Gate G4 evidence (Phases 8 and 9: Tier 2)

Gate: G4, end of week 19 (Sun 21 Feb 2027).

  • Question (PHASES.md): are E3 and E4 complete or honestly reported, do prereg-v2 and prereg-v3 predate their results, and is this file written?

  • Stop rule: if Tier 2 is incomplete, the rest becomes future work and there is no Tier 3.

Prepared: 21 Oct 2026, by Claude Code (engineering lead). Sources: docs/CLAIMS.md (E-40 to E-58, P-17, P-18), reports/tables/e3_*.csv and e4_*.csv, and the reviews in docs/reviews/.

The gate decision is the researcher’s (CLAUDE.md). This file states which criteria the evidence meets and what is open.

Decision (researcher)

G4: PASS, decided by the researcher on 21 Oct 2026 after reading this file.

The researcher also reported:

  • Phase 9 hours: “10 to 18 h”, given as a range.

  • Cumulative hours: within plan.

  • Recent hours: at least 9 h per week over the last four weeks.

  • Capstone deadline: earlier than Sun 28 Mar 2027 (end of week 24).

The Tier 3 deadline rule therefore fails. Under the Phase 10 prompt, Phase 10 runs in polish mode (MODE B) with no discussion. The decision is recorded in reports/G5-tier3-decision.md.

Summary

All three G4 criteria are met on the evidence.

  • E3 and E4 are complete: E3 has 56 of 56 cells and E4 has 76 of 76, with none failed.

  • Registration order: both registrations predate their first results.

  • This file exists.

Tier 2 was run in calendar week 2 of the plan, against a plan of weeks 15 to 19, so no stop rule is near.

E4’s result is negative for the monitors. That is a reported result, not an incomplete one.

  • No monitor beat periodic retraining in any stratum.

  • In nine rows a monitor was slightly worse.

  • When refits happened made no difference, to within the 0.01 margin.

  • The monitors alarmed in most slots.

Criteria

#

Criterion (PHASES.md G4)

Evidence

Met

1

E3 complete or honestly reported

Both datasets have 28 of 28 cells, none failed. The Main cells have seeds 0 to 9 (the LAMDA seeds 5 to 9 were added on the researcher’s go, after the 5-seed rows were read, which is stated). Results and caveats are in chapter 7 section 7.5 [E-40 to E-47]

Yes

2

E4 complete or honestly reported

Both datasets have 38 of 38 cells, none failed, seeds 0 to 9 decided before any result [E-48]. Further evidence: the periodic arm reproduces E3’s random arm exactly [E-49]; the registered contrasts M-P and M-C are reported with intervals and verdicts [E-59 to E-62]; monitor quality is reported [E-63, E-64]; the results are in chapter 7 section 7.6

Yes

3

prereg-v2 and prereg-v3 predate results

prereg-v2 was created before the first results/e3/ file [E-40, G3]. prereg-v3 was created on 1d302ea at 20 Oct 2026, 17:52 IST; the first E4 ledger event is 20 Oct 2026, 20:24 IST [E-48]. driftwatch grid refuses E4 evidence cells without the tag

Yes

4

reports/G4.md written

This file

Yes

What was cut

Nothing registered for Tier 2 was cut. The registered reductions, stated in the addenda, are these:

  • E3 ran a reduced grid: two delays, two budgets, and the ablation and score cells at 5 seeds.

  • E4 ran one budget (1%), two delays, and one alarm level per monitor (alpha 0.01, frozen, never tuned). It has no label-based detectors and no monitor-triggered labelling.

These are design limits, not cuts made after results.

Reviews

  • Leakage, E4 code (docs/reviews/leakage-review-e4.md): no blocker and no major finding. Six minor findings and four notes are fixed or recorded, including:

    • a label-invariance test;

    • a complete cache key;

    • the matched control’s settings check;

    • refusal of evidence runs without the tag.

  • Statistics, E4 (docs/reviews/stats-review-e4.md): no blocker; six major findings, all on interpretation (saturated monitors and M-C, repeated BODMAS arms, reversals inside the margin, BODMAS drops in one window, replicate handling, missing chance baselines), all resolved; CLAIMS E-59 to E-65 supersede or extend E-50 to E-57.

  • Synthetic acceptance of the monitors (code behaviour, never evidence): reports/tables/e4_synthetic_checks.csv.

    • Every monitor alarms in the shift slot.

    • The no-drift false-alarm rates were stated before the tag (ADR-0010).

    • Two monitor designs were fixed before any real-data run.

Tier 3 go checklist (Phase 10; written to reports/G5-tier3-decision.md when Phase 10 starts)

Item

Rule

Value

Met

G4 passed

G4 PASS, with Tier 2 complete or honestly reported

PASS (researcher, 21 Oct 2026)

Yes

Cumulative hours within plan

at or under 151 h at the end of week 19 (plan Table 4.1)

within plan (researcher; exact figure not given)

Yes

Recent hours

at least 9 h per week averaged over the last four weeks

at least 9 h per week (researcher; exact figure not given)

Yes

Capstone deadline

not earlier than the end of week 24 (Sun 28 Mar 2027)

earlier than 28 Mar 2027 (researcher; exact date not given)

No

If any item fails, Phase 10 runs in polish mode (MODE B) with no discussion.

Open items for the researcher

  1. Gate G4 itself.

  2. Hours for Phase 9, and the three Tier 3 numbers above.

  3. The Holm floor. With 32 contrasts and 10 seeds, no E4 contrast can reach a Holm-adjusted p below 0.0625 [P-18]. The registered decision rule is the bootstrap interval. The floor was not written in addendum v3; chapter 7 says so.

  4. Reading the result. The practical-significance wording in chapters 7 and 8 is the researcher’s to write. Chapter 8 does not yet discuss E4.

  5. PR #4 (Phases 6 to 8) is still open. This branch (phase-09-monitors-e4) is stacked on it.