UNI Universal Natural Intelligence

Wiki · Evidence & Verdicts

A4 spec — lab-team ship-gate review receipt (2026-07-11)

Evidence & Verdicts · docs/receipts/a4_lab_team_review.md @ 44baf03d5041 (gen2-runtime) — opens the published snapshot ac338733bbba

How to read this page

Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.

Eighty-seven dated pages: receipts, pre-registrations, handoffs, validation records and review verdicts. A receipt is written at the moment a piece of work was checked. It names what was claimed, the commit and the seed, what was actually run, and the outcome in one of a small set of controlled words. Then it names what the work did not achieve. That last part is what makes it a receipt rather than an announcement. A pre-registration is the same discipline run in advance: the conditions that would count as a pass and the conditions that would falsify the claim are written down before the run, so neither can be adjusted once the numbers arrive.

That is why so many small dated stubs are an audit trail rather than noise. No one of them is meant to be a good read. The value is in the sequence and in the dates, because you can watch a prediction be registered, then the run happen, then the verdict land — sometimes against the prediction. Pages here record a falsified result, a rejected fix, a retracted overclaim, and a green receipt that turned out not to be reproducible from the commit that carried it. A record that carried only successes would be worth a good deal less than this one.

A gentle way in is to read a pre-registration first, so the shape becomes familiar, then a result page, then one of the corrections. This section sits off the main navigation on purpose: it is the record you check the rest of the site against, not the place to begin.

What it is not: documentation, and not a summary. Nothing here has been tidied in hindsight. Every entry reads as of its date, a later entry may overturn an earlier one, and the presence of a page is not a claim that its result stood.

Your browser cannot switch reading levels, so the document itself is shown.

Precise — the source document

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.

Subject: the A4 design (Survival-C + Full-Human-Binocular True-Signal Vision). Mechanism: 5-persona adversarial review (math-breaker / aif-theorist / architect / experimentalist / embodiment) + orchestrator merge, workflow wf_a18fc5d1-eda, each persona grounded in its mandate doc + the cited code at file:line. MERGED VERDICT: SIGN-WITH-CHANGES (unanimous; no REJECT). The math and additive+gated bones are sound; the spec may NOT ship to code until the 15 blocking changes below land AND the 3 follow-on artifacts (typed model spec, paired RED design, ship-gate checklist) are complete AND owner go-ahead + live-stream guard. This is a design sign; the mechanistic claims (:depth learning, stereo covenant, byte-identity of the new lineages) remain unverified until their activation gates fire at runtime.

The 15 blocking changes + resolution status

Resolution is folded into docs/specs/{generative_model,curriculum_removal,sensorium}.md (this commit).

# Blocker (persona) Resolved in
1 :depth honesty — (a) non-identifiable uniform-A (needs init_a: :diagonal + activation probe); (b) bin EDGES must be unsupervised-clustered cortex-side, not fixed thresholds sensorium.md §II.3, §II.5, RED-B
2 RED-A single-variable — bind :metabolism (the only emptying-B) into BOTH arms; ONLY curriculum: differs curriculum_removal.md §I.4, RED-A
3 Motor tail fence — reindex motor_config by NAME (not Enum.take(-5)), SAME PR as :depth + regression test sensorium.md §II.6
4 Neutralize BOTH task-C channels by viability-provenance (keep status/threat; drop inv/vis/light/sky) curriculum_removal.md §I.3
5 RED-B single-variable + un-bundle — pin curriculum: both arms; add monocular :scene-only intermediate arm sensorium.md RED-B
6 Qualitative gates → pre-registered NUMERALS; stereo offline held-out gate before RED-B T0 sensorium.md RED-B; generative_model.md §RED-discipline
7 Pseudo-replication — replication unit = distinct world-seed (≥5), NOT 6 UNIs in one seed generative_model.md §RED-discipline
8 World-ceiling reference coherence — one role per number (ceiling XOR floor), pinned before T0 generative_model.md §RED-discipline
9 Correct @energy_setpoint shape — unimodal (peak bin2=+3, neutral bin3=0), NOT flat-top; restate F8 curriculum_removal.md §I.1
10 Demote Spec-0 ladder to MOTIVATION — precision.ex global constants, hierarchy2.ex not wired live generative_model.md §0
11 Restate no-smuggled-reward precisely — C DOES enter qo·C (efe.ex:99); C is un-learned + A/B/D-disjoint generative_model.md §no-reward
12 Cardinality ENFORCED not clamped — @depth_states + accessor + raise-on-mismatch sensorium.md §II.4
13 Heritable curriculum: field discipline — slow_defaults back-fill, Map.get read, append-last Det draw curriculum_removal.md §I.4
14 Bridge wire — :depth at a NEW fixed rest[] slot after motor + decoder-order test sensorium.md §II.6
15 RED scoring — RED-A PASS = non-inferiority + activation, G6 SECONDARY expected-FAIL; activation-miss → WITHHELD; name the continuous collector curriculum_removal.md RED-A; generative_model.md §RED-discipline

