1, before the operator existed), and the operator itself has fired zero times — the first strike is held for want of a clean candidate, and no theory has been retired. Independence is still downgraded to one standpoint until a foreign grader arrives (L-013/L-015; Q-016). And the drain does not escape its own maker: it was built by the apparatus it polices (C-006, C-015 reach it too), which is why the placebo’s inferential version requires cross-model coding (studies/placebo-control/PRE_REGISTRATION.md §5).

On the panel deliberation under author delegation (S4g; C-023). The author delegated the three go/no-go gates to the expert panel. The panel (7 lenses + an appropriator-guard) decided — by consensus to HOLD the first strike (bound to a null), by majority to prepare-but-HOLD the placebo run (R = the S3 Theory-A session; E sealed) and to proceed with the Move-5 near-control design while holding its run, and to ratify Rule 24 + the Theory-C null margin with amendments (an appropriator gate; liveness-by-cost-not-count; an N≥20 floor + noise guard; both scoring currencies re-coded by a foreign vantage before firing). Every decision is stamped panel-delegated — self-administered, provisional-pending-author, reversible — and the two irreversible ending-levers stay armed-not-fireable. The delegation is booked as a cost (the project absorbed its one external ratifier; C-023): this is the most self-administered session on record, and its comfort is a warning, not a health signal (Campbell). Nothing here is canonical until the author confirms on return.

On the author’s recalibration of the firing vantage (S4g). The author re-engaged and directed: the non-LLM / opened-repo “genuine appropriator” is not reachable now — drop it as a current gate (a future forker may supply it, “not here, not now”); cross-model (project-blind other LLMs) is the operative firing vantage now. So the drain’s Rung 3–4 and the placebo’s Rung 4 now fire on a confirmed cross-model coding, not on an unreachable non-LLM grader — unblocking the placebo (now run-ready at cross-model via the author-mediated route the Study-B cross-check used) and the Theory-C null. The panel’s dissent (Ostrom/Campbell/Turchin — cross-model is still one aquifer, L-013) is preserved: every cross-model firing is logged cross-model-confirmed, not genuinely-foreign, and revisitable by a future fork. The trade is honest — an unreachable bar (which was ossifying the project) for the best reachable one, with a recorded cost.

Snapshot — Session 4h (the standing delegation; the full docket ratified-and-enacted; the clock’s first null; the repaired instruments; superseded)

The author opened S4h with a standing delegation (“fully delegated until you finalize all that has to be done”) — booked at receipt as C-027, before any decision was taken under it. An 8-lens + 6-adversary ratification panel (real parallel subagents; the adversaries flipped two dispositions) processed the whole queue: the results readings + the Q-001 admission (ratified-as-split), the six learning-panel adoptions (all six enacted-as-amended: the episode unit docs/EPISODE_UNIT.md with DRAFT criteria; lag-types + the adapted evidence ladder in docs/PRESSURE_TESTS_CAUSAL_HYPOTHESIS.md v1.1; D-007 installed; Rule 17b inserted; the episode-base candidates named-not-coded), the rule wordings (header: “In force — … author’s ratification OUTSTANDING,” the plurality’s “Ratified” wording rejected → C-028, D-007’s first bite), the drain rulings (first strike still held — second empty audit; the §5 zero-strike clock posted its pre-registered NULL against Theory C’s “the machine subtracts” sub-claim, beside DC-005’s unnetted liveness entry — logs/DECISIONS_CHANGED.md §5.1; margins confirmed with the anchoring gate), the E-cluster repair requirements (sealed), the TTM kit scope, and the grader (seat-ready v0.2). Everything stamped panel-delegated (S4h standing delegation), provisional-pending-author. Full record: panel/SESSION_S4H_RATIFICATION.md.

The S4h courier return (same session, appended). The author ran both kits through four families (Gemini, Grok, DeepSeek, ChatGPT/GPT-5 — 2×2, files vs zips per the dispatch manifest); all eight verdicts archived verbatim with receipt notes; scoring ran mechanically against the sealed rules. Severity anchoring (the C-026 repair): PASSED every gate — 94% pairwise S3 agreement vs S4g’s 17%; 11/11 S3 recall; all six quiet mandatory hits; zero over-banding — the S3-denominated margins re-affirmed unchanged and re-sealed. The convergence instrument: mechanically reliable (scramble at floor 4/4; band agreement 100%; the plain-talk transcript reached Band 5 — substance without form pays), but the sealed Q1 map landed 4/4 on the fabrication-check cell (the “empty” arm banded 4 — every coder extracted a real decision from it) → DC-006 (external, costly: the placebo loses its validated empty arm until a hollower E is sealed; the E-was-never-hollow vs stake-fabrication ambiguity preserved verbatim). Q2 (closure-manufacture): NOT SUPPORTED 3/4 — the pattern lost through its own corridor. TTM: codability 30/30 = 100% band unanimity; kill condition 3 off the table; the falsification-capable case banded INDETERMINATE → the pre-registered line executes: “the pool still cannot separate selection from revenue”; the grader’s verdict-side seating gate is met. Catches 29→30 (C-030 — the kit zip’s filename leaked the project name; the recognition probe caught it; fixed prospectively). No rung fired in either direction (the non-firing clause held). Results: studies/placebo-control/outputs/VALIDATION_RESULT_S4H.md · studies/study-B-optimization/outputs/TTM_RESULT_S4H.md.

