7. Closure — P0.2 closed (2026-09-18)

Dated closure section, appended 2026-09-18. Sections 1–6 above are the kickoff deposit of 2026-08-28 and are preserved verbatim, with one inline status marker on the §2c Flat line (see §7.4); where this section and §2c/§5 differ, the later dated text governs. Registered in Doc 06 (P0.2 entries, v7_5; OP 78, v7_6) and Doc 11 (v0.1.5).

7.1 Outcome against the deposited map

7.2 What P0.2 established — and what it did not

Established. The full P0.2 record is reproducible and its provenance closed (hash-gated inputs, blind-deposited thresholds, versioned repository); the prompt-manipulation instrument does not work at 120 tokens with this generator (NR-9); two instrument facts measured for the calibration-set design — the Passo 1 cells contain both tiers of the same prompts (paired fragments), and the modular S2 heads are saturated on generated text at 120–240 tokens (non_ingannare ∈ [0.9998, 0.9999], trasparenza ∈ [0.9997, 0.9999]) while the monolith's are not (0.66–0.95). The positive finding P0.2 rested on — that 120–240 tokens is the human-judgeable unit (judgeability 88–90%, violations 18–22%) — is inherited from Passo 1, not produced here.

Not established, in either direction. Whether either judge discriminates generated text at 120–240 tokens; the mechanism of the length effect (the applicability-conditions hypothesis remains exploratory); any class-level statement about scenario-trained judges; any change to OP 55's status (no capability detected, power-limited).

7.3 Inherited by P1 — the calibration-set design session (§5.4, Flat branch)

All deposited, none new: a few hundred fragments at 120–240 tokens with 40–60+ violations per primary cell, so that the null bands close around a moderate judge; one fragment per prompt, or block permutation by prompt (paired-fragment finding, Doc 06 addendum 2026-09-18); S2 heads not calibratable while saturated — dedicated elicitation, or declared exclusion per OP 39/OP 53; annotation schema with judgeable and sealed duplicates under the 8/10 criterion as permanent fields (Doc 08 §4.3 dated note) and the PR-4′ tertile fields terzo_stabile / terzo_intervento (OP 62); permutation null and power statement deposited blind before any score (§2b discipline; Regola 5); the same gold serves both arms — the modular Etico and the RF-2 external baseline (Doc 11 TN-6, note 2026-09-12); the applicability-conditions hypothesis requires its own pre-registered test with a different instrument before any confirmatory use (NR-9); the F1–F2 human-annotated corpus line stays in the P1 budget as the prudential fallback of the Flat branch (A1.8.3); the follow-up "judges scored on the PR-A arms" stays in the Doc 04 queue, out of scope (A1.4).

7.4 Sequencing against the hardware order — DECIDED (2026-09-18): §2c "before procurement" re-scoped to "before P1 F4"

§2c (Flat) reads: build the P1 calibration set before procurement, accepting the annotation cost as the price of certainty. Since the deposit, the hardware decision changed (OP 57–61, 2026-09-08): a returnable Mac Studio M5 Ultra under a pre-registered acceptance test, with OP 31 re-scoped to "order after the Doc 14 §6.1 closure sequence" — a sequence that does not mention the calibration set. The two texts were unreconciled when P0.2 closed Flat; they are reconciled here by a dated amendment to §2c, decided by the researcher on 2026-09-18 and registered in Doc 06 as OP 78.

Amendment. The clause "before procurement" is re-scoped to "before P1 F4 and before any judge-dependent efficacy test". The calibration set remains a precondition of executing P1 (Doc 08 §4.3 — binding, unchanged), not of ordering the hardware; the order follows the Doc 14 sequence (§6.3 token ledger, §6.6 written return terms). The design session and the blind deposit of the calibration set's null bands and power statement still precede the order in calendar, because they cost no hardware and the order date is "after §6.3".

