Registration and result of C61: a paired-check audit of derived statuses in an external agent specification (dsh-superagent-prototype, commit 6337fc8), version 2
收藏资源简介:
Version 2 (1 November 2026, CST) adds the result of the registered run to the registration of version 1, whose five files are carried over unchanged. Verdict: supported. By the frozen adjudication script (adjudicate.py, committed before the run; c61-adjudication.json): P1 holds, the pinned suite passed in full on re-run (1197 tests in 59 files, typecheck and build, exit 0; a first run under a sandbox that denied loopback failed 8 tests on listen EPERM 127.0.0.1 and is kept beside, see the amendment); P2 holds, none of the eight statuses read as adequate was inadequate when run (statuses 2, 4 and 8 are reported as specification ambiguity or consumer-policy cases and do not count); P3 holds, status 10, the health entry of an effect kind, is inadequate by a witness demonstrated by running, in the class of partial coverage read as a full verdict (it reads non-error for a kind that never processed anything, and for the quiet delivery kind after five consecutive failures). Beyond the predictions, status 11, the outbox receipt sent, is inadequate by running, in the class of delivery read as receipt (the receipt comes from the send call's return value alone). The registration's inventory read statuses 3 and 9 wrongly: the unavailable decision has a wider input than one path, and auto-merge readiness is not a host derivation. The run was a fresh headless session given only the registration, sections 4 and 6 of Protocols for Continuity v0.7, the pin manifest, the suite result and the pinned clone; its seven witness tests (17 tests, all passed) ran against the subject's own in-memory harness under sandbox-exec. Files added: c61-result.pdf and c61-result.md (the result document), c61-findings.json and c61-report.md (the runner's output), the seven witness-.spec.ts files with witness-README.md and witness-run.log, suite.json and suite.log (the loopback-allowed run) and suite-run1-loopback-denied. (the first run), c61-adjudication.json and adjudicate.py, the amendment, FREEZE-PINS-v2.json and MANIFEST-v2.sha256. Registration of conjecture C61 of the Epistemic Science and Engineering corpus, frozen on 1 November 2026 before any part of the subject was installed or run anywhere in the author's estate. Reserved DOI of this record: 10.5281/zenodo.23875233. Subject: dsh-superagent-prototype, version 0.1.0, by the DeepSeek Harness team, MIT licence, a proactive coding agent for one GitHub repository. Pin: GitHub deepseek-ai/dsh-superagent-prototype at commit 6337fc823ce20251c8de8085948165c81aaa701e (2026-10-31T12:05:39Z), the repository's only commit at registration. A clone at that commit was hashed on 2026-11-01: 216 files, manifest SHA-256 00f7442bb521092322252e3f5e9bbd618e0642ca20c37a5c873a77d9bfa699d6. Its suite: 60 test files, 794 tests by grep, not run. Method: the paired check of Protocols for Continuity v0.7 (status adequacy; witness pairs, two histories that give a derivation the same input and differ in the property its name asserts; evidence classes 1 to 4), applied for the first time to a specification the author did not write. The registration holds the inventory of eleven derived statuses with the property asserted and the input read for each (eight read adequate in the specification's documents, three run-only), the five-step procedure, the adjudication rule, and three predictions frozen with their consequences: P1, the pinned suite passes in full on re-run; P2, none of the eight read statuses is inadequate; P3, at least one of the three run-only statuses is inadequate with a class-1 witness, in the class of a completion signal read as attainment or of partial coverage read as a full verdict. The run is done later by a fresh session given this registration, the pinned clone and sections 4 and 6 of the paper only. Files: c61-registration.pdf and c61-registration.md (the registration note); FREEZE-PINS.json (the pin, the manifests, the suite count, the design statement, the deposit); MANIFEST-dsh-superagent-prototype-6337fc8.sha256 (the 216 file digests of the pinned clone); MANIFEST.sha256 (this bundle). Nothing in this record is a result. No efficacy claim, no deployment result; one specification at one pin. The reading of the eight statuses was not blind; the run is. Amendments after registration are dated files; the note is never edited.



