Cross-interface Acceptance Evidence For RC5¶
Status and identities¶
The historical automated runtime campaign passed before RC5 preparation. Actual Codex and current-specification human acceptance passed on the exact RC5 main candidate. Publication workflows and public verification passed. See the publication evidence. This document records what actually ran, without treating a previous artifact as a newly built RC5 wheel. The plan and release process still apply.
Exact RC5 candidate checkpoint¶
On 2026-10-16, PR #623
merged as frozen main 890925afdff68e97cafe38a28605d2874c9dad79. Its tree
a2f57bba64603b4ae2a39c97f958c506315e630c matches repaired reference
cd4f1cdba042862909fc9ccfe06fc52e4389e622. The following are candidate checks;
public release verification passed separately. Later documentation commits do not
change these source identities or inherit their results.
| Exact-main check | Result and evidence |
|---|---|
| Full local gate | 4197 passed, 23 classified skips, no failures, 91.49% coverage; schema, lint/types, compile, licenses/dependencies, budgets and quickstart passed. |
| Documentation and OpenSpec | Strict MkDocs passed; all 30 then-current OpenSpec items passed strict validation. |
| CI | Main workflow: Python 3.11–3.14 each passed 4197 tests with 23 skips; coverage 91.49%, 91.49%, 91.49%, 91.44%. Eighteen installed package profiles, four PostgreSQL routes and six Trino routes passed. |
| Containers | Main workflow: six actual hardened workloads across three targets and AMD64/ARM64 passed. |
| Documentation | Main workflow passed. |
| Security | Main workflow passed. |
| Installed matrices | Fresh exact-main runner/fixture: 600/600 logical cells on each of MCP 1.30.0 and 2.2.0, two seeded repeats per cell; workers completed and MCP processes were reaped. |
| Independent AI source review | Exact-main technical approval recorded in PR #623, subject to separate manual and release gates. Decision record SHA-256: 2ff4128de6dfe767f878b0347d61db38b4cf328276fcf08bcb4cb5eb6f4f030b. This is not human dataset consent or a GitHub-required human approval. |
| Ubuntu no-publish preflight | Exact-main run passed with Python 3.12.15: 4197 passed, 23 skips, build and installed-wheel checks. Downloaded wheel/sdist hashes were independently verified. |
The focused SDK suites ran separately on cd4f1cdba042862909fc9ccfe06fc52e4389e622:
847 passed and six legacy combinations excluded per SDK. Their 24 referenced
test/configuration files and all 157 package payloads were verified identical
at main. Those focused results retain their actual cd4f1cdb execution identity;
they are not described as fresh main execution. The 600-cell matrices above
were rerun for main. Their acquisition and finite-coverage limits remain those
described below.
Ubuntu publication-preflight artifacts:
| Artifact | SHA-256 |
|---|---|
agent_paranoid_android-1.6.0rc5-py3-none-any.whl |
969f969d57c0c87784cf7708d165507d8c3cd34eb41dccccaff43bdd08dea6d1 |
agent_paranoid_android-1.6.0rc5.tar.gz |
d2011e802d64157e9351805e94f97e2202b4c52a7c14511f6b6973b1079d3ae8 |
The separate macOS manual-client wheel has SHA-256
b32e459cd7a346bba2dc47e2caf61760c7fb8d35327cd486d286667389551368.
All 162 wheel members, including 157 package files, match the Ubuntu wheel in
content; only ZIP timestamps differ. The macOS hash is not the publication
preflight hash.
Actual Codex and human specification acceptance¶
On 2026-10-16, the actual Codex CLI 0.162.0-alpha.17.2 conversation continued
with its existing fictional plan. After the authored specification was installed,
the client inspected it and displayed the complete current file and its hash.
The human then explicitly attested full-file review and gave new specific consent
before generation. Earlier consent predating that file was excluded. Before
approval, the client inspected and reread the unchanged file. The observed
sequence then completed approve_dataset_plan, validate_dataset and a final
inspect_dataset_plan reporting phase completed.
| Manual acceptance binding | Value |
|---|---|
| Exact source | 890925afdff68e97cafe38a28605d2874c9dad79 |
| Reviewed specification byte SHA-256 | b691f8c82d2f0d28f8a70064ba414f269846ae88f98b781d06364a6ba97254e2 |
| Reviewed typed specification SHA-256 | 2e299faba7ac1dac3f18a385922c709f4294f3de5d77ac7fa8ba1b0a0eef365d |
| Manual evidence SHA-256 | 22105ba0a2dcd9ed7654898c669832c7eee2c85969c79542ad23aa2f61db1270 |
| Final observation SHA-256 | 59da184f674d6f90433e6b8596b35b4fe90492fb2292bb59d85b823822bd70d7 |
| Artifact audit SHA-256 | 425e5d066f09626995d257e15655e30fdf00aa9280ed9d1e1858f5d53074f719 |
The accepted result contained two synthetic customer rows and five synthetic
order rows. Validation and the separate artifact audit passed for primary/foreign
keys, native constraints, exact DECIMAL values and six business rules. The final
inspection reported source_rows_copied: false. This manual result uses the
separate macOS wheel identified above.
The evidence records the actual user self-attestation after complete current-file display; automated observation alone does not prove human reading. The same conversation continued without a second planning call, but the exact resume command and resumed binary version were not independently captured beyond the preserved client metadata. Separate trust/tool UI prompts and normal interactive exit/disconnect were not captured. No process bound to the acceptance MCP environment remained at the final observation; that does not establish a normal client exit status. No transformation or preservation authority was exercised. Public release acceptance passed separately; publication facts are recorded in the release evidence.
