Skip to content

Add a TypeSafe-backed branch policy jev_api behind the BranchPolicy seam that asks TypeSafe Jev to choose among the local JevPolicy's offered actions from ids and numbers only, validates every answer, - #147

Merged
sykosyber merged 2 commits into
mainfrom
syberlabs/f138ebd0
Oct 28, 2026

Conversation

@sykosyber

@sykosyber sykosyber commented Oct 28, 2026 •

Copy link
Copy Markdown
Member

Objective: Add a TypeSafe-backed branch policy jev_api behind the BranchPolicy seam that asks TypeSafe Jev to choose among the local JevPolicy's offered actions from ids and numbers only, validates every answer, falls back to the local JevPolicy on a timeout, an invalid answer or an unavailable API and switches to local after three failures in a row, records every decision, and adds the evolve_jevapi benchmark arm with network-free tests over recorded responses.

Class: R
Agent-platform: claude@opus-5.5
Admitted-by: sykosyber
Reviewed-by: rules@2 (critic check critic.rules: high 0, medium 0, low 0)
Evidence: tested

Accepted in SyberLabs thread f138ebd0-ef17-4a31-8501-79ae897f0042 as candidate c1 (commit 608b9509525c36d1de6648ce0efc00c5017aedeb).
Host checks on that tree: conformance passed, critic.rules passed, scenarios passed, tests passed.

Reviewed-by: claude (independent verifier session, SyberFactory workflow)

… offered actions from ids and numbers, validated, with local fallback; evolve_jevapi arm; recorded fixtures

SyberLabs-Thread: f138ebd0-ef17-4a31-8501-79ae897f0042
SyberLabs-Candidate: c1
SyberLabs-Provider: claude@opus-5.5
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Oct 28, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-10-28T09:57:03.594019Z 608b950 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 608b950952

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread syberlabs/jev_api.py


def _number(value) -> bool:
return type(value) in (int, float) and math.isfinite(value)

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Handle oversized JSON integers without aborting fallback

When a TypeSafe response contains an integer too large to convert to a float in probabilities or confidence, math.isfinite raises OverflowError. _ask only turns JevFailure from validation into a local fallback, so this escapes decide and aborts the search instead of recording jev_invalid and taking the local action. Treat this conversion failure as invalid input.

AGENTS.md reference: AGENTS.md:L210-L211

Useful? React with 👍 / 👎.

@sykosyber

Copy link
Copy Markdown
Member Author

Scope: owner direction in session 2026-10-28 ("retake the JEV evo work but with actual typesafe jev integration with key TYPESAFE_API_KEY in system vars and continue from the research we were doing on it"). Packet jev-api, thread f138ebd0, candidate c1 (head 608b950952, tree de6669846b), base main at 0b03040fe0. Built through contract v5 (thread_setup_v5.py --base main --max-dollars 5), class R, unattended: the first syberlabs run --once checked c1 acceptable and the second accepted it. No Anthropic model was called. TypeSafe's Jev API was called 28 times, for discovery only (below); the tests make no network call.

What changed

  1. TypeSafeJevPolicy (new syberlabs/jev_api.py, name jev_api, revision "1") sits behind the existing BranchPolicy seam.

    • The local JevPolicy builds the offered actions and every stop. Jev chooses one offered action through a single choice question, next_action.
    • Stops (a passing individual, no room, nothing offered, the floor, the generation bound) stay the local rule's and ask nothing. Jev cannot end a search, extend one, or add an action.
    • A decision with only one offered action is taken without a call.
    • Selected by construction only: EvolutionaryProvider(..., policy=TypeSafeJevPolicy(...)). Stdlib urllib, no new dependency.
  2. The request carries ids and numbers only.

    • The state holds slots (never candidate ids), scores, the remaining budget, OperatorStats, and the local estimates of each offered action.
    • The instructions and criteria are module constants, and the option keys are action kinds.
    • The key is read from TYPESAFE_API_KEY at call time and sent only as the Authorization header.
  3. Validation before an answer can become an action:

    • status 200, and one strict JSON object (a duplicate key or NaN is refused);
    • model exactly the pin, jev-1.13.0. The constructor refuses jev-latest, jev-preview and anything that is not jev-X.Y.Z;
    • answers holds exactly next_action, of type choice;
    • the choice is one of the options sent;
    • the probabilities cover exactly the options sent, each finite in [0, 1], summing to 1 within 0.02;
    • the choice carries the top probability;
    • confidence is finite in [0, 1], and usage holds non-negative integers.
  4. Fallback. A failure records fallback: {reason, detail}, with detail from a closed list, and the local JevPolicy action is taken. The reasons:

    • jev_timeout: a timeout (30 s by default, at most 120 s);
    • jev_unavailable: no key, a non-200 or a transport error;
    • jev_invalid: a body that fails validation;
    • choice_outside_options.

