This is the operational contract for agents modifying the repository.
Keep runtime code, schemas, profiles, evidence sidecars, maintenance control, current documentation, and historical-document status semantically aligned without turning agent narrative into renderer or scientific truth.
Read docs/03-maintenance-and-audit/DOCUMENT_STATUS.md before broad documentation or governance maintenance.
For repository-level open-research positioning or independent research-production work, read OPEN_RESEARCH.md. Use RESEARCH_TEMPLATE.md only for prospective bounded records when no more specific native template owns the surface. Neither file supersedes implementation, MANIFEST.yaml, machine configuration, subject-specific figure/communication contracts, or the native maintenance/frontier-research/ specification/history.
Use docs/03-maintenance-and-audit/history/JULES_CORRECTION_RECORD.md only when interpreting early Jules task/PR prose as historical renderer/backend/publisher/accessibility/scientific evidence. It is dated correction evidence, not a current authority layer.
historical snapshot != current contract
later renderer capability != permission to rewrite history
historical agent PR narrative != current communication contract
current merged main implementation
> MANIFEST.yaml / metadata / profiles / quality / recipes / machine configuration
> latest relevant dated repair or current maintenance record
> docs/03-maintenance-and-audit/DOCUMENT_STATUS.md
> AGENTS.md
> active subject-specific contracts
> docs/03-maintenance-and-audit/MAINTENANCE_CADENCE.md / maintenance/cadence.yaml
> current Architecture / README explanation
> historical snapshots / superseded plans / PR-task narratives
current implementation
> current schema/profile/configuration/machine contract for the subject
> active subject-specific contract
> executable/operational evidence for supported use
> current explanatory documentation
> maintenance evidence
> historical records
A newer maintenance record does not become stronger renderer, publisher, accessibility, or scientific authority merely because of date.
Record, when applicable:
repository
+ owning surface / task
+ logical period or evidence window
+ producer / maintainer
+ exact base revision
+ run identity when available
Before any write, inspect open PRs/live branches for the same owning surface and logical period.
overlap -> COORDINATE
no confirmed defect -> NO_CHANGE_REQUIRED
confirmed current drift -> REPAIR
unsafe or unrecoverable evidence/access -> BLOCKED
Do not create repository objects to test write access. Write never probes.
recipe
-> P0 schema
-> P1 runtime + claim communication audit
├─ assertion basis
└─ dimensional communication coverage
-> backend capability resolution
-> render
-> render-manifest / provenance / a11y
-> P2 artifact integrity
-> P3 publisher-target alignment
-> figure-evidence
-> optional communication-transfer
└─ explicit non-inheritance constraints
repository state
-> daily / weekly / monthly maintenance
└─ current-document / calendar / stage / delivery reconciliation
Use stable unversioned project-owned profile names. Preserve real WCAG/external/runtime versions when genuinely known. Do not invent decorative internal versions.
- Schema does not imply backend support.
- Do not silently change DPI, format, uncertainty semantics, or declared data.
- Claim bindings are explicit only and never inferred from titles, legends, pixels, prose, filenames, or data values.
- Claim relation labels are communication declarations, not verified entailment.
- Process disclosure is bounded and does not adjudicate authorship, peer review, or truth.
- Unknown provider/model/version/review/source metadata remains unknown.
- The canonical renderer does not infer AI authorship/use from text or pixels.
- Assertion basis records how a field entered evidence and never upgrades a value to correctness.
- Communication coverage remains dimensional and must not become an unsupported aggregate quality score.
- Coverage ratios are not probability, entailment, evidence sufficiency, or provenance soundness.
- Bounds alone are not a confidence interval or probability model.
- WCAG scope stays exact; project all-pairs checks remain extra safeguards.
- CVD simulation is extra and not normative WCAG certification.
- Publisher presets are snapshots/config targets, not acceptance validators.
- Optional backend source does not prove runtime availability.
- Figure evidence is handoff, not truth.
- Communication transfer may copy declared context but must not inherit scientific validity, entailment, evidence sufficiency, statistical validity, peer review, publisher acceptance, or WCAG conformance.
- Unsupported experimental methods fail explicitly rather than fabricate output.
- Calendar/month/stage status comes from actual date/configuration.
- A worked maintenance demonstration is not a clean runtime result unless the scanner actually ran and output was preserved.
- Do not add GitHub Actions, CI, CodeQL, dependency bots, branch-protection assumptions, or merge-gate architecture as routine maintenance.
- Agent task text/PR bodies/summaries/completion claims are proposal/delivery metadata, not automatic repository authority.
- Historical
tests passed, backend execution, checksum,complete,fully aligned, orfixedrequire current re-verification before reuse. - Correct historical overstatement forward; do not rewrite old PR history.
- Path relocation does not change semantic status.
- Scanner/checker source or configuration inspection is not execution.
