Navigation
iteration 01 · 2026-07-14
DONE — PR #143 OPEN v4 cohort MINTEDprereqs ①②③ DISCHARGEDtolerances LOCKED

Bootstrap — mint the v4 cohort and pin the substrates

Branch feat/odb-ext/b5-10-bootstrapterrylica/mql5 PR #143 (loop never merges). All gates green: targeted cargo test · rust:check · check:all-gates · pre-push suite.

In plain terms

The campaign's ground rules are now frozen: a sha-anchored list of the 10 feature names (the v4 cohort), a locked per-feature table of the exact Python reference call and the numeric tolerance the Rust port must meet, proof that the frozen test fixture carries every input these features consume, and — new this campaign — a measured demonstration that every wrong way to build each feature produces a visibly different number on real data (so a mis-port cannot hide). Nothing here touches the production schema.

Pre-build prerequisites — all three discharged

#prerequisiteverdictmachine SSoT (sha256)
Agnostic TEST-B convergence (today-slate 107/108/98/80/78)PASS 5/5 540/540 windows each · #107/#108 EXACT (worst_abs 0.0) · #98/#80/#78 ≤ 4.3e-14 rel · converse control FAILs as designed (0.0056)agnostic_b5_10_result.json · 5c29316a…
Adversarial build-today decision — INLINE bounded probe (§6.S), no open-ended AttackerGO-ALL-5 7 rejected forks all breach on real windows · H4 bounds hold 300/300 · RG8 witnesses pinnedadversarial_build_today_b5_10_result.json · da46956d… (byte-identical on C&H re-run)
Substrate-map §2.1 + RG8 gap-policy pinningPASS all 5 substrate classes covered column-for-column · zero blockedsubstrate-map_impl_b5_10.md (fixture c106c51e…)

H2 — every rejected fork has measured teeth (frozen fixture, real windows only)

cardrejected fork (pinned convention)breachmax-abs
107NSRPS tie → latest occurrence (pinned: earliest)37/602.513e-02
107Rust f64::signum 0→+1 (pinned: np.sign 0→0)10/651.010e-02
108grammar size excl. rule S (pinned: incl. S)60/605.051e-03
98row-stochastic normalization (pinned: joint f=C/total)60/601.135e+00
80argument-swap sign negation (the #7 lesson class)60/601.403e+00
80ordinal ranks (pinned: average-tie rankdata)80/802.285e-03
78average-tie grid copula (pinned: stable mergesort ordinal)73/801.502e-03

All breaches ≥ 7 orders above the locked tolerance tier. Ties in logret/dur/spread are COMMON on the fixture (rank-convention forks breach even on untargeted windows); zero-sign bars exist at 5 dbps only — the #107 signum fork is witnessed there, never fabricated. Anomalies checked, not rationalized: the EURUSD@25 #107/#108 witness coincidence was verified benign first-hand (different arrays, coarse integer collision; #98 discriminates the pair).

Locked tolerance tiers (never relaxed to pass)

itercardtierH4 analytic bound (observed range on 60 disjoint windows)
02107EXACT (integer/(L−1))[0,1] (obs 0.2864 … 0.3434)
03108EXACT (integer/L)(0,1.02] (obs 0.49 … 0.6)
0498counts EXACT + ε 1e-9[0,1] (obs 0.0 … 0.0201 — zero-flux windows are REAL)
0580ε 1e-9 value-parity[−2,2] (obs −0.7013 … 0.3127)
0678ε 1e-9 value-parity[0,1.05] (obs 0.0050 … 0.0526)

Env pins: numpy==2.5.0 · scipy==1.18.0 · Python 3.14 · threads=5 (RG2 value-parity where float machinery is numpy-version-sensitive). Full table incl. clean-room attestations: python-ssot-and-tolerance-table_impl_b5_10.md.

Challenge and Held (§6.S bounded — no open-ended agents)

sidetallyoutcome
Attacker (bounded, 2m16s)6/6 PASSprobe re-run byte-identical; anchor triple-equality e5473634…; witnesses exact
Defender (bounded, 1m6s)7/8 + 1 RESOLVEDthe FAIL was a checklist-expectation error — cohort names are alphabetical by documented convention; build order lives in the JSON implementation_order (verbatim §1)

Conceded gaps: none. One self-caused drift gate (check:cli-spec) fixed in place before PR-open. Record: challenge-and-held-impl-b5-10-01-bootstrap.md.

What lands in PR #143

  1. v4 cohort: orthogonal_extension_cohort_v4.rs (10 None-stub slots) + JSON anchor + .sha256 + LOCKED_V4 anchor-pin test (652b5d0)
  2. §6.S subagent-discipline codification in the contract (47e523c)
  3. prerequisite ② decision: probe + machine JSON + decision doc (9c881e5)
  4. SSoT/tolerance table locked (57458e6)
  5. cli_spec.json regen for the two probe CLIs (5f4c1bf)
  6. Challenge-and-Held record (854894f)

No BarRow / schema.sql / feature_metadata.yaml changes — test_schema_contract green at every commit. Next: iter 02 (#107) stacks on this branch.