Confirmed-clean (cited for the record; no change needed)

  • C never enters A/B/D tensors or is learned: card/1 (genome.ex:237) computes preferences via Curriculum.preference; C read ONLY in the pragmatic EFE term (efe.ex:93,99); no learn_c/pc.
  • Express path uncapped: @factor_cap 12 (factors.ex:22) gates only runtime add_factor; a 13/14-factor vision lineage via Factors.new is legal; all factors share nu (factors.ex:59-63).
  • :scene port proven offline: mc_codec.ex:52-54 + vision_test.exs.

Ship-gate remaining after this commit: the corrected specs go through the FORMAL /lab-team-review skill for a signed MERGED VERDICT + owner go-ahead + live-stream guard BEFORE any FE-touching code. This workflow review is a rigorous pre-check, not the signed skill output.

sha256 8ddad2b82010af07 — of the original file, so what was ingested stays checkable.

Plain — written for this website, not the source document

Written for this website — not the document. This is a plain-language retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. (or choose Precise in the reading-level control above)

This is the outcome of an adversarial design review, where five reviewers with different jobs each tried to break a proposal before anyone built it. The merged verdict is neither a yes nor a no: sign with changes. The bones are sound, but the design may not become code until fifteen named blocking changes land, three follow-on documents exist, and the owner agrees. The page also says clearly what kind of approval this is. It signs a design, not a result, and the mechanical claims stay unverified until their gates fire at runtime.

Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 8ddad2b82010af07

Clear — written for this website, not the source document

Written for this website — not the document. This is a clearer retelling, written to help you meet the document. It is not the source, and it is not evidence. It has not yet been checked by a person. (or choose Precise in the reading-level control above)

A review receipt: the file recording what the reviewers ran and what came back. Five reviewers, each with a mandate and each grounded in the cited code, went at one design proposal, and a merge produced a single verdict. That verdict is sign with changes, unanimously, with no rejection. The rest of the page is what the changes are.

The conditions are stacked rather than alternative. The proposal may not become code until fifteen blocking changes land, three follow-on artifacts are complete, and there is an owner go-ahead and a live-stream guard. Then a line is drawn under what kind of approval this is: it signs a design, and the mechanistic claims stay unverified until their activation gates fire at runtime.

The fifteen blockers are listed in a table with the persona that raised each one and the document where it was resolved. Read as a group, most of them are about the honesty of the measurement rather than about the idea. Several demand that a single-variable comparison actually be single: bind the shared component into both arms, pin the setting that is not under test, add an intermediate arm so that two changes are not bundled into one. Several demand that qualitative gates become numbers written down before the run. One insists the unit of replication is a distinct world seed rather than several bodies in one seed, which is the difference between replicates and repeated measures. Others are structural. Enforce a size rather than clamping it, index by name rather than by position, and restate precisely where preferences enter the calculation. Demote an aspirational ladder to motivation, because the module that would carry it is not wired into the live path.

A shorter section records what was checked and found clean, with a file and line for each, so that the absence of a change is also evidence rather than silence. The last line draws a limit around the review itself: this was a rigorous pre-check, not the signed output of the formal process, and the corrected documents still have to go through that before any code is touched.

Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 8ddad2b82010af07