Build frozen Jev trading case sets and paired scorer - #1150

Scope

Build the frozen FPB-50, TFN and FOMC case sets and the paired Jev trading scorer. Both versioned IDs, jev-1.14.0 and jev-1.13.0, share one run. The combined review correction reserves costs before dispatch, uses the adopted native SDK, refuses scored startup outside the credential runner, and makes tied-label ECE consistent with the accepted prediction.

• Original build base: 0f4ee075b2ab3af1b099d5ca1ecb8cae02ee3f5b
• Current rebase base: fda24cfe436310658755a75e5b6b7865512c8df1
• Rebased head: b967bdc77ab9bbd06de4731b52cbd64d3b5b6a33; tree: c2983342b1510d3b2732cdd4d49f49a90aa6c7a4
• Lane: lane:shared — shared P1 tooling for the trading-owned comparison; trading-cc assigned the build through the brief and retains scored acceptance.
• Owned paths: P1 extension modules/guide; trading, SDK and provenance tests; the dated decision; evidence/artifacts/jev-trading-build-20261015/; corresponding evidence manifest rows.
• The CC-directed history cleanup first preserved the reviewed b118d0f2a051489bc5d70601b3448e3525325558 tree in one commit. The subsequent registry-only rebase onto the pinned main above retains one commit and the same lander change ID, eda95e3378c09a31fe345d9ca043bcd9418f95c5, with manifests/evidence.json excluded. The only conflict resolution was the registry; 35 branch rows were re-registered from merged bytes.

SOTA sources

• Private recipe, cited by name/path only: us-equities-trading/docs/decisions/2026-10-11-jev-harness-case-sets.md, decisions 1–10; docs/decisions/evidence/jev-harness-sources-20261011.md, JV1–JV52; coordination/trading-cc/specs/jev-trading-20261015/JEV-HARNESS-ADDENDUM-01.md. Private authority and reviewer digests are checked privately and never exported.
• TypeSafe SDK 0.7.2, _core/retry.py, _core/transport.py and _core/response_types.py. Reuses tools/sota-convergence/gap_crosswalk.py::call under the adopted transport decision: five attempts, only 429/529, 1/2/4/8-second backoff, no jitter/Retry-After/connection retries. Qualified public wheel SHA256: 0a961148187d52e18276ed7f2d02617cfac48e3b97673cb631a8749397d43d1e.
• TypeSafe API, retained public capture 6b760275f89341fe15cff80c28aacae94d7d0c0c99afcd987ec0bca680b5713a: Choice requests, model identity, probabilities, input usage and default context ceiling. No live currency check occurred in BUILD.
• scikit-learn 1.9.1, metrics/_classification.py, model_selection/_split.py, linear_model/_logistic.py; SciPy 1.18.1, stats/_resampling.py, stats/_binomtest.py; NumPy 2.5.3. Existing P1 locks and primitives are reused. The ECE correction follows p1_scoring.py::expected_calibration_error at the original build base: correctness is accepted prediction accuracy.
• Dataset revisions: Financial PhraseBank 8d3fe0c36d5feec6b3cc5e455b0fcb4820fb9964; TFN ccbe24de388e287beb92dd393a335c376b350ac3; FOMC 6b0283f55f0005a6d38d49f271d795c21fccc1a3. Parsing/source/split checks remain frozen.
• finBERT: retained Hub prefix 0bc2ac10dda6, captured by coordination/ns2604-coop/work-20261012/jev-screen-20261015/CLAUDE-ARM.md; full ID unresolved offline. Research comparison USE, cc-by-nc-sa-4.0, no redistribution and no licence gate. Nemotron 1B and CLM-8B retain their pinned native references. Haiku structured-output/pricing captures remain private JV20/JV21/JV52; its ledger belongs to its separate workload.
• Native scored-start parsing follows tools/credentials/credential_run.py::parse_args/run_command at the original build base. The guard inspects public source/process metadata, not credential stores; it prevents accidental startup outside the runner and does not authenticate a hostile same-user process.

Evidence-class table

Local commands run

Scientific tests and SYN used bubblewrap with live / read-only, private /var/tmp, fresh /dev and /proc, and --unshare-net --die-with-parent. Each scratch directory was removed.