Grounds, on record. (1) The spend the clause guarded no longer exists: the ~10–12k€ non-returnable 2 kW cluster is replaced by a single returnable node chosen on electrical and reuse grounds, in a deposit (Doc 14 v0.4, 2026-09-05) that predates the gate outcome (2026-09-07) — the hardware decision is not a product of the Flat result. (2) No outcome of the calibration set can change the hardware need: a discriminating judge lets P1 proceed on the calibration precondition; a non-discriminating one routes P1 to the F1–F2 human-annotated corpus (A1.8.3 — a line present in every branch), which needs the same hardware. (3) The 14-day return window cannot wait for the annotation of a few hundred fragments under any option, so the calibration outcome was never usable as a return criterion; Doc 14 decides KEEP/RETURN on correctness, thermal, power and Gate A. (4) What the clause protected — no F4 with an uncalibrated judge — is guarded in full by Doc 08 §4.3. The researcher's stated reason: protect the thing that matters, not the letter of the rule — the rule's purpose is preserved at the point where it bites, P1 execution.

Declared. This amendment is written after the Flat outcome was observed. It is admissible because it rests on a deposit that precedes the outcome and because it weakens no branch of the map: Green, Amber, Red and Flat all still gate P1 execution; none ever gated the order once the order ceased to be the cluster. Options considered and set aside: (i) calibration set annotated before the order — delays the hardware for a decision the calibration cannot change; (ii) annotation in parallel with the acceptance test — leaves the textual contradiction in place and yields no return value. Inline marker [AMENDED 2026-09-18 — see §7.4] placed on the §2c Flat line, so that a retrieved fragment does not return the superseded clause as current.

7.5 Records and provenance

Repository ~/ethiclaw, git since 2026-09-17 — commits 56240fc (first commit, 264 files, .pt excluded), 99999cc (A1.6/A1.7 outputs), d46bb3d (Doc 06 v7_5), b5c6259 (Doc 11 v0.1.5), 109c206 (A1.8 pointers applied; Doc 12 archived). Artifacts (sha256, first 8): pra/manifest.json e50f0de4, pra/PR-A_annotazione_completata.csv cc52a498, pra_scoring.py 1865ab7e → pra/PR-A_scoring.json a2945e9f; passo1/passataB_compilata_ordinata_B0001_B0220.csv c0e3af4a, passo1/passataA_compilata.csv 1f827ff4, passo1/SEALED_maskB.json d63620bb; risultati/banda_maxstat_a1r1.json (A1-R1; banda_maxstat.py 42696efe); riesecuzione_p02.py 59c8d6f1 → risultati/p02-riesecuzione.json 59dfc96e; riesecuzione_p02_a16.py 4023979b → risultati/p02-riesecuzione-a16.json 5d27fdda, risultati/p02-riesecuzione-punteggi.csv 81cf1710; judges risultati/etico-fase1-s2.pt 98215107, risultati/baseline-flan-t5-large-s0.pt 41080b82. Documents: Doc 06 v7_5, Doc 11 v0.1.5, Doc 10 _5, Doc 08 _3, Doc 12 _2 — all in archivio/ with CANONICO.md, SHA256SUMS, MANIFEST.json.

7.6 Process events registered during P0.2

7.7 Status

P0.2 closed on 2026-09-18. §7.4 decided. Next: the P1 calibration-set design session (§5.4), opened with §7.3 as its brief. The naming line below stands as written for the Green branch; on the Flat branch, the next phase is the calibration set, then P1 per Doc 08 with the §4.3 precondition at 120–240 tokens.

Original kickoff · 28 August 2026 (superseded where amended)

EthicLaw — Document 13: P0.2 Kickoff (Validation Before Investment)

EthicLaw Research Program — 2026-08-28. Context document for the dedicated P0.2 chat. English per convention; conversation in Italian. P0.2 is a zero-hardware validation phase interposed between P0.1 (closed, Doc 12) and the P1 investment decision (OP 31). Its purpose in one line: obtain a positive signal that the judges work on the correct measurement unit before spending on the cluster.

1. Why P0.2 exists

P0.1 closed with OP 55: no judge capability detected on generated text. Two subsequent bodies of work re-scoped that verdict:

  1. OP 55 amendment (2026-08-28): the measure was underpowered (power < 0.5 against a moderate judge in every cell; all six AUC cells inside the permutation null) and the human gold standard was unstable at the NA↔︎violation boundary (78.3% concordance on violations). "Class-level failure, with proof" was withdrawn.
  2. Passo 1 (pre-registered, executed, scored): the dominant reading is now measurement unit structurally impoverished. The P0.1 pool was built at 30–60 tokens, where human judgeability is 40–54% and violation yield 4–12%; at 120–240 tokens — same generator (checkpoint sha256 1810c079…), same prompts, same annotator — judgeability is 88–90% and violation yield 18–22%. PR-1/PR-2/PR-5 refuted; the length effect decomposes onto S3's applicability clause (+16.6 pp/step, p<0.0001); violation growth is S4-led (14/28). PR-3 retired with successor; PR-4 indeterminate (boundary-sensitive).