    No response text, header or exception message is kept, and nothing is retried. Three failures in a row set switched_to_local, and the rest of that search runs local with no call. A success resets the count.

  5. Record. Every decision's trace entry adds the following to JevPolicy's audit fields:

    • source: jev_api, local or rule;
    • model, answers_digest, request_digest, latency_ms, fallback and failures;
    • local_action, the shadow local choice;
    • on a Jev decision, also jev: {choice, probabilities, confidence, usage, dollars}.
  6. JevPolicy.decide is split into a pure _plan and a _take. The traces of fixed, empirical and jev are byte-identical to main: digests were pinned on 0b03040 before the change and are asserted.

  7. New evolve_jevapi arm in benchmarks/model_vs_patch.py.

    • It is not in the default arms, so a default run makes no TypeSafe call.
    • Its row adds jev_api: calls, decisions by source, fallbacks, agreement with local_action, input tokens, Jev dollars and median latency.
    • With --write, it writes the decision entries beside each exported journal.
  8. Docs and gaps.

    • docs/EVOLUTION.md: the policy, the live facts, and the pre-registered measurement plan (below).
    • docs/AUTOSKILL.md: the adapter status, and three corrections from the live calls.
    • spec/GAPS.md: entry 61.

Files (22):

  • new: syberlabs/jev_api.py, tests/test_jev_api.py, and tests/fixtures/jev_api/*.json (15 recorded bodies; request ids, cookies and headers scrubbed);
  • changed: syberlabs/evolve.py, benchmarks/model_vs_patch.py, docs/EVOLUTION.md, docs/AUTOSKILL.md, spec/GAPS.md.

No change to the kernel's schema, event kinds, golden traces or cli.py.

Phase 1: live discovery (28 of the 40-call cap; about $0.0006)

The calls used synthetic Situation states (ids and numbers only) and ran from Windows. The redacted transcripts stay outside the repository.

Fact Observed
Response shape {model, answers: {q: {type, choice, confidence, probabilities}}, usage: {input_tokens, output_tokens}}. The body has no request id; the header x-typesafe-request-id carries it.
Model jev-1.13.0 is served as jev-1.13.0. jev-1.14.0 (released 2026-10-14) and jev-latest are served as jev-1.14.0. jev-1.13 returns 400 Unknown model. GET /v1/models lists only the aliases.
Latency 20 answered calls: 142 to 338 ms, median 170 ms; about 705 to 760 input tokens each.
Probabilities Two decimals, summing to 0.99 to 1.00. confidence = (p_max − 1/k)/(1 − 1/k) on 20 of 20 answers. A separate prototype's calls the same day found the same formula. The policy records confidence and does not threshold on it.
Repeatability The same request sent three times gave mutate at 0.58, 0.56 and 0.53: not deterministic.
Errors Bad key: 401. No key: 403. Unknown model: 400. NaN: 400. Missing criteria: 422, which echoes the input back.
Limits No rate-limit or usage headers. TypeSafe's docs give per-organization limits in tokens and requests per second, 429 and 529, and 64k tokens per request.
Price $0.042 per million input tokens; output is free.
Data terms A Data Processing Agreement, and a privacy policy that says no training on user data. Retention periods are not stated on the legal page; zero data retention is enterprise-only.
Behaviour On these states Jev always chose mutate or evaluate, and gave crossover 0.02 to 0.15.

Commands and results

  • Targeted, local (Windows 11, CPython 3.12.6). Modules: test_jev_api, test_jev, test_evolve, test_gaps, test_providers, test_package, test_eval_core, test_model_comparison. Result: Ran 140 tests in 448.246s, OK. After strengthening one purity test, test_jev_api and test_gaps gave Ran 30 tests, OK.

  • Mutation check (on a copy of the clone, since removed): 16 mutations of jev_api.py and evolve.py.

    • 15 caught. They are: model unchecked, sum unchecked, key set unchecked, choice outside the options accepted, argmax unchecked, no switch, no reset, candidate ids sent, exception text kept, HTTP status mapped to transport, Jev allowed to stop, local action not taken on fallback, timeout unmapped, _plan taking state, and JevPolicy tie order (caught by test_jev).
    • 1 survived: "non-finite accepted". It is equivalent: the strict parser already refuses NaN, and the [0, 1] range refuses ±inf.
  • Full suite, GitHub Actions, scratch branch syberlabs/f138ebd0-ci (2cd52ca, tree de6669846b, identical to the head's): https://github.com/SyberLabs/SyberWork/actions/runs/40665383090, success.

