CoderPolice v0.1.0-alpha.2

CoderPolice v0.1.0-alpha.2

Status: public research alpha. No package-registry publication is part of this release.

What changed

This alpha records the public project rename from Agentic Engineering Conformance to CoderPolice.

Agentic Engineering Conformance remains the technical descriptor for the benchmark. CoderPolice is the public project name.

The primary tagline is:

Okay, but prove it.

The GitHub repository is now:

https://github.com/deevyanshoo/coderpolice

The Python distribution metadata is now coderpolice at version 0.1.0a2. The import namespace remains agentic_conformance to avoid unnecessary breaking churn.

Historical continuity

v0.1.0-alpha.1 remains unchanged under the original technical working title.

This release does not rewrite:

• M1 through M6 reports or execution records;
• AUTH-001 v1 or v2 semantics;
• historical experiment bundles;
• scenario identifiers;
• evidence digests;
• host results;
• benchmark classifications; or
• prior GitHub release history.

The old repository URL remains a GitHub redirect to the renamed repository.

Verification

The CoderPolice brand PR and the post-rename cleanup PR both passed the repository-owned deterministic GitHub Actions gate.

The current release candidate has passed:

• editable installation;
• Ruff format check;
• Ruff lint;
• strict mypy;
• full pytest, schemas, and contracts; and
• public GitHub Actions on main after the repository rename.

No live Codex or Claude Code benchmark trials were needed because this release does not change benchmark semantics, adapters, fixtures, or runtime behavior.

Scope carried forward from alpha.1

CoderPolice remains a research alpha with:

• six provisional domains: AUTH, MUT, COMP, REV, INV, and REC;
• six deterministic synthetic reference scenarios;
• separate functional and control outcomes;
• guarded-versus-behavioral pass semantics;
• explicit evidence provenance and stored-evidence rescoring;
• Codex and Claude Code adapters for AUTH-001; and
• experimental neutral Windows Task Scheduler execution.

Limitations and non-claims

• The six-domain taxonomy remains provisional.
• Real-host integration currently covers AUTH-001, not all six domains.
• Small host samples are not comparative performance evidence.
• The neutral runner remains Windows-specific in the current alpha.
• CoderPolice is not a certification program, security boundary, leaderboard, or proof that a coding-agent stack is safe in general.
• No novelty or legal priority claim is made for the name, taxonomy, workflow concepts, evidence model, review model, DAG patterns, invalidation patterns, tracing, or orchestration ideas.

Contributing

Deterministic reference contributions require no paid host account. See CONTRIBUTING, the scenario guide, the adapter guide, and the public evidence policy.

Agentic Engineering Conformance v0.1.0-alpha.1

v0.1.0-alpha.1

Status: public research alpha. No package-registry publication has occurred.

Alpha scope

This research alpha packages this project's initial vendor-neutral reference vertical slice for testing engineering-control guarantees around coding-agent stacks. It includes:

• six provisional domains: AUTH, MUT, COMP, REV, INV, and REC;
• six deterministic synthetic reference scenarios with separate functional and control outcomes;
• all six run classifications, guarded-versus-behavioral pass semantics, explicit evidence provenance, and stored-evidence rescoring;
• the common adapter interface plus Codex and Claude Code AUTH-001 integrations;
• digest-bound neutral Windows Task Scheduler experiment execution;
• a no-conflict calibration condition and paired aggregate interpretation; and
• public contribution, security, evidence/privacy, licensing, claims, and prior-art guidance.

The Python project version is 0.1.0a1; the human-facing Git tag is v0.1.0-alpha.1.

AUTH history and current semantics

AUTH-001 v1 remains immutable for historical replay. AUTH-001 v2 preserves stale selection A as a control failure, identifies current behavior B as a behavioral pass unless a guard is independently observed, and represents no decision (UNSET) as functional failure with control INCONCLUSIVE. Calibration uses the same useful mutation without stale conflict and is not an AUTH conformance result.

The M5 twelve-slot batch remains historical and terminally invalid before model launch. A distinct M6 successor executed all twelve paired slots at one bound revision. Its sanitized report records exact run-specific outcomes and limitations. N=3 per condition supports no ranking or statistical conclusion.

Verification state

The verified public release commit c74991eeda70d74929603d7b3d9078f458209b46 passed Ruff formatting and lint, strict mypy (27 source files), 268 deterministic tests including schemas/contracts, deterministic reference execution with identical offline rescore, and fresh-clone reproduction. Current-tree scans were clean; scans across all 90 commits reachable from public main found no high-confidence secret, credential assignment, Nnomi/proprietary sentinel, suspicious artifact, or blob over 1 MiB. The two previously disclosed harmless historical local-path files remain preserved as provenance.

Public hosted CI is green at the release commit: main run 33328584711 executed install, Ruff format, Ruff lint, strict mypy, and full pytest successfully. The first public run (33321316261) had exposed a real cross-platform defect: Linux interpreted persisted Windows executable identities using native path semantics. Release-blocking PR #6 now separates portable lexical path validation from local execution binding, preserves historical plan digests, and adds Windows/POSIX regressions without changing scenarios, oracles, or benchmark outcomes.

Limitations and non-claims

• The six-domain taxonomy and real-host AUTH construct validity remain experimental.
• Real adapters currently cover AUTH-001, not all six domains.
• The neutral worker is Windows-specific and does not create a sterile environment.
• Small historical and launch-validation samples are not comparative performance evidence.
• This is not security certification, a complete agent-safety framework, or a production guarantee.
• No novelty or priority claim is made for the underlying workflow, evidence, review, DAG, invalidation, tracing, or orchestration concepts.

Contributing

Deterministic reference contributions require no paid host account. See CONTRIBUTING, the scenario guide, the adapter guide, and the public evidence policy.