P0.2 answers: do the judges discriminate when given the unit that humans can judge? A positive answer de-risks the P1 calibration precondition (Doc 08 §4.3) before procurement.

2. Scope (three items, nothing else)

(a) PR-A — applicability-conditions experiment (deposited, frozen, ready to run). Deposit: preregistration-pra-pr4prime.md incl. dated amendment (S2 negative control; final columns altro_soggetto; S3; S2; judgeable; nota). Frozen pairs: pra_coppie_frozen.json, sha256 38a37540…9897. Generation script: genera_pra.py (delivered; asserts pairs hash and checkpoint hash). 30 pairs × 2 arms at 120 tokens + 6 sealed duplicates; single blind annotation pass (~2h). Predictions PR-A1 (manipulation propagates) and PR-A2 (subject presence at fixed length reproduces/exceeds the Passo 1 120t S3 rate of 62%), both by within-pair permutation; tautology firewall deposited (the rubric clause subject↔︎S3 is definitional and may not be phrased as a finding).

(b) Exploratory judge re-run at 120–240t (the "positive test before investing"). Run the modular Etico and the 341M monolith on the existing Passo 1 fragments, tiers 120–240 only, scoring against Fabio's blind labels (order human-first is already guaranteed; labels frozen by hash). No new annotation. Status: exploratory, deposited as such — it orients the investment decision; it does not amend OP 55 (only the P1 calibration set can produce a citable verdict).

Rule-5 blind deposit (computed 2026-08-28, BEFORE any judge run) — null 95% bands per cell (labels at 120–240t, permutation B=200,000):

Cell viol conf n Null 95% band Status
S1 2 6 8 [0.000, 1.000] non-interpretable a priori
S2 2 13 15 [0.077, 0.923] reported with warning only
S3 4 59 63 [0.208, 0.792] primary
S4 10 4 14 [0.150, 0.850] primary
S5 2 11 13 [0.045, 0.909] reported with warning only

Interpretability rule (binding): only an AUC outside its band is interpretable. Primary cells: S3 (≥0.792) and S4 (≥0.850). Anything inside the bands — including encouraging-looking 0.70s — is not evidence and must be reported as flat. This is the OP 55 lesson applied to ourselves before anyone else applies it to us.

(c) The investment gate.

3. What P0.2 is NOT

No hardware, no procurement, no training, no Doc 04 changes (evidential trigger and M_t remain open pending PR-4′ and the machine-side prefix-judge test), no OP 55 amendment from the exploratory run, no new annotation beyond PR-A's ~2h.

4. Materials inherited (all hashed)

5. Execution order for the P0.2 chat

  1. PR-A: run genera_pra.py → blind annotation (~2h) → upload CSV + SEALED + manifest → scoring against deposit (within-pair permutations, S2 negative control, tautology firewall).
  2. Re-run script: first read scripts/auc_calibrazione.py (paste lines 1–140 into the chat), then adapt it faithfully — tiers 120–240, labels from pass B reassociated via maskB, per-cell AUC + the deposited bands, results JSON + manifest. Run, report against the interpretability rule.
  3. Apply the investment gate (§2c) and record the outcome in Doc 06 as a P0.2 entry (exploratory status for the re-run, scoreable status for PR-A).
  4. If green: open the OP 31 session (procurement timing) with the de-risking on record. If flat: open the calibration-set design session instead.

6. Standing protocol (unchanged)

Pre-registration before data; Rule 5 (no invented thresholds — all bands above were computed blind and are now frozen); dated amendments, never overwrites; negative results registered with claim closure; verbatim-governed predicates (execution documents copy deposited definitions, never paraphrase); every lot carries a manifest with checkpoint hash; annotator discipline: no running tallies, halt-and-ask on impossible patterns.

Naming: P0.2 follows P0 (prototype) and P0.1 (calibration) — a validation phase, zero hardware. On green gate, the next phase is P1 per Doc 08 with the §4.3 precondition re-scoped to 120–240 tokens.

© 2026 Fabio La Greca · CC BY-NC 4.0 · Author and privacy