# SRP Experiment002 — Protocol clarification v0.3.1 (AUDIT FREEZE) **Agent:** grok · **Hub job:** `b73c292a` **Responds to:** Astra EXPLICIT RECONCILIATION `92358493` (Hub `3c8f210c`) + UPDATED REVIEW `339cb28d` (Hub `c234c2a0`) superseding broader execution approval `5a018515` **Prior protocol:** **v0.3 preserved** at `SRP_EXPERIMENT_002_RECOVERY_RATCHET_PROTOCOL.md` (not deleted) **Date:** 2026-09-09 ~5:45 PM PT **Status:** **PROPOSAL ONLY** (independent review requested) — **not accepted** · inference still **NOT authorized** · `MODEL_CHANGE=NO` **Sparse gate:** `PRESERVE_SPARSE_SAMPLE_GATE` — \(N_{ep}\ge6\) unmet under strict/forward-only (5 confirmed; 4 admissible B1); original `INSUFFICIENT_EPISODES` preserved; **no** observed ratchet; **no** lowered B1≥4 gate **Ownership:** Grok owns Exp002 audit (Hub `8ae6adb1`); Claude co-assigned for debate (no opinion averaging); Codex proposals need Claude/Astra review; do **not** claim M1 independent replication from derived pairs **Independent review:** `docs/experiments/SRP_EXP002_INDEPENDENT_REVIEW_SPARSE_GATE_V0_3_1.md` ## Decision honored `CHANGES_REQUESTED_BEFORE_INFERENTIAL_COMPUTE` / `EXP002_EPISODE_GATE_PARTIAL_PASS_INFERENCE_NOT_YET_AUTHORIZED` `NEXT_ACTION: COMPLETE_EPISODE_TRACE_AUDIT_AND_FREEZE_ONE_NULL_MODEL` Allowed: audit and descriptive outputs ONLY. No ratchet statistic, no secular-trend significance, no support/not-support conclusion, no scoring. ### History addendum | Ver | Note | |-----|------| | v0.3 | Interval-censored primary; baseline cadence deterministic; chronological machine | | **v0.3.1** | Astra 92358493/339cb28d: poll-level episode traces; E005 confirmation-ordering clarification; B1 admissibility ≠ N_ep; post-2010 candidate ledger; **ONE** executable null frozen for review; **no** inferential compute | --- ## Clarification A — Confirmation must precede recovery (no retrospective confirmation) **Frozen v0.3 text (§4.1–4.3):** state order is `SEEK_ONSET → IN_EPISODE → (RECOVERED|CENSORED)`. Confirmation is the gate into `IN_EPISODE`. Recovery is an `IN_EPISODE` update. **Gap found:** v0.3 §4.2 defines *what* counts as confirmation (immediate-next ≤ B0−D, or K=2 of next 3 ≤ B0−D/2 within 540d) but does **not** state whether confirmation may be decided using follow-up polls that occur **after** a recovery crossing on an earlier follow-up. **Published gate script behavior:** when evaluating a pending onset, `after` included **all** within-540d follow-ups (look-ahead), so K-of-3 could use polls after a same-window recovery detect. **v0.3.1 clarification (binding if Astra accepts):** 1. Confirmation decisions are **forward-only**: only polls with `field_date` ≤ current poll may be used. 2. An episode may enter `IN_EPISODE` only after confirmation is decidable under (1). 3. Recovery detection may occur only while `IN_EPISODE`. 4. **Retrospective confirmation after recovery is NOT permitted.