Historical campaign identities¶
| Identity | Value |
|---|---|
| Final campaign source | 408ec527795146cb0971dc12f7eb75cb9415b0a6 |
| Installed matrix/SDK execution source | 2eecfd6980bd88cf1c8a4152ed3f8792f6c30895 |
| Campaign wheel version | 1.6.0rc4 development build, distinct from published RC4 |
| Byte-identical wheel SHA-256 at both sources | 5eaddcebc83961d4c3e74daa0d12fdd25657d77734a1745456c80e2106a108e6 |
| Final campaign sdist SHA-256 | a31db8a1b5f8e57504ef04fd52bdabbf63013916101bd92815aba8e359ba1d4d |
| CI synthetic merge commit | 9e0c01dcc70a37bb3a7027cbf09de005d07f5b8d |
The final campaign commit changes only an isolated MCP inventory test. Rebuilt
wheels are identical, and the matrix runner, fixture and selected SDK test
definitions were checked for equivalence. Reports retain their real execution
SHA. The full source gate and corrected isolated test ran on 408ec527.
GitHub's comparison between that head and the CI merge commit has no changed
files. These are earlier runtime proofs, not fresh RC5 release gates or hashes.
Independent RC5 review subsequently found unbounded DECIMAL fixed-point
formatting in typed Parquet and its controlled-invalid fallback. The correction
uses one bounded converter across validation and output adapters. That runtime
change superseded version-only equivalence and required fresh installed-matrix
and full release gates on the corrected RC5 candidate. The candidate checkpoint
above records those reruns. Results below remain the historical campaign
baseline, not acceptance of the correction.
Historical campaign results¶
| Check | Result and scope |
|---|---|
| Full source gate | 4121 passed, 23 classified skips, 91.46% coverage; schema, lint/types, compile, licenses/dependencies, budgets and quickstart passed. |
| Hosted full Python matrix | Python 3.11–3.14 each passed 4121 tests with 23 skips; coverage 91.46%, 91.46%, 91.46%, 91.41%. |
| Installed matrix | 600/600 logical cells per MCP SDK major, two seeded repeats per cell, completed workers and reaped MCP processes. |
| SDK permissions/options | Per SDK: 771 passed, 6 excluded legacy preservation combinations, including 21 real stdio cases. |
| Finite option pairs | 464/464 defined compatible pairs; 52 behavioral cases plus an invariant over nine representative axes. |
| Isolated registration | Five passed: three common CLI flag placements, configured console tool inventory and offline piped OAuth refusal. |
| Additional delegated CLI | 22 passed: direct format-editor settings/refusals and six operations across three JSON/debug positions on expected-refusal paths. Configuration only, no approval/receipt. |
| Live fictional acquisition | Four PostgreSQL and six Trino routes passed with restricted readers. |
| Hardened containers | Six actual workloads: three targets on AMD64/ARM64; non-root, read-only, network-none, bounded resources and dropped capabilities. ARM64 uses CI emulation. |
| Installed package profiles | 18 passed: base, six optional extras, seven minimum-dependency profiles and four Python wheel-compatibility profiles. |
| Documentation and specifications | Strict MkDocs and all 29 then-current OpenSpec items passed. |
Hosted evidence: CI, 25 successful jobs, containers, Security and Documentation.
The installed MCP 1 environment used Python 3.12.13, MCP 1.30.0, Pydantic 2.14.1 and PyArrow 26; MCP 2 used Python 3.12.13, MCP 2.2.0, Pydantic 2.13.4 and PyArrow 25. All 157 installed package files matched the campaign wheel. Independent AI review closed the confirmed runtime findings; this is separate from human dataset approval and must be renewed for the final RC5 main SHA.
Coverage limits and unresolved steps¶
The 600 logical cells combine ten source handoffs, four outputs, five modes and Python/CLI/real-stdio MCP. Acquisition is prepared once per source route, then an independently authored reviewed specification exercises generation semantics. Database aggregates in this matrix are stand-ins; separate live integrations establish actual acquisition. This is not 600 live profiling/inference runs.
The 179 typed model properties, 225 top-level CLI flag occurrences, delegated operation inventory and 131 default MCP argument occurrences are navigation inventories. Finite pair coverage does not establish every possible value, all triples or all option combinations. Supplemental CLI refusal tests do not claim arbitrary internal-exception/debug behavior or per-leaf help coverage.
The 23 full-gate skips comprise ten live integration cases exercised separately, five isolated composition cases exercised separately, six explicitly excluded legacy preservation combinations and two inapplicable retained-destination combinations. Excluded/N/A cases are not counted as passed. Applicable bounded preservation refusal and destination-refusal checks passed separately.
The required six-container baseline passed. Two supplemental live-Trino runs inside final containers remain unexecuted after local dependency/network setup failed; they are outside that baseline and distinct from the successful live database and installed-process tests. Owned database/container resources were cleaned up. Temporary fictional test fixtures are retained with local evidence.
Actual Codex inspection, current-specification human review, new consent and the completed synthetic plan are recorded above, with explicit UI and client-exit observation limits. Automated SDK/PTY cases and permission to publish RC5 do not provide that dataset consent. The current exact-main gates, Ubuntu hashes and manual result do not replace publication or post-publication verification. The separately gated transformation/session campaign also remains open; this ordinary synthetic run grants no preservation authority or claim of complete campaign acceptance.