<P1-python> -B -m unittest tests.test_jev_trading tests.test_jev_p1_casepack tests.test_jev_p1_scoring tests.test_provenance_digest_custody tests.test_published_digest_privacy tests.test_sota_sources_gate
# exit 0; 190 tests, one existing host-name opt-in skip
<P1-python> -B -m unittest tests.test_jev_trading_sdk tests.test_gap_crosswalk_sdk
# qualified offline SDK install; exit 0; 12 tests, no skips
<P1-python> -B blueprints/native-skill-practice/p1/jev_trading_mutants.py
# initial 106/107; isolated assertion correction plus final accounting-delta rerun: 108/108
<P1-python> -B blueprints/native-skill-practice/p1/jev_trading_run.py syn --out /var/tmp/syn
# exit 0; COMPLETE; 161 fake calls
python3 scripts/validate.py
python3 -B scripts/evidence_manifest.py --check
git diff --cached --check

Review corrections and fail-first evidence

• The GPT exact original-head witness reproduced 619 timed-out Haiku dispatches and $4.0235 of synthetic admitted exposure against a $3.3216 cap. The fixed timeout and malformed-usage cases each stop at 511 dispatches with $3.3215 reserved. Unknown actual spend stays unknown; first cap rejection latches across all conditions/sets.
• Jev preflight/running limits enforce $0.386568 per model and $0.773136 combined. Before dispatch, each possible SDK attempt reserves 65,536 input tokens under the vendor ceiling. Missing usage and ambiguous retries retain reserves. Above-planning usage stops the model; above-ceiling usage invalidates the upper bound. Per-arm costs also expose unknown retry charges; their regression failed first. Its fail-first regression fell from 112 dispatches to two.
• ECE's tied-label point regression failed at 0.4 instead of 0.6; the actual clustered interval failed at 0.2 instead of 0.4. Both now use accepted-label accuracy. Local P2-2 below records the remaining preregistered tie-contract issue. The actual model-minus-baseline argument-swap mutant is killed.
• Startup, private-digest custody, native transport and the finBERT prefix update have named fail-first tests. Unsupported historical failure/source/mutant hashes are omitted with explicit public projection metadata. Original private provenance stays private, and earlier measurements retain clear supersession.
• Decisions 5 and 10 now name implementation/tests for executable obligations. Review, scored execution and admission remain process/future obligations.

The retained external Jansen 8-K comparison supports the existing joint F1/Brier BEATS rule; it is not this harness's measurement. ai-hedge-fund and QuantDinger remain rejected for order-path use. T0–T10 are proposed readiness inputs requiring a separate sealed addendum; T8/T9 retain separate admission requirements.

Decision record

2026-10-15 Jev harness build records the layer resolution, sources, alternatives and overturn condition. JEV-TRADING.md maps each record decision to code/tests. Validation retains the correction evidence and verification limits.

Registry rebase validation

The registry-only rebase passed python3 scripts/validate.py at 12,092 registered hashes, the sorted-manifest check and the 19-test merged provenance suite. Native lg_change_id is eda95e3378c09a31fe345d9ca043bcd9418f95c5 before and eda95e3378c09a31fe345d9ca043bcd9418f95c5 after, excluding manifests/evidence.json. git merge-tree --write-tree returned 0 against both the pinned base and the observed newer main 39b3fcaf5bcb96c8a672fb6198972dffd3d9937a. Branch-history private-value match counts remain zero for the record, JV evidence, addendum, brief and private repository pin; no values are published.

History cleanup

The original commit ea96de7aceb6088314753a94dc88b7232d1cc5f0 may remain reachable by SHA on GitHub until garbage collection. A purge is an owner-only GitHub Support request; this branch rewrite does not claim server-side purge.

Follow-ups

• REQUIRED-BEFORE-SCORED-RUN — separate budget follow-up PR: at jev_trading_budget.py:42 and jev_trading_budget.py:59, settle holds when attempts resolve, retain unknown outcomes until reconciled, and test that a model is not stopped below the real cap. Current holds can trigger CAP_STOP after about 28 unknown outcomes or 135 retried attempts while charged usage remains below the cap. The CC permits landing this build but blocks every scored run until this follow-up is accepted.
• CC P3 — over-planning usage: jev_trading_budget.py:82 discards an answer above 1,000 input tokens and stops its model. Reconcile that behavior with the planning assumption and actual cap accounting.
• CC P3 — finBERT revision: jev_trading_baselines.py:24 and jev_trading_baselines.py:43 enforce a 12-hex prefix. Resolve the full Hub revision, change the pin and verify cached weights/config. Older captures do not establish the new revision.
• CC P3 — installed SDK attestation: jev_trading_custody.py:119 compares the supplied wheel pin to a constant; jev_trading_run.py:100 checks the installed version. Verify the installed distribution against the qualified wheel before treating this as installed-byte attestation.

The GPT fix1 read is ACK with these five P3s; the patched local review reports the three P2s that follow. These are recorded follow-ups, not fixes claimed in this body edit.