- Unrun checks are
NOT_EXECUTED; unobserved scheduler/workflow execution isEXECUTION_NOT_OBSERVEDwhen material. - A Draft PR is a review boundary, not renderer/test/scientific/merge success.
- No confirmed maintenance defect means no activity-only branch or PR.
assertion basis != correctness
communication coverage != entailment
coverage ratio != probability
supports evidence-context coverage != evidence sufficiency
coverage != provenance soundness
Keep aggregate_score: null unless a future validated evaluation design explicitly justifies a composite score.
transfer != entailment
upstream ref != inherited validity
uncertainty metadata != statistical validation
publisher target != acceptance
accessibility metadata != WCAG certification
human review != peer review
Destination and purpose are caller-declared when present and must not be inferred.
Before reusing a historical agent statement, distinguish:
what the PR body claimed
what actually entered merged main
what current main does now
what current machine contracts say
what was actually re-run or re-verified now
Historical backend/test/render/checksum/publisher wording does not establish current runtime success, publisher acceptance, WCAG certification, scientific validity, entailment, or independent reproduction.
Temporal-date rule: scanner/report as_of is an observation date; MANIFEST.yaml.current_temporal_status.as_of is the last explicit calendar-state reconciliation represented by the static manifest and is not a daily heartbeat. Do not bump the manifest date solely because render/evidence or longitudinal maintenance ran while the calendar state is unchanged.
The active maintenance system is jointly owned by:
docs/03-maintenance-and-audit/DOCUMENT_STATUS.md
docs/03-maintenance-and-audit/MAINTENANCE_CADENCE.md
docs/03-maintenance-and-audit/README.md
docs/03-maintenance-and-audit/independent-gpt/README.md
maintenance/cadence.yaml
core/maintenance_cadence.py
The .py scanner is executable implementation. Governance prose changes do not by themselves require rewriting it, and source/config inspection is not scanner execution.
Local scanner:
python core/maintenance_cadence.py daily
python core/maintenance_cadence.py weekly
python core/maintenance_cadence.py monthly --as-of YYYY-MM-DDDaily corrects demonstrated recipe/evidence/backend/profile/document/governance drift only and permits NO_CHANGE_REQUIRED without branch/PR churn.
Weekly reconciles implementation, machine contracts, active communication contracts, README/Architecture, Agent/Contributor guidance, examples, document status, maintenance configuration/records, checker ownership, and upstream profile names.
Monthly/phase-close derives calendar/phase state from the actual date and never converts closure into reproduction or scientific validation.
If Daily and Weekly own the same real correction, one branch and one final Draft PR should carry it whenever practical.
scanner source present != scanner executed
scanner executed != scanner passed
historical render/test pass != current pass
contract inspection != runtime verification
| Goal | Primary files | Synchronize |
|---|---|---|
| recipe field | metadata/recipe.schema.yaml |
runtime + docs + evidence semantics |
| claim binding/audit/coverage | claim-audit / figure-evidence implementation | communication contracts + Manifest + examples |
| communication transfer | transfer implementation | transfer contract + machine contract + examples |
| maintenance cadence / agent provenance | core/maintenance_cadence.py, maintenance/cadence.yaml |
Maintenance Cadence + Document Status + maintenance README + Independent GPT router + Agent Guide + current dated maintenance record; synchronize runtime scanner/Manifest only when executable/profile/path semantics actually change |
| accessibility | accessibility implementation/runtime | a11y + backend truth |
| backend | corresponding adapter | Manifest/provenance capability truth |
| publisher profile | profiles/*.yaml |
source status + acceptance false |
| public capability | README / Architecture / active contracts / Manifest | update together when semantics change |
auto-doc-engine/artifact-record
auto-doc-engine/artifact-lineage
epistemic-pipeline/claim-verification
epistemic-pipeline/claim-transfer
epistemic-pipeline/evidence-envelope
sci-render-kit/communication-transfer
These are optional handoff references, not direct coupling or inherited scientific validity.
Metaphorical filenames are not scientific capability evidence. Experimental/importable modules are not canonical capability until intentionally integrated.
Render manifests, figure evidence, communication transfers, accessibility sidecars, maintenance reports, demonstrations, and canonical hash baselines never count as independent reproduction. R3 requires an actual separate rerun plus a declared comparison criterion.
For a confirmed maintenance repair:
- branch from exact observed current
main; - synchronize owning control surfaces and true dependencies only;
- inspect aggregate branch diff;
- refresh current-main/open-PR overlap;
- record executed/unexecuted checks separately;
- open one bounded Draft PR;
- stop for maintainer review.
Do not auto-merge, force-push, or write maintenance repairs directly to main. Final doctrine and merge authority remains with the maintainer.
Do not publish private Jules prompts, repository memory, hidden reasoning, credentials, or unrelated operator context. Public governance may preserve the effect of a rule without copying private control text.