Release Support¶

Release support binds a source identity, publishable distributions, scientific evidence, public language, and refusal state into one review. A build proves that archives can be produced. It does not prove that package boundaries are coherent, generated governance is current, a workflow can be rerun, or a scientific claim is ready for publication.

Release Evidence Pipeline¶

flowchart TB
    source["tagged source identity"] --> packages["package build + metadata validation"]
    packages --> truth["repository truth report"]
    truth --> dossier["scientific release dossier"]
    dossier --> language["public-language validation"]
    language --> consequence["runtime + decision + lab consequence"]
    consequence --> gate{"all required evidence passes?"}
    gate -->|yes| publish["publication may proceed"]
    gate -->|no| refuse["preserve failures and stop"]

make release-preflight evaluates documentation clarity, package boundaries, test collection, benchmark assets, runtime reproducibility, consequence coherence, and artifact hygiene in a fixed order. Every stage reports its own failures; a later pass cannot erase an earlier refusal.

Candidate Assembly¶

flowchart LR
    revision["clean source revision"]
    version["coherent package versions"]
    distributions["wheels · sdists · containers · docs"]
    inventory["hashes, metadata, SBOMs, attestations"]
    evidence["gate results and scientific dossier"]
    language["requested and allowed public language"]
    candidate["identified release candidate"]
    revision --> candidate
    version --> candidate
    distributions --> candidate
    inventory --> candidate
    evidence --> candidate
    language --> candidate

Publication review starts only after these identities agree. Rebuilding one distribution, regenerating an attestation, changing documentation, or rerunning a gate on another revision creates a different candidate or requires explicit proof that the governed inputs remained identical.

Scientific Release Dossier¶

build_scientific_release_dossier() in release/governance/scientific_readiness.py resolves each declared workflow from configs/package-governance/scientific-release-workflows.toml. The related configs/package-governance/flagship-workflow-manifest.toml records flagship workflow ownership and evidence paths.

For dda, dia, lfq, multiplex, ptm, and targeted, the dossier binds:

owning package and workflow identifier;
dataset or public benchmark-package locator;
builder symbol and package test evidence;
public documentation and claim-limit route;
scientific limitation that prevents broader wording.

The DDA evidence chain begins at benchmark-assets/flagship-public-packages/dda_reviewable_run/package_manifest.json, identifies the dda-maxquant-pipeline-corpus, and records comparator_path:msfragger_imported_dda_review. Those identifiers make the comparison inspectable; they do not turn imported search results into native engine execution.

Repository Truth¶

build_repository_truth_report() composes evidence from package ownership, public APIs, benchmarks, Runtime, generated governance, and package substance. It depends on these validators rather than replacing them:

Validator	Release question
`validate_ssot_readiness()`	do structured models and public symbols have coherent canonical owners?
`validate_generated_governance_freshness()`	do checked governance records match their generators and inputs?
`validate_public_language()`	does wording stay within the allowed evidence posture?
`validate_workflow_consequence_coherence()`	do grounding, recommendation, and Lab consequence agree across families?
`validate_workflow_public_scrutiny()`	can an external reviewer reach the artifacts that support or block each claim?

Repository truth should not be cited without those shared consequence surfaces. Open workflow-consequence-maps.md, what-changed-the-recommendation.md, outcome-learning-loops.md, and workflow-refusal-handbook.md with the report.

Runtime And Consequence Gates¶

Runtime black-box evidence is implemented in bijux_proteomics_runtime/workflows/black_box_reproducibility.py and reviewed through:

runtime-execution-boundary.md for exact manifests and entrypoints;
black-box-run-verification.md for installed public behavior;
raw-versus-import-execution.md for native and imported custody;
runtime-rerun-refusals.md for evidence required to widen a rerun claim.

The runtime flagship rerun gate answers whether the declared lane can be reopened under its recorded environment and artifact contract. The lab-consequence gate answers whether an advisory result has a controlled, feasible, and informative downstream path. Neither gate substitutes for Core scientific acceptance.

Public Scrutiny Routes¶

Review the following pages before using repository-wide readiness language:

flagship-release-candidate.md for the candidate bundle and current vetoes;
elite-readiness-scorecard.md for family evidence completeness;
workflow-claim-limits.md for allowed and blocked family language;
why-multiplex-stops-at-internal-support.md for the multiplex boundary;
public-artifact-index.md for externally inspectable evidence;
public-artifact-role-matrix.md for artifact authority and non-authority;
package-substance.md for thin-module and package-boundary pressure.

The generator modules workflow_lab_consequence.py, workflow_consequence_chain.py, workflow_consequence_docs.py, workflow_public_scrutiny.py, hostile_review_pages.py, and release_narrowing_protocol.py keep the shared family and hostile-review surfaces synchronized. final_preflight.py composes their results into the ordered release decision. Fresh output is required, but freshness alone does not mean the evidence passes.

Publication Refusals¶

Failure	Required response
unresolved release identity	stop; establish the source version and tag relationship
wheel or sdist version differs from source	reject the artifact and rebuild from the tagged revision
duplicate canonical ownership	resolve the owner; do not waive the SSOT gate
stale generated governance	correct the generator or input and regenerate
requested language exceeds allowed language	apply the release narrowing protocol
runtime lane cannot support the claimed rerun mode	preserve the refusal and narrow the claim
consequence chain is incomplete or contradictory	withhold repository-wide readiness language
package-root artifacts or caches remain	clean the output and correct its producer

Do not weaken a validator, hand-edit generated evidence, or discard failure output to unblock publication. Resolve the owner-level cause, rerun the narrow gate, and then repeat make release-preflight from a clean repository state.

Decision	Minimum condition	Required record
publish	every applicable gate passes and publication authority approves the identified candidate	candidate identity, channel actions, reviewer, rationale, timestamps
narrow	artifacts are publishable but requested public language exceeds evidence	allowed language, blocked wording, owner, replacement public surfaces
withhold	candidate evidence is incomplete or temporarily blocked	missing evidence, failed prerequisite, owner, reconsideration condition
refuse	a contract, security, scientific, provenance, or integrity gate fails	retained failure, affected artifact or claim, required correction

Signing and attestation prove artifact custody and build assertions; they do not override a scientific refusal, stale governance, or incompatible schema. Likewise, a publishable documentation site does not authorize package or container publication unless the same candidate passed its own channel gates.

Release Decision Record¶

A release decision is reconstructable only when the record binds the following fields to the same candidate:

field	required evidence
source identity	commit, clean-worktree result, tag, and source version
distribution identity	wheel and sdist names, versions, hashes, metadata validation, and package inventory
validation identity	exact commands, environment, timestamps, results, and retained failure output
generated authority	generator inputs, checked outputs, and freshness results
scientific authority	workflow dossier, primary and companion benchmark identities, runtime mode, and acceptance result
public authority	requested and allowed language, consequence posture, and unresolved vetoes
publication action	published, withheld, or refused; reviewer; reason; and supersession link

Passing results from different commits or environments cannot be assembled into one release decision unless the record demonstrates that the candidate inputs are identical. A rebuilt artifact receives new hashes and a new distribution review even when its version string is unchanged.