Technical extensions

This document is the audited register of architectural and technical measures in the current AI engineering system that are not explicitly prescribed by Developing-in-Agentic-AI-Systems-Learning-Paths.md.

The learning guide remains authoritative. An item belongs here only when it fills an implementation gap left by the guide. An extension may make a guide control more concrete or stricter, but it may not replace, weaken, or reinterpret a guide requirement.

Classification rule

Use this order for every architecture decision:

1. If the guide requires the behavior or names the technology, implement it as guide conformance.
2. If the guide offers multiple valid patterns, document which pattern is used and why; the choice itself is not a new architecture.
3. If the guide leaves the mechanism unspecified, record the chosen mechanism here before describing it as part of the system.
4. If a proposal conflicts with the guide, reject or redesign it. It cannot be legalized as an extension.

Extension identifiers are stable. Do not renumber them when an extension is retired.

architecture-lock.json carries a machine-readable mechanism-to-extension coverage manifest. The verifier checks that every registered ID has named mechanisms and existing evidence paths. That structural check does not prove that reviewers noticed every new mechanism; completeness remains an explicit architecture-review responsibility.

Register

ID	Extension	Scope	State
EXT-001	Versioned contracts and digest binding	Canonical control plane	Active
EXT-002	Deterministic risk floors	Canonical control plane	Active
EXT-003	Plan publication and branch topology	Reference policy	Active
EXT-004	Exact pre-tool authorization	Reference enforcement	Active
EXT-005	Payload-minimized local audit	Reference observability	Active
EXT-006	Evidence envelopes and readiness decisions	Canonical evidence model	Active
EXT-007	Trusted publication and validation-authority maintenance	Hosted trust boundary	Active
EXT-008	Split GitHub App identities and secret placement	Hosted identity boundary	Active
EXT-009	Governance drift audit and evidence retention	Hosted/source governance	Active
EXT-010	Failure-signature repair budget	Recovery	Active
EXT-011	Reference validation toolchain	Reference implementation	Active
EXT-012	PostgreSQL idempotency proving workload	Reference application	Active
EXT-013	Framework/reference repository separation	Framework maintenance	Active
EXT-014	Repository-local gh-aw MCP adapter	Reference developer tooling	Active
EXT-015	Supplemental security and merge checks	Reference evaluation	Active
EXT-016	Responsible AI Agent Hooks comparison	Experiment	Isolated
EXT-017	Guide conformance and extension ledger	Framework maintenance	Active
EXT-018	Single-source agent instructions	Reference context control	Active
EXT-019	Controlled bootstrap and deferred activation	Adoption process	Active
EXT-020	Documentation generation and publishing	Framework bookkeeping	Active; non-normative

“Active” means present in source, not proved conformant. At Northstar b65c2de5c8224342c72c37eeed7ef9f965ad8a2c, the following gaps block release. They are implementation defects, not compatible extensions to the guide:

Mechanisms	Current limitation / evidence
EXT-001, EXT-018, EXT-019	Missing cloud task cache and stale local task context; no completed cross-surface bootstrap
EXT-002, EXT-003, EXT-004	Catch-all risk precedence, approved-plan requirement at every risk, local-only branch rule, incomplete host payload/path normalization
EXT-005	Failure event may log success; malformed-input diagnostic echoes raw payload
EXT-006, EXT-007	Missing artifact/null digest can pass; stale Stop report and incomplete base/attempt binding; hosted acceptance absent
EXT-009	Missing live authentication; audit requires both alternative branch controls and optional capability files
EXT-010	Tested helper is not integrated into actual Stop-loop retry accounting
EXT-011, EXT-015	Five Poutine errors despite exit zero; unpinned zizmor image; malformed SARIF and unreadable scan inputs mishandled
EXT-012	Existing acceptance suite is in-process; additional audit-only separate-process proof is not a committed regression

Evidence, guide citations and repair ownership: AUDIT-2026-09-23.md and webmaxru/northstar-orders-api-demo#14, webmaxru/northstar-orders-api-demo#16, webmaxru/northstar-orders-api-demo#24, webmaxru/northstar-orders-api-demo#22, and webmaxru/northstar-orders-api-demo#20. Descriptions below state intended mechanisms; this limitation table qualifies their implementation claims until independently accepted repairs exist.

Later pull-request candidates and their local/hosted evidence are summarized in GUIDE-CONFORMANCE.md. A passing local test or reviewed candidate does not change the baseline classification or prove that an extension is accepted in the hosted system.

EXT-001 - Versioned contracts and digest binding

Guide gap. The guide requires inspectable task inputs, outputs, success criteria, plans, approvals, workflow outputs, artifacts, and traceability to a specific run and commit. It does not prescribe one serialization format, schema namespace, canonicalization algorithm, or tamper-evident binding between those artifacts.

Implementation. Northstar uses versioned JSON records such as northstar/plan/1, northstar/check-evidence/1, and northstar/execution-report/3. SHA-256 digests bind the live issue body, plan, base commit, approval, evidence, and maintenance manifest.

Compatibility. This makes the guide’s traceability and source-of-truth requirements machine-verifiable. It does not replace the issue, pull request, workflow run, review, or human decision.

Trust and operational effects. A semantically changed contract is a new contract even if its display name is unchanged. Schema evolution must be explicit, and stale or differently serialized artifacts cannot be silently accepted.

Rollback. Revert the schema change and its producer/consumer updates together. Never accept both old and new shapes through an unbounded fallback.

Northstar evidence. scripts/task-contract.mjs, scripts/plan-contract.mjs, scripts/plan-approval.mjs, scripts/evidence-record.mjs, scripts/build-execution-report.mjs, scripts/maintenance-manifest.mjs, and their focused unit tests.

EXT-002 - Deterministic risk floors

Guide gap. The guide requires autonomy to be explicitly bounded by risk and recommends a low, medium, high, and critical classification model. It does not define a portable algorithm for resolving conflicts between a requested risk level and sensitive paths or capabilities.

Implementation. Northstar derives a minimum risk from changed paths and requested operations. A narrative assessment may raise risk but cannot lower the deterministic floor.

Compatibility. The extension implements the required risk-based autonomy concept, adopts the guide’s recommended four-level model, and does not reduce any approval boundary.

Trust and operational effects. Policy files and validation code become high-risk because changing the classifier can change the controls applied to future work.

Rollback. Revert the policy and classifier together, then re-evaluate every open plan whose effective risk could change.

Northstar evidence. .github/governance/policy.json, scripts/risk-policy.mjs, scripts/plan-contract.mjs, and tests/unit/risk-policy.test.ts.

EXT-003 - Plan publication and branch topology

Guide gap. The guide permits both plan-first and plan + execution workflows and does not prescribe branch names, plan serialization locations, or how a read-only planner publishes without gaining repository write access.

Implementation. The planner writes only a local proposal through its stop hook. A human publishes the plan. High and critical work uses a plan-only pull request and a separate agent/implement/<task-id> branch created from the approved base SHA. Lower-risk work may use the guide’s plan + execution mode when policy allows it.

Compatibility target. Both guide-defined workflow modes must remain available; the current all-risk authorizer requirement does not prove the lower-risk mode. Separating publication from planning preserves the required read-only planning boundary.

Trust and operational effects. Approval binds to a plan-only commit and cannot be reused after the plan or base changes. Exact branch names are a Northstar convention, not a universal guide requirement.

Rollback. Close the implementation pull request, return to the approved base, and republish a new plan when scope or assumptions change.

Northstar evidence. .github/agents/plan.agent.md, .github/agents/implement.agent.md, scripts/plan-stop.mjs, scripts/publish-plan.mjs, scripts/plan-approval.mjs, and .github/workflows/plan-gate.yml.

EXT-004 - Exact pre-tool authorization

Guide gap. The guide requires hooks, tool allow lists, capability limits, path scope, and least privilege. It leaves repository-specific command grammar, tool payload parsing, base ancestry checks, and precedence rules unspecified.

Implementation. Northstar’s native PreToolUse hook resolves a trusted live task contract and approved plan, denies shell metacharacters and non-allowlisted command shapes, extracts paths from supported tool payloads, checks issue and plan scope, verifies the implementation branch and approved base ancestry, and categorically denies privileged operations.

Compatibility. This is a concrete, stricter implementation of the guide’s pre-action and least-privilege controls. It remains defense in depth; complete diff checks, GitHub policy, and human acceptance are still required.

Trust and operational effects. Unknown commands and unknown tool payload shapes fail closed. Hook timeout behavior supplied by the host remains a documented platform limitation and is not represented as complete mediation.

Rollback. Revert policy and authorizer changes together. A failing authorizer must be repaired through the approved control-plane maintenance path, not bypassed by renaming or disabling hooks.

Northstar evidence. .github/hooks/agent-boundary.json, scripts/authorize-tool.mjs, scripts/check-scope.mjs, and tests/unit/tool-authorization.test.ts.

EXT-005 - Payload-minimized local audit

Guide gap. The guide requires post-action/error observability and warns against exposing secrets, but it does not define a privacy-preserving local hook record.

Implementation. Northstar records actor, task, contract, plan, event, tool, paths, outcome, and SHA-256 digests of command arguments and results. It does not store raw prompts, request bodies, idempotency keys, commands, or tool results in the local audit record.

Compatibility. The record supplies attribution and correlation while preserving the guide’s secret boundary. Durable GitHub workflow artifacts and audit events remain the acceptance evidence.

Trust and operational effects. Digests support equality and correlation, not reconstruction or proof that the hidden payload was safe.

Rollback. Remove or rotate local session artifacts. Do not replace the scheme with raw payload logging.

Northstar evidence. scripts/audit-hook.mjs, .github/hooks/agent-boundary.json, and tests/unit/audit-hook.test.ts.

EXT-006 - Evidence envelopes and readiness decisions

Guide gap. The guide requires workflow outputs, artifacts, provenance, fan-in, and human/platform acceptance, but does not prescribe one evidence schema or local-versus-hosted decision vocabulary.

Intended implementation. Every producer emits a versioned evidence envelope bound to repository, pull request, base SHA, head SHA, workflow, run, job, actor, and artifact digest. Strict fan-in must reject missing, failed, skipped, duplicate, stale, cross-run, or cross-commit evidence. ready_for_review means local proof is complete; ready_for_acceptance additionally requires current hosted review and policy evidence.

Compatibility. The extension prevents self-reported success from replacing the guide’s system signals. Neither readiness state means that an agent approved its own work.

Trust and operational effects. Producers and the fan-in consumer form part of validation authority. Hosted evidence must be re-resolved for the immutable head instead of copied from an earlier run.

Rollback. Revert producer, schema, required-check policy, and fan-in changes together. Missing evidence remains a failure during rollback.

Northstar evidence. scripts/evidence-record.mjs, scripts/build-execution-report.mjs, scripts/check-human-review.mjs, .github/workflows/governed-change.yml, and the execution-report tests.

EXT-007 - Trusted publication and validation-authority maintenance

Guide gap. The guide requires independent acceptance controls, protected environments, workflow evidence, and human approval. It does not specify how a pull request that changes its own evaluator avoids self-certification.

Implementation. Pull-request-controlled code cannot produce the stable acceptance verdict for a validation-authority change. A default-branch workflow_run publisher imports only allowlisted evidence. Control-plane changes additionally require an immutable maintenance manifest, a protected system-maintenance environment, and revalidation by default-branch code before the stable trusted-acceptance status is published.

Compatibility. This preserves the guide’s independent human and platform acceptance boundary when the proposed change can alter its own checks.

Trust and operational effects. The default branch, protected environment, reviewer, workflow definition, evidence allow list, and publisher identity are part of the trust root. The first installation requires one explicit audited bootstrap; it is not a reusable bypass.

Rollback. Revert to the last trusted default-branch publisher and policy. Do not let pull-request code publish a substitute success status.

Northstar evidence. scripts/check-validation-authority.mjs, scripts/import-evidence-artifacts.mjs, scripts/maintenance-manifest.mjs, .github/workflows/publish-evidence.yml, and .github/workflows/system-maintenance-approval.yml.

EXT-008 - Split GitHub App identities and secret placement

Guide gap. The guide names GitHub App tokens, least privilege, protected environments, and scoped secrets, but it does not prescribe identities for maintenance dispatch versus acceptance publication.

Implementation. Northstar uses separate GitHub Apps: a dispatch-only identity that cannot publish status and a trusted-publisher identity that cannot dispatch maintenance. Private keys are available only in explicitly allowlisted protected environments and are forbidden as repository or organization Actions secrets.

Compatibility. Splitting identities narrows capability and prevents one credential from both initiating privileged validation and declaring its result.

Trust and operational effects. App installation permissions, integration IDs, environment branch restrictions, reviewers, and secret locations require live GitHub verification. Repository configuration alone is not proof.

Rollback. Disable both App installations or revoke their keys, restore the last known policy, and keep acceptance blocked until live identity checks pass.

Northstar evidence. .github/governance/policy.json, scripts/governance-audit.mjs, .github/workflows/publish-evidence.yml, and .github/workflows/system-maintenance-approval.yml.

EXT-009 - Governance drift audit and evidence retention

Guide gap. The guide requires continuous governance, permission reviews, and artifact-retention awareness. It recommends weekly, monthly, and quarterly review cadences and documents GitHub’s 90-day default retention. It does not define exact hosted queries, owner names, or a machine-readable desired-state policy.

Implementation. Northstar stores desired governance in .github/governance/policy.json, audits source and live repository settings, uses named weekly/monthly/quarterly review subjects, and adopts the guide’s documented 90-day default for canonical governance/evidence artifacts.

Compatibility. Northstar adopts the guide’s recommended cadence and documented retention baseline. The JSON shape, exact reviewers, and API queries are implementation choices.

Trust and operational effects. A source audit can verify intended policy and workflow structure. Rulesets, environments, App installations, security features, reviewers, variables, and secret locations remain not verified until live API evidence succeeds.

Rollback. Restore the previous desired-state policy and audit code together. Do not convert an unavailable hosted check into a pass.

Northstar evidence. .github/governance/policy.json, scripts/governance-audit.mjs, .github/workflows/governance-review.yml, and tests/unit/governance-audit.test.ts.

EXT-010 - Failure-signature repair budget

Guide gap. The guide requires bounded retries, rollback, and escalation and gives a two-failure escalation example. It does not prescribe a failure normalization or signature algorithm.

Implementation. Northstar classifies the responsible failure layer, normalizes volatile text, hashes the normalized signature, and stops automation when the same required failure repeats or the finite attempt budget is exhausted. Policy, security, and unknown failures escalate immediately.

Compatibility. This mechanizes bounded retries without treating repeated execution as recovery.

Trust and operational effects. Normalization must not collapse unrelated failures into one signature or strip information needed for diagnosis.

Rollback. Revert the classifier and signature logic together and retain a finite, explicitly documented retry limit.

Northstar evidence. scripts/repair-budget.mjs, docs/RECOVERY-POLICY.md, and tests/unit/repair-budget.test.ts.

EXT-011 - Reference validation toolchain

Guide gap. The guide names GitHub Actions, CodeQL, dependency/security signals, SARIF, and Agentic Workflows, but it does not mandate a language, runtime version, test runner, declaration format, workflow compiler version, or third-party workflow scanners.

Implementation. Northstar uses Node.js 22, ESM JavaScript control scripts, TypeScript declaration files, Vitest, npm scripts, a pinned gh-aw compilation path, and Poutine plus zizmor workflow analysis.

Compatibility. These tools produce or validate the guide-required signals; they do not replace GitHub checks, reviews, or security controls.

Trust and operational effects. The package lock, compiler version, scanner versions, generated lock workflow, and validation scripts are part of the reference validation supply chain.

Rollback. Revert tool and lockfile changes together and rerun the complete validation suite. A scanner unavailable in one environment must be reported as not run, not passed.

Northstar evidence. package.json, package-lock.json, scripts/run-zizmor.mjs, .github/workflows/daily-repository-status.md, .github/workflows/daily-repository-status.lock.yml, and tests/unit/.

EXT-012 - PostgreSQL idempotency proving workload

Guide gap. The guide requires observable success criteria and evidence appropriate to the software under change, but it does not prescribe an application domain or distributed-systems test workload.

Implementation. The reference is a TypeScript/Fastify Orders API whose idempotency behavior must hold across two stateless service instances. PostgreSQL supplies shared durability and advisory-lock-based concurrency control. Idempotency records retain hashes rather than raw keys or serialized request payloads; ordinary order business fields are persisted in orders.

Compatibility. This workload makes the guide’s requirement for system-grounded evaluation concrete. It is an example, not a universal application architecture.

Trust and operational effects. Unit tests cannot prove the cross-process criterion. PostgreSQL acceptance evidence is required whenever that behavior is in scope.

Rollback. Revert application and migration changes together and preserve a cross-instance acceptance test for the behavior claimed by the reference.

Northstar evidence. src/services/postgres-idempotent-order-service.ts, migrations/, tests/acceptance/, and the application ADRs.

EXT-013 - Framework/reference repository separation

Guide gap. The guide describes governance inside software repositories but does not prescribe how the architecture specification itself can evolve when the current system version is defective.

Implementation. webmaxru/ai-engineering-system is a non-self-governing specification repository. Every normative system change is first implemented and tested in webmaxru/northstar-orders-api-demo. Northstar knows about the framework through exactly one link in its root README.md; no instruction, workflow, application, or test file may depend on the framework repository. The framework keeps an inert, commit-locked snapshot under templates/northstar/; potentially active instruction files are renamed, including AGENTS.snapshot.md.

Compatibility. Northstar still demonstrates the full guide-defined system. The split prevents framework maintenance from requiring a bypass of the controls being redesigned.

Trust and operational effects. The two repositories can drift unless the source commit, normalized file hashes, and one-way reference boundary are checked. The snapshot is educational and never the behavioral source of truth.

Rollback. Restore the last lock that matches a validated Northstar commit. Do not activate copied hooks, agents, or workflows in the framework repository.

Evidence. AGENTS.md, docs/MAINTAINING-THE-SYSTEM.md, templates/northstar/reference-lock.json, tools/verify-reference.ps1, Northstar’s root README.md, and its context-architecture test.

EXT-014 - Repository-local gh-aw MCP adapter

Guide gap. The guide requires MCP registry and allow list governance but does not prescribe the local command adapter used to inspect or compile GitHub Agentic Workflows.

Implementation. Northstar declares a local gh aw mcp-server adapter and enables only the named compile, audit, logs, inspect, status, and audit-diff tools.

Compatibility. The named-tool list narrows access as required by the guide. Use remains contingent on organization or enterprise MCP policy; the repository file does not prove registry approval.

Trust and operational effects. Installing or expanding the adapter changes tool capability and is treated as a high-risk dependency and policy change.

Rollback. Remove the server declaration or reduce the named-tool list. Never replace registry approval with an unrestricted local wildcard.

Northstar evidence. .github/mcp.json, AGENTS.md, and .github/governance/policy.json.

EXT-015 - Supplemental security and merge checks

Guide gap. The guide requires security signals, dependency review, traceability, small pull requests, and current merge gates. It does not prescribe a deterministic source secret scanner or a merge-base compatibility algorithm.

Implementation. Northstar adds a payload-safe source secret scan, merge-base validation, and rename-aware complete-diff scope checks alongside CodeQL, dependency audit, GitHub secret scanning, push protection, and review.

Compatibility. These checks add signals; they do not substitute for the guide-named hosted controls.

Trust and operational effects. The supplemental scanner has a narrower claim than GitHub secret scanning. Merge validation must use the approved base and immutable head.

Rollback. Remove only the supplemental check and continue to require the guide-named security and merge controls. Report any unavailable hosted signal as not verified.

Northstar evidence. scripts/secret-scan.mjs, scripts/check-merge.mjs, scripts/check-scope.mjs, and their unit tests.

EXT-016 - Responsible AI Agent Hooks comparison

Guide gap. The guide requires lifecycle hooks but does not require the Responsible AI Agent Hooks project or define an adapter between that project’s eight interception points and GitHub Copilot’s exposed lifecycle events.

Implementation. Northstar branch reference/ai-engineering-system-agent-hooks contains an experimental partial adapter and compatