Universal Natural Intelligence

This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.

Adverse results, refuted predictions, partial runs, not-run cells, and blocked externals. These are first-class evidence. They are reported alongside every later result, never instead of one.

1. Retained adverse results

id	result	status
ADVERSE-LOGNORMAL	A simple lognormal (M2) out-predicts the two-timescale UNI mixture (M3) on held-out data under motor-equal NLPD, in both cohorts. Event-pooled gap 0.03687 / 0.03872 reproduces B2.	RETAINED — HEADLINE. Not superseded by any later work. Predictive superiority is never promoted to mechanism.
ALL-CONTRASTS-INCONCLUSIVE	All 8 motor-equal M3 contrasts are `INCONCLUSIVE` at 19 holdout motors. B4C04 shows this is seed-stable across 5 seeds × 2 replicate counts × 2 rules × 2 cohorts.	RETAINED. A sample property, not a seed artifact. Underpowered is not equivalence.
RULE-DISAGREEMENT	M2 ranks first under NLPD and near-last under CRPS-seconds.	RETAINED. Scoring-rule dependence is real.
M7-MEAN-ONE-FRAGILITY	The frozen M7 uniform-grid mean-one trapezoid mis-integrates the near-zero `y^(a-1)` singularity (0.9999942). The property holds exactly (analytic weight sum 1 to 1e-15).	RECORDED VERBATIM. Tolerance not loosened.

2. Refuted frozen predictions (B4)

cell	prediction	observed	outcome
B4C03	STABLE	M8 rank 2→3 at +1 bandwidth grid step	REFUTED
B4C06	BOUNDARY-STABLE	top-3 reorders under drop-longest	REFUTED
B4C08	LOMO-STABLE	2 non-primary contrasts flip under single-motor removal	REFUTED
B4C10	UNIDENTIFIED_OR_WEAK	U2/U3/U4 all OK at 100/2000	REFUTED_PARTIAL — being re-run at full N
B4C11	PROFILE_FLAT_OR_WEAK	U2 OK (full contract); U4 OK at 30/2000	REFUTED_U4_PARTIAL → U4 WITHDRAWN (D1)

A refuted prediction closes as refuted and stays visible.

3. Withdrawn claims

claim	reason	route back
B4C11 `U4_OK`; τ CI `[0.17658, 0.27020]`	D1 cluster-collapse bootstrap: 80 draws → 46 groups	corrected C11 full N=2000
`RESOURCE_BOUND` justifications for C01/C02/C10/C11	D2: overstated 17–29× vs measured	reclassification issued; runs scheduled
Track D resolution floor ≈0.064 nats	wrong narrowest contrast (M2 vs actual M4) + D7	corrected floor ≈0.042 nats

4. Not run / blocked

item	status	why
B4C01, B4C02, B4C09 (original)	`NOT_RUN` historical	recorded `RESOURCE_BOUND`; justification since withdrawn (D2). C02 now running; C01 queued; C09 still `NOT_RUN`
B4C06 `analysisStartIndex ∈ {3400, 3600}`	`BLOCKED_EXTERNAL`	raw MAT archive absent
Leave-one-study-out, leave-one-condition-out	`NOT_RUN` by construction	single study; no separable condition label
Raw archive re-derivation	`NOT_LOCATED_RAW_ARCHIVE`	`data/remodeling_data.mat` absent
`P4` transfer, `P7` replication	`NOT_ESTABLISHED`	no independent dataset — not closable by modelling
`P5` intervention	`NOT_ESTABLISHED`	no perturbation data — not closable by modelling
G-side biological policy selection	`DESIGN_ONLY_UNTIL_INTERVENTION`	passive dataset; action set structurally empty
Prospective mark-process claims	`RETROSPECTIVE_ONLY`	D5 — holdout mark channel burned

5. Self-inflicted process negatives

Recorded because a laboratory that hides its own procedural failures cannot be trusted with its scientific ones.

id	what	cost
D5	A subagent brief with no split boundary caused a read of the held-out mark channel	Permanent within this dataset. Mitigated only by the channel having had little resolving power (contrast `NOT_ESTABLISHED`, sign flips with the smoothing constant)
Claim-guard self-reference	Guard scanned its own phrase dictionary (19 false hits); then its own FAIL report (124)	Fixed via documented exclusions
Claim-guard cross-line masking	A `"forbidden"` on a previous line masked a real claim on the next	Fixed: negation must be same-line. Would have passed a genuine parity claim under a forbidden-wording table
Test conflated "small" with "unresolvable"	A uniform paired offset is genuinely resolvable; the test asserted otherwise	Split into two tests; distinction documented
TDD ordering not strictly red-then-green	D1/D3 corrected paths were authored with their tests	Disclosed in `H-AIF-G4-RUNNER-FIX-REPORT.md` §5 rather than presented as compliance

6. Carried Phase-1 limitations (must reach the final ledger)

D1 honest headline is 20/23 credited, not 22/23. AC4 FAILS.
D1A10B_LATTICE_SINGLE_BOND_DOUBLE_COUNTED: real detection, no strict prospective ancestry → prospective status NOT_ESTABLISHED.
Live stator integrality remains NOT_ESTABLISHED.
Mojibake in three Unicode diagnostic regexes in the corrected D1 runner (do not use it to create new evidence).
package-manifest-addendum-v2.json scope limitation (lists 6 artifacts, not itself).

7. Repository-level verdicts (unchanged, not authored by this program)

cross-study X16_FULL_BIOLOGICAL_PARITY = FAIL
  X06, X11 FAIL; X10, X12 NOT_ESTABLISHED; X13, X14, X15 BLOCKED_EXTERNAL; X01-05, X07-09 PASS
science-gates overall = PARTIAL_PARITY_ONLY  (G03, G05, G06 FAIL)
observed-experiment H1/H2 SUPPORTED_WITHIN_PROTOCOL; H4 INCONCLUSIVE

Gates move only by real evidence, never by relabeling.

sha256 d1e52519a00bae8a — of the original file, so what was ingested stays checkable.

Negatives and Partials — preserved, never absorbed

How to read this page