• GPT fix1 P3-1 — private filename prefixes: jev_trading_custody.py:99 exports the private basename, while jev_trading_custody.py:23 recognizes only full-length digests. A synthetic filename retained an eight-character content-digest prefix. Project stable authorization/reviewer role names or suppress verified prefixes too, with a synthetic basename regression. No current protected-data disclosure was identified.
• GPT fix1 P3-2 — interrupted Haiku accounting: jev_trading_run.py:455 writes exposure to the final report, but the partial receipt at jev_trading_run.py:470 lacks it. The interruption probe produced a partial journal without a report or Haiku snapshot. Retain known/reserved exposure and an interrupted-call marker, then reconcile with the owning ledger before another attempt. Local P2-3 below specifies the corresponding receipt repair.
• GPT fix1 P3-3 — exactly 1% failures: JEV-TRADING.md:139 and jev_trading_scoring.py:80 use “exceeds 1%”; addendum A3 also says “below 1%.” The frozen sizes 1,454/2,388/496 cannot realize exactly 1%. Resolve the governing wording before extending populations.
• GPT fix1 P3-4 — historical manifest locators: repair the older references at langgraph.json:40, recheck-verification.json:109 and gpt6-cross-family-review.md:5 with historical revision context. These references already missed their subjects at the merge base; they are not new PR regressions.
• GPT fix1 P3-5 — missing primary captures: the SDK/vendor-limit claims at source-review.json:7 and source-review.json:17, screen/external-comparison record at screen-inputs.json:4, and private recipe references at JEV-TRADING.md:204 lack some primary bytes in the reader packet. Supply pinned SDK, model/limit, Haiku, finBERT, screen/comparison and related dataset/source captures through the co-op custody path, then reproduce the native checks. Missing captures are verification gaps, not proof that the claims are false.
• Local P2-1 — SCORED-content binding: jev_trading_run.py:174 checks frozen row IDs and a source-pin attestation, but does not bind the supplied labels/text to verified source content. A relabelled TFN scored row was accepted while its ID/date hashes stayed unchanged. Compare pack manifests against frozen content hashes or bind packs to verified source bytes before dispatch, with relabelling/preprocessing refusal tests.
• Local P2-2 — preregistered ECE tie contract: the patched review identifies decision 6's sorted-label argmax convention as still binding under ADDENDUM-01 A4. The accepted-label calculations at jev_trading_scoring.py:54 and jev_trading_scoring.py:104 give 0.6 for the reported tie instead of the frozen convention's 0.4. Restore that convention in both point and bootstrap statistics, or record an authoritative amendment before SCORED execution. The generic P1 correctness parameter does not itself amend the experiment.
• Local P2-3 — partial-receipt Haiku exposure: jev_trading_run.py:470 retains only Jev budget state after an interruption/fatal error. Initialize the Haiku budget before protected execution and always retain its known spend, reservations, workload and cap in the final receipt. Test interruption after dispatch so the outstanding reservation survives even when no report is produced; preserve owning-ledger reconciliation.

Independent verification gaps remain: fix1 could not execute the full nested-namespace mutant campaign or native SDK tests without its qualified package/captures; this host's 108/108 mutant and 12 SDK results remain separately recorded. Broader reader portions had archive, socket, native-scanner and historical-Git limitations. The full1 WSL flake was not reproduced in fix1's supplied genuine-worktree result; that does not establish all merged/native acceptance.

Native weights, scored Jev/Haiku/CLM, masking/auditor selection, acquisitions and paid cost/p95 remain NOT_RUN. CC DELTA is PASS and GPT fix1 is ACK; their read carries to this head through the verified tree/change identity. The land cue goes to the CC after CI is green at this head, with trading-cc retaining acceptance. Scored execution remains NOT_RUN, belongs solely to trading-cc through credential_run, and is blocked by the required budget follow-up. BEATS confers no lane admission.

Host evidence

Not applicable: no evidence/hosts/ change or platform-status promotion.

Checklist

• No new workflow or scripts/*.py; source-policy grant requirements are unaffected.
• No secrets or private authority/reviewer digests in the current public files/body; no new paid hosting/subscription.
• Scientific runs used fake transports; no provider/network call or credential-store/environment-file read.
• Current evidence is hash-listed; historical projections and supersession are explicit.
• Peer-owned files/worktrees preserved; scientific scratch uses /var/tmp and is removed.
• Repository no-reply identity; history cleanup explicitly directed by the CC, with a lease pinned to the reviewed old head.
• GPT fix1 ACK and CC DELTA PASS recorded; carry verified by tree/change identity.
• CI green at this head and CC/trading-cc landing acceptance complete.
• REQUIRED-BEFORE-SCORED-RUN budget reconciliation follow-up accepted.