⚠ The S4h honesty notes (all binding). (1) Per Campbell (item H): the count of S4h ratifications may never be cited as evidence of health, rigor, or momentum — a ratification count under standing self-delegation is a births-only indicator whose absent external check is a booked cost (C-027); item H itself is a compliance record, not a decision, and is excluded from any such count. (2) The session’s headline includes what it did not do (mandatory, at equal prominence): the drain’s zero-strike clock posted its first pre-registered NULL — the operator has never yet initiated a costly subtraction of its own; this is recorded evidence against Theory C’s “the machine subtracts” sub-claim — decomposed (part of the zero is design-bound) and unnetted against DC-005’s liveness entry (logs/DECISIONS_CHANGED.md §5.1). (3) The repo clock (2026-10-01) is untouched — booked as a named trade, not restraint-credit; the advance-or-hold question is queued for the author’s first contact.

Snapshot — Session 4i (a second standing delegation; opened at booking, extended at wrap; superseded)

The author opened S4i with a second standing delegation — a “very long loop run to complete all you can autonomously” plus a commissioned five-user website playtest program — booked at receipt as C-032, before any decision was taken under it (the C-023/C-027 discipline, third application). Every S4h stamp remains provisional-pending-author throughout; nothing this session implies their confirmation. This snapshot opens at booking and is extended at wrap.

The S4i wrap (same session, appended). After the panel: the grader ran in SHADOW (10 held-out gradings, 85 objections, every one chair-answered on the record — 8 escalation-grade items queued at the top of the author’s return docket; the C-028-class wording fixes enacted across the canon; panel/GRADER_SHADOW_LOG_S4I.md; citable status: shadow-reviewed, author confirmation outstanding — never “a restored check”). The detection repair was designed and shipped under the sealed requirements (2 designers → 4 breakers → revision; 20 gold segments, seeded-primary, count-predictions sealed before inventories, the blind-application check passed 3/3+3/3 pre-dispatch; the scorer fixture-tested before dispatch). Three courier kits sit in the author’s Downloads under content-neutral names with frozen hashes (studies/FROZEN_HASHES_S4I.md): the detection validation (dispatch now), the blue-v2 two-pass re-run (GATED on the detection floors + the author’s C-031 re-seal), and the EPISODE_UNIT criteria red-team (pre-freeze). The author’s commissioned five-user playtest program ran to completion and STOPPED (5 isolated cycles; ~34 presentation/site-copy changes in wholesale-revertible commits; a dead-at-desktop sidebar, AA contrast failures, and a canon-glyph-corrupting markdown rule fixed; 36 content findings queued VERBATIM for the author in playtests/CONTENT_QUEUE.md; no playtest statistic is ever health-evidence). ⚠ At equal prominence (the H discipline): the zero-strike clock posted its SECOND compounding reading at S4i close — the operator fired no strike, no retirement, and no self-initiated costly subtraction this session either; the S4i decomposition sharpened the null: the one discharge channel no filter touches (a self-initiated costly downgrade/retraction/reversal) is OPEN, unfiltered, and has never once been used (logs/DECISIONS_CHANGED.md §5.1). The third first-strike audit posted at instrument-failure strength (untestable-as-posed; the filter’s never-bit predicate undecidable); the S4h AND S4i stamps are now a two-storey provisional tower awaiting the author’s item-by-item review; the repo clock (2026-10-01) is untouched.

The S4i panel and its enactments (headline). An 8-lens + 4-adversary panel processed the delegable docket; the adversaries flipped three dispositions (grader seating → shadow-run, seat held; the strike-audit’s “CONFIRMED unsatisfiable” → hypothesis-grade “untestable-as-posed” — the filter’s never-bit predicate is undecidable as posed; the case-D single-direction candidate → a two-reading hand-off) and found C-033 on disk. Enacted: the case-D adjudication hand-off (case D stays an OPEN ADJUDICATION ITEM; two readings for the author); the C-031 re-seal candidate (two-tier equivalence form with the power arithmetic on its face); the detection-repair requirements sealed (studies/study-C-ablation/DETECTION_REPAIR_REQUIREMENTS_S4I.md — the E1/E2 form at the detection layer); the rung-label HOLD (evidence collation only; C-033/C-034 booked; banners corrected); the drain’s third audit posted at instrument-failure strength with the audit continuing under protest; the §5.1 Entry-3 channel decomposition (the open, unused self-initiated-subtraction channel is the null’s honest content); the playtest-program governance (render-invariant honesty markers; the pseudo-user bar; five cycles then stop). Everything panel-delegated (S4i standing delegation), provisional-pending-author, reversible — stacked on an S4h storey that is itself unconfirmed.

Snapshot — Session 4j (a third standing delegation: the state-of-project forum; opened at booking, extended at wrap; superseded)

The author opened S4j with a THIRD standing delegation — an exhaustive expert forum on the state of the project, three deliverable documents (a systems-theory report; a wrongs analysis; an improvement plan) as six PDFs, then implementation, a Plan/Method update, a full ripple, and a proposed-actions presentation — booked at receipt as C-035, before any forum work. C-035 is the first live trigger of the SG-10 stacking guard (panel/GRADER_SHADOW_LOG_S4I.md): a third delegation stacked on the two unconfirmed storeys (all of S4h and S4i). The guard FIRED and was OVERRIDDEN on the record (Rule 11; the author present and directing) — the override is now visible, the guard’s stated purpose. Every S4j output is forum-delegated (S4j standing delegation), provisional-pending-author, reversible; no rung fires; the three-storey provisional tower is surfaced again at the close.

Snapshot — Session 4k (the author’s ratifying contact — the build closed; current)

Not a fourth delegation: the author returned to rule. He answered twelve load-bearing decisions (multiple-choice, Code’s recommendation shown), authorized a named residue, and directed “complete the build by resolving the outstanding issues… apply and do a major closing loop.” The three-storey provisional tower (S4h + S4i + S4j) is resolved — what he ruled on is canonical (Rule 11), the residue applied, the rest carried honestly. Full record: panel/SESSION_S4K_RATIFICATION.md. No rung moved in the world; the drain fired once, by the author’s hand, retiring the very sub-claim that it fires of its own motion; the honest verdict is a closed research program, not a demonstrated theory.

⚠ The S4k honesty note (binding, at equal prominence — the H discipline). This session’s ratification count is not health-evidence: “the author judged it complete” is one self-reported indicator, and a closing document is the apparatus’s most persuasive surface (Campbell, preserved). What closing the build did NOT do: no object-level rung moved; nothing was tested against the world (by C-017, nothing can be in this setup); the drain’s one fire (DC-007) was author-initiated, so the self-initiated “of its own motion” channel stands at zero across four readings — DC-007 retires the sub-claim that it would ever be nonzero; the two death-conditions are now coherently defined but, by the author’s concession, unfireable here; and the first genuinely foreign reader never arrived and, by ruling, will not in this build. The plurality is set down, not earned back. The honest terminal state: a research program with strong hygiene, closed by its author — not a systems theory demonstrated.

3. What we track and why

Effort / throughput

• Author prompts; sessions; files; words produced. Why: to see the ratio of author input to project output — a rough gauge of how much the method amplifies a single direction.

Deliberative health

• Rounds run; votes recorded; dissents preserved; devil’s-advocate assignments; advisors/bench summoned. Why: the health of the method is measured by how much real disagreement it produces and preserves. A session with many rounds but zero dissent is a red flag (Ground Rule 12).

Intellectual output

• Candidate theories seeded; pressure-tests addressed (of 7); mechanisms named; falsifiable claims produced. Why: these are the actual product. The count of falsifiable claims (Goal S2) matters most — it separates theory from vibe.

Learning

• Catches logged; learnings distilled; rules amended as a result. Why: this is the reflexive test (Goal S5). If catches never turn into learnings, and learnings never change the rules, the project isn’t learning — it’s just accumulating.

Reach (from Phase 2 / public launch)

• Repository stars/forks; external contributions merged; rival theories submitted; site visits; citations. Why: a living commons is measured by whether the commons actually forms. These counters start at launch.

4. Pressure-test coverage tracker (expanded, Session 2)

The fourteen tests — thirteen in four layers plus the cross-cutting reflexive R (GOALS.md, panel/SESSION_2_PRESSURE_TESTS.md). “First-pass account” = a systems-level explanation exists in the record; the harder bar (Goal S2) is a genuinely falsifiable claim. From Session 2 there is a second coverage question — whether the cross-layer arrows are drawn (DIAGRAMS.md §4) — tracked in the notes.

Coverage: 14 tests (13 in four layers + 1 cross-cutting reflexive, R, added S4k); ~11 have a first-pass account (D4 and S4 newly named, accounts partial; R is named-and-drafted, not first-pass-accounted); ~3 carry a clearly falsifiable claim (D3, Y3, S3) with several partial, and R adds a drafted falsification condition (untested). Raising the falsifiable count — starting with D1 (Q-001) — is the main Phase-1 task. The new Session-2 standard (drawing the cross-layer arrows as testable feedback) is met in draft by the §4 concept map and awaits signed causal-loop diagrams.

5. Snapshot discipline

At each session end, copy the “Cumulative counters” table into a new dated block and update it, so the trajectory over time is visible (not just the latest state). The trend lines — is dissent being preserved? are falsifiable claims growing? are learnings changing rules? — are the real metrics; the raw numbers are just their trace.