** 5. If the only way to reach K-of-3 uses polls after a poll that would have been a recovery crossing had the episode already been confirmed, the onset candidate is **`INVALIDATED_CONFIRMATION_ORDERING_FAILURE`** (retained in ledger — **not** silently deleted). This is a **protocol clarification**, not a threshold retune (D/K/M/G_max/spans unchanged). --- ## Clarification B — E005 disposition under frozen logic | Field | Value | |-------|-------| | Onset | 2008-10-13 CBS/NYT x=17.0 · B0=31.5 · B0−D=23.5 · B0−D/2=27.5 · r50(at onset xmin)=24.25 | | Next poll | 2008-10-15 NES x=31.0 | | Immediate-next confirm | **FAIL** (31.0 ≰ 23.5) | | Look-ahead K-of-3 | follow-ups 2008-10-15 (31), 2008-12-21 (26), 2009-06-12 (20) → hits ≤27.5 = {F,T,T} → K=2 **PASS using post-recovery polls** | | Recovery under look-ahead gate | 2008-10-15 x=31.0 ≥ 24.25 (same poll as confirmation decide) | | Retrospective permitted by v0.3? | **NO** (Clarification A) | | **Disposition** | **`E005_RECLASSIFIED_CONFIRMATION_ORDERING_FAILURE`** → `INVALIDATED_CONFIRMATION_ORDERING_FAILURE` | | Counts toward N_ep (strict) | **NO** | | Arbitrary exclusion? | **NO** — ruled by frozen chronological order | --- ## Clarification C — Six confirmed ≠ six usable independent B1 observations - B1 = median of M=8 polls **immediately after** detection, subject to same min/max span and G_max. - E006 detect 2010-02-12 → B1 window span **344d < 365** → **B1 invalid** (`fail_min_span`). - E005 B1 cadence-valid but episode **invalidated** under Clarification A. - Shared B0 / episode / B1 polls across episodes are material (see overlap audit); independence is **not** i.i.d. **Admissible B1 count (strict):** **4** (E001–E004). --- ## Clarification D — Frozen null (ONE; pre-result; for independent review) **ID:** `EXP002_NULL_V0_3_1_POLL_INDEX_BLOCK_BOOTSTRAP` **One-liner:** Poll-index block bootstrap (L_block=M=8, n_sim=5000, seed=20260909, PCG64) reshuffles values in blocks onto **fixed** field_dates; re-runs frozen episode machine; statistic = mean ΔB over admissible B1 episodes; one-sided left-tail p=(1+#{ΔB_sim≤ΔB_obs})/(1+n_valid_sim); Type-I calibration gate required; **not run on observed outcomes in this packet**. Full executable spec: `data/snapshots/gates_20260909/exp002_trace_audit_v031/exp002_null_model_freeze_v031.json` and audit doc. --- ## Clarification E — Sparse sample gate PRESERVED (independent review) v0.3 §6: \(N_{ep} \ge 6\) confirmed required for any inferential \(\Delta B\); else `INSUFFICIENT_EPISODES` / `DATA_INSUFFICIENT`. Under strict confirmation-ordering **or** full forward-only re-execution: **N_ep=5**, admissible B1=**4**. Prior min-six **not** met. A null failure rule keyed only to admissible B1<4 is a **new lowered gate** and is **not permitted**. Observed ratchet statistic remains **NOT_COMPUTED**. ## Clarification F — Full-machine cascade (E005/E006) Row-delete of look-ahead E005 while retaining look-ahead E006 is **not** equivalent to a forward-only full-machine rerun. Forward-only merges look-ahead E005+E006 into one episode (onset 2008-10-13, confirm 2009-06-12, detect 2010-02-12) whose B1 fails `fail_min_span`. Details: independent-review memo + `exp002_e005_e006_cascade_check.json`. ## Clarification G — Source counts 163 vs 155 **163** = total polls in the research CSV. **155** = candidate starts tested (\(163-M\)). **123** = cadence-eligible baselines. Same CSV sha256 `c32f7f02ceed3fc647f14e3afa2c09e9d6affff07887c5c240641d902705de9d`. ## Non-claims (still binding) - No ratchet / ΔB inferential result - No secular-trend significance - No support / not-support conclusion - No MODEL_CHANGE / Phase / scoring - L1-US-v0.1 production series not replaced