I1–I5 evidence
The current 0.3.0rc2 release requalifies I1–I5 against
the corrected source: 113 conformance tests and all five scenarios pass,
including the complete toy and 521-reference audit. Its
5,944-file archive,
relocated replay and
independent toy calculation pass.
The accepted distributable source is
sha256:a7a67e09548e40442f2ff72ba63898f66e02368046132a183588ebebdbc1a196.
The v0.3 ledger and comparison
cover the subsequent three-task native scope. The evidence below is the retained
7 October v0.2 audit, not a claim that old receipts describe the new source.
Target: the four adopted specifications in docs/specification/, originally
dated 6 October 2026. Noemetric’s scientific experiments are a separate study.
The post-delivery self-audit found and corrected 13 groups of defects or missing requirements. The requirement map identifies the evidence and limits for each obligation. Earlier passing tests did not cover all of these cases; their original evidence remains preserved.
| Milestone | Required demonstration | Audited status |
|---|---|---|
| I1 | Same contracts and artifacts through generic, Codex and Claude Code entrypoints; no native hooks or subagents | Passed within the declared packaging scope |
| I2 | Swap context method, reviewer adapter and recipe without rewriting other modules | Passed |
| I3 | One-module contribution, targeted offline conformance, lifecycle and ownership | Passed |
| I4 | Exact bundles, configuration provenance, update reconciliation, migration, rollback and historical resume | Passed |
| I5 | Actual toy raw data, recomputation, sensitivity, figure, claims, manuscript, success/failure/interruption | Passed |
Validated on 7 October 2026 (Asia/Seoul). The fresh acceptance receipt records 75 passing conformance tests with no skips and all five executable acceptance scenarios. The generic toy retains 26 successes, one failure and one interruption. Its deterministic audit verifies 521 distinct evidence references. A separate limited-budget study produces an explicitly partial manuscript.
Independent schema/packaging validation checks eight schemas, 643 record occurrences and 290 skill files. RFC 3339 timestamp validation is installed. A separate 1,557-case differential check agrees with the independent JSON Schema implementation.
The independent scientific calculation
uses rational arithmetic without importing the study implementation. It verifies
eligibility, primary/sensitivity tables, figure data, claims and manuscript
numbers. The original mean MSE difference remains 25/384. All 24
leave-one-seed-out means are positive, ranging from 0.06182065 to 0.07269022.
The SVG figure was also visually inspected.
The complete evidence archive contains 5,839 files. Its inventory and digest cover the exact bundles, raw attempts, comparisons and update/migration cases. Archive reproduction passed after extraction elsewhere using the pinned helper, with the new review retained. The original evidence describes the earlier source only.
Accepted distributable source revision:
sha256:fe61419ac78231ab0fcc0d70f77714f6946cbdb1bd72ce3a360313850855e100.
Study examples and conformance sources have separate captured inventories.
English technical documentation, contribution materials and the
commit-and-push instruction remain in the repository.
The default kit uses Python 3.11+ on POSIX, without paid APIs, GPU or third-party Python runtime dependencies. Historical bundles keep their old helpers; adopt the corrected source explicitly for new campaigns. Native Codex/Claude activation, model quality and independent scientific peer review were unperformed at that audit checkpoint.
Subsequent native Codex qualification
Section titled “Subsequent native Codex qualification”The weight EMA diffusion study adds real
native Codex evidence on 8 October 2026. Six distinct CLI 0.160.1 sessions
cover interruption, fresh-session recovery, pilots, confirmation, analysis and
reporting revision. The inherited model was gpt-6-astra with no project model
pin. The exact audited framework source above remains the study’s frozen
source; PyTorch and scientific dependencies belong only to the separate study.
Four pilot and ten confirmation trajectories completed on MPS. Every trajectory retains raw and EMA0.99/EMA0.999 checkpoints at 5,000 and 10,000 updates. One 100-update pilot interruption remains excluded. The deterministic audit and independent scientific verifiers recompute results and reproduce selected saved-weight samples. A global 1,800-second ceiling counts scientific checks, retries, execution, analysis and recomputation. See the host qualification report and machine-checkable receipt for actual cost, tested methods and limits. Claude activation, general model research quality and independent scientific peer review remain unqualified.
Canonical source · Generated from the maintained repository document at this build’s commit.