    • test (3.11), test (3.12), test (3.13), G8 and G8 strict: Ran 1180 tests, OK (skipped=2).
    • windows: Ran 1180 tests, OK (skipped=25).
    • compose: success.
  • Bite, GitHub Actions, scratch branch syberlabs/f138ebd0-bite (60148fd: only tests/test_jev_api.py and the fixtures, on main): https://github.com/SyberLabs/SyberWork/actions/runs/40665382783, failure, as intended.

    • Every leg: Ran 1155 tests, FAILED (errors=1). The one error is ModuleNotFoundError: No module named 'syberlabs.jev_api'.
    • That is an import error, not one failure per test. The per-property bite is the mutation check above.
  • PR CI: all 25 checks on Add a TypeSafe-backed branch policy jev_api behind the BranchPolicy seam that asks TypeSafe Jev to choose among the local JevPolicy's offered actions from ids and numbers only, validates every answer, #147 at 608b950952 pass (gh pr checks 147).

    • The first attempt of run 40667906908 failed one leg, windows, and with it G8. The cause was test_resolution.AuthoritativeResolution.test_site_task_needs_authoritative_choice_and_owner, which hit ConnectionAbortedError [WinError 10053] on the case study's loopback HTTP.
    • That test is untouched by this change and passed on the same tree on the scratch branch.
    • The failed jobs were re-run once, and attempt 2 passed.
  • python -m conformance.run: matched 23 golden traces.

  • python benchmarks/run_scenarios.py: scenarios: 11 replayed, exit 0, digest b41f7cbb… (unchanged).

  • python -m conformance.clean_install: exit 0, build-thread complete, evolve complete, 47.8 s.

  • Thread checks on c1 (22 files, 135,312 diff bytes):

    Check State Time Output
    conformance passed 13.4 s
    critic.rules passed 0.5 s "no findings over 22 changed path(s)"
    scenarios passed 0.4 s
    tests passed 363.9 s Ran 1180 tests, OK (skipped=25)

    findings c1: no findings.

  • certify c1: result 608b950952 (tree de6669846b) on repo-change.v5, class R; four checks passed; chain head f4872499c9 at 14 events.

    • Signed with key 4b6ae2deea. Before signing, I checked that the public key of $SYBERLABS_CERTIFICATE_SEED equals the factory's test-integrity-2-74e7d4ac.cert.pub. The written .pub is byte-identical to it.
  • verify against that .pub with --repo .: signature ok, repository ok (22 paths, diff digest match), journal ok (14 of 14 events). Verdict pass.

  • Publish. push_branch succeeded.

    • open_pull_request failed three times with effect_unknown, destination_error. A diagnostic wrapper that changed nothing else showed the cause: URLError SSLCertVerificationError … unable to get issuer certificate on the kit's GitHub API read. The failure was intermittent; the same read from the same machine succeeded in between.
    • Each attempt was settled with syberlabs recover (not_applied after the 120 s window) before the next. The fourth attempt opened this pull request.
    • Verification was never disabled and no trust store was changed. The thread journal therefore has 69 events; the certificate covers the first 14.

Pre-registered measurement (not run; needs the owner's go through the lead)

This is in docs/EVOLUTION.md.

  • Arms: single, evolve, evolve_jev (local) and evolve_jevapi.
  • Set-up: the four v2 fixtures, K = 6, Claude Sonnet 5.5 at medium effort; one repeat, then three.
  • Rule 1, carried over verbatim: local Jev becomes default if evolve_jev solves at least as many as evolve with at most half the dollars.
  • Rule 2, new: TypeSafe's Jev replaces the local chooser only if evolve_jevapi solves at least as many as evolve_jev, with Anthropic plus Jev dollars at most evolve_jev's.
  • Not measured, if: more than 10% of the asked decisions fell back, or any run switched to local.
  • No difference claimable, if: every asked decision agreed with local_action.
  • Uninformative, if: single solves every v2 fixture in repeat 1, as it did on 2026-10-21. Then no default changes and the owner picks v3 or v4.
  • Cost: ceiling about $2.0 to $2.5 of Anthropic spend (228 calls at most, at the v2 pilot's $0.0086 a call), and about $0.006 of Jev (about 135k input tokens over at most 180 calls). Budget line asked: $3 plus $0.01.
  • Not a SyberEval spec: ARM_KINDS in syberlabs/eval/spec.py is closed without evolve_jevapi (GAPS 61 (5)).

Claim ceiling

Implemented and tested: on CI (Linux 3.11 to 3.13, Windows) and locally, against recorded TypeSafe responses and stand-in transports.

  • The live API shape is observed, from 28 discovery calls.
  • Not measured: whether Jev chooses better than the local heuristic, on any fixture or with any model.
  • Jev's probabilities are a provider's signal, not evidence.

Known limitations

  • Decisions are not journaled by the host. PROVIDER_FACTS in build.py is closed to lookup. The decision entries live in the trace, and the arm writes them beside the journal. The smallest fix is a policy_decision key (GAPS 61 (1)).
  • Jev's dollars are outside max_dollars. At about $0.00003 a decision this does not move a bound.
  • Jev sees only the numbers the heuristic reads. That was a deliberate choice: no text leaves the process.
  • The pin jev-1.13.0 is not jev-latest (now jev-1.14.0). TypeSafe states no deprecation policy, so a retired pin would fall back on every decision and switch to local after three.
  • The bite is an import error, not per-test.

Overlap and numbering

Unresolved questions

  • Whether the owner wants jev-1.14.0 pinned instead.
  • Whether the measurement runs on v2 as asked, or on v3 or v4.
  • Whether the build owner adds policy_decision to PROVIDER_FACTS.

Next permissible packet

The paid measurement above, after the owner's go. Separately, a build-owner change that journals policy decisions (GAPS 61 (1)).

Never merged, never force-pushed, nothing pushed to main, no setting changed, no Anthropic call made.

@sykosyber

Copy link
Copy Markdown
Member Author

Independent verification: #147 (jev-api, thread f138ebd0, candidate c1)

Verdict: PASS. Mergeable as is; four doc corrections recommended; GAPS 61 must be renumbered by whichever of #146 / #147 lands second.

Head 608b950952 (tree de6669846b), base main 0b03040fe0, 22 files. Fresh clone; no Jev, model or GitHits call made; tests run with TYPESAFE_API_KEY removed from the process environment.

Findings, ranked

  1. Medium: spec/GAPS.md:123 entry 61 collides with Add the SyberEval safety gate, scorer and paired report, with the hidden_feedback spec field, hidden-check integrity re-hashing and an append-only spend ledger. Each is computed from the journal and t #146's entry 61.
  2. Medium (docs): the measurement plan's premise is not supported by the repository.
    • docs/EVOLUTION.md:198 says "The plan is written for v2 because that is the measurement the owner asked to continue."
    • But docs/MODEL_COMPARISON.md:172 records that v2 is uninformative at Sonnet 5.5 (16/16) and Haiku 5.5 (8/8). It also says the paid Jev-against-fixed comparison "goes to … the v3 fixtures … by the owner's decision".
    • So the plan pre-registers a run that its own "When v2 cannot decide" clause expects to be uninformative.
    • Either cite the newer owner direction, or point the plan at v3/v4. The handoff lists this as an open question; the doc states it as settled.
  3. Low (docs): docs/EVOLUTION.md:189 "gave crossover 0.02 to 0.15" is wrong on one recorded call.
    • The two-option answer (tests/fixtures/jev_api/choice_two_options.json, live call 09) gave crossover 0.37.
    • The range holds only for the 19 three- and four-option answers. The same claim is in the handoff's Phase 1 table.
  4. Low (docs): the cost ceiling counts three repeats, but the plan says "one repeat, then three".
    • The plan (docs/EVOLUTION.md:193) runs one repeat first, then three; the ceiling (docs/EVOLUTION.md:199) counts only three repeats.
    • If the first repeat is a separate run before --repeats 3, that is four repeats. It is 304 Anthropic calls, about $2.6, or about $3.3 with feedback. That is above the $3 line asked for.
    • Also, "at most about 15 asked decisions per run" is not a code bound. A Jev-chosen crossover uses none of K = 6, so asked decisions are not capped by K. The Jev cost stays negligible either way.
  5. Low: the 30 s timeout is a per-socket-operation timeout, not a wall-clock deadline.
    • syberlabs/jev_api.py:111 calls opener.open(request, timeout=timeout), which is urllib's per-operation timeout. A server that drips bytes could hold one decision longer than 30 s (reads are bounded by MAX_RESPONSE). DNS resolution is not covered either.
    • The docs (docs/EVOLUTION.md:170) and the docstring read as a total deadline.
    • Either state the semantics, or enforce a deadline around the call.
  6. Info: fallback.response_digest is recorded but not documented.
    • syberlabs/jev_api.py:339-340 adds it (SHA-256 of the raw response body).
    • It is not in the record list at docs/EVOLUTION.md:170-172 or in the module docstring.
    • It is a digest, so a 422 that echoes the input leaks nothing. The echoed input is the module's own question in any case.
  7. Info: the original exceptions are suppressed, not discarded.
    • raise … from None (syberlabs/jev_api.py:120-126) only suppresses display. __context__ still holds the urllib exception, and the traceback frames still hold the Request with the Authorization header.
    • This is harmless today: decide keeps only reason, detail, latency and digest, and drops the exception. It would matter only if a future caller logged the exception with frame locals.
    • Similarly, the closed-word check in JevFailure is an assert (syberlabs/jev_api.py:88), which -O strips. Every call site passes literals, so it holds.
  8. Info: the fixtures' recorded requests are from discovery calls with an earlier request shape.
    • They use population id: "c1", an earlier instructions text, and no options. That is not what TypeSafeJevPolicy.state() sends.
    • Only the recorded responses are used by the tests, and the ids are synthetic.

Per check

  1. Scope: OK.

    • The diff is exactly 22 files: jev_api.py, evolve.py, model_vs_patch.py, EVOLUTION.md, AUTOSKILL.md, GAPS.md, test_jev_api.py, and 15 fixtures.
    • Nothing changes in schemas, event kinds, conformance/ goldens or cli.py. --policy still offers only fixed|empirical|jev.
    • evolve_jevapi is in ALL_ARMS and not in ARMS (the --arms default).
  2. Refactor: OK.

    • _plan and _take reproduce the old decide/_stop exactly: same audit, same stop generations, same state mutations.
    • I recomputed the seven UNCHANGED trace digests independently on main 0b03040 with tests.test_jev.run. All seven equal the constants asserted in tests/test_jev_api.py.
    • conformance.run and the targeted modules pass (below).
  3. jev_api.py: OK.

    • Egress. The state holds only numbers, booleans, None, and fixed key words. Population entries are slots. A Mutate is sent as parent_slot, a Crossover as slots, an Evaluate as seeds_left, never ids. Request.test_the_state_holds_no_text… walks the state against a key allowlist with planted canary ids, paths and seeds.
    • Key. It is read only in http_transport, from TYPESAFE_API_KEY, and only into the Authorization header. A key containing CR or LF is refused. Redirects are disabled, so a redirect never makes a second send.
      • It is never stored on the policy. There is no dataclass in the module.
      • JevFailure holds only reason, detail, latency and digest.
      • The Rejected message is constant.
      • The Key tests plant a canary through http_transport and check it is absent from the trace, the policy, the signals, and the exception's str and vars.
    • Validation. Every listed item is present, in this order (validate, lines 156-201):
      • status 200;
      • size;
      • strict JSON (duplicate keys and NaN/Infinity refused; 1e400 overflows to inf, which the finite check then refuses);
      • an object;
      • model equal to the pin (the constructor regex refuses aliases);
      • answers key set exactly {next_action};
      • type choice;
      • choice is a string and within the options (choice_outside_options);
      • probabilities have the exact key set, are finite, are not bool, lie in [0, 1], and sum to 1 within 0.02;
      • the choice is the argmax (ties allowed);
      • confidence is finite in [0, 1];
      • usage.input_tokens and usage.output_tokens are non-negative ints (not bool).
    • Fallbacks. The four reasons are present, and each detail is a closed word.
      • A non-200 body is never parsed, so a 422 echo cannot reach a record. Only its SHA-256 can.
      • Three failures in a row set _local_only, and a success resets the count (tested).
      • request_too_large counts as a failure but not as a call.
    • Timeout. Default 30.0 s. The constructor refuses anything outside 0 < t <= 120. See finding 5 for the semantics.
  4. Fixtures: OK.

    • The 15 files contain no Bearer, ts-… key shape, cookie, CF-RAY or request id.
    • I extracted all 27 x-typesafe-request-id, 27 set-cookie and 27 cf-ray values from the live transcripts: none occurs in any fixture, in full or as a 12-character prefix. The two Authorization values in the transcripts are redacted.
    • Every fixture response is byte-equal, as canonical JSON, to its live transcript body: 15 of 15 matched.
  5. Docs. The live facts table matches the transcripts:

    • 20 answered calls, 141.8 to 337.5 ms, median 170.1 ms, 704 to 760 input tokens;
    • sums off by at most 0.01;
    • repeats 0.58, 0.56 and 0.53;
    • /v1/models lists aliases only;
    • status codes 401, 403, 400, 400, 422.

    The decision rules are as stated, with rule 1 verbatim from the existing text. The 228-call and $0.006 arithmetic is right. ARM_KINDS, PROVIDER_FACTS, GAPS 41, FixedPolicy revision 2 and SYBERLABS_MODEL/SYBERLABS_EFFORT are confirmed in code. The exceptions are findings 2 to 6.

  6. Tests: OK (local and CI below).

  7. Certificate: OK (below).

  8. Rebuild: OK (below).

Local (fresh clone, Windows 11, CPython 3.12.6, TYPESAFE_API_KEY removed from the child environment)

  • python -m unittest tests.test_jev_api tests.test_jev tests.test_evolve tests.test_model_comparison: Ran 105 tests in 427.807s, OK. The child process confirmed that TYPESAFE_API_KEY was absent.
  • python -m conformance.run: matched 23 golden traces.
  • The UNCHANGED digests were recomputed on a main worktree at 0b03040: all 7 match.

CI

  • Full run 40665383090 (syberlabs/f138ebd0-ci at 2cd52ca): the tree is de6669846b, identical to the head's.
    • Results: 5 × Ran 1180 tests OK (3.11, 3.12, 3.13 and G8 strict with skipped=2; windows with skipped=25), plus compose success.
    • 0 ERROR/FAIL lines.
  • Bite run 40665382783 (60148fd: only the test and the 15 fixtures on main; the tests are identical to the head's): five legs at Ran 1155 tests, FAILED (errors=1).
    • Every one of the five errors is test_jev_api failing to import: ModuleNotFoundError: No module named 'syberlabs.jev_api'.
    • G8 failed only as the aggregate.
  • PR run 40667906908.
    • Attempt 1: only windows failed (and G8 as the aggregate). It was Ran 1180, FAILED (errors=1), the single error being test_resolution…test_site_task_needs_authoritative_choice_and_owner. The cause was ConnectionAbortedError [WinError 10053] in case_studies/enterprise_procurement.py:329 urlopen against its own ThreadingHTTPServer on 127.0.0.1. That file is untouched by this PR.
    • Attempt 2: all 7 jobs passed. Each test leg was Ran 1180, OK.
  • gh pr checks 147: 25 of 25 pass.

Certificate and journal

  • syberlabs verify: --public test-integrity-2-74e7d4ac.cert.pub --repo . gave:
    • signature ok, ed25519 4b6ae2deea, thread f138ebd0, c1, result 608b950952, repo-change.v5, class R;
    • repository ok, with 22 paths and a matching diff digest;
    • journal ok, 14 of 69 events certified, head f4872499c9;
    • verdict pass, exit 0.
  • Receipt key: the receipt's .cert.pub is byte-identical to the factory .pub (sha256 1972eda6…).
  • Journal events 15-69:
  • Wording note: the handoff says open_pull_request "failed three times" and "the fourth attempt opened" it. The journal shows four failed attempts and a fifth that succeeded; one of those failures followed the denied proposal.

Rebuild preview

Reviewed-by: claude (independent verifier session, SyberFactory workflow)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sykosyber
sykosyber merged commit d85f3d2 into main Oct 28, 2026
25 checks passed
sykosyber added a commit that referenced this pull request Oct 28, 2026
…e-registered jev_api measurement to the (#149)

Objective: Close the independent verification findings on #147: re-target the pre-registered jev_api measurement to the v3 fixtures with four paid repeats and a sourced cost ceiling, correct the recorded crossover range, make the Jev timeout a wall-clock deadline per decision, document fallback.response_digest, and replace the assert in JevFailure with a real check.

Class: R
Agent-platform: claude@opus-5.5
Admitted-by: sykosyber
Reviewed-by: rules@2 (critic check critic.rules: high 0, medium 0, low 0)
Evidence: tested

Accepted in SyberLabs thread `7ca8f3c4-534e-492e-bc25-2f5440e2146b` as candidate `c1` (commit `78c116c90fc9c0262e3c6906d96ddf2c176a201c`).
Host checks on that tree: conformance passed, critic.rules passed, scenarios passed, tests passed.
Reviewed-by: claude (independent verifier session, SyberFactory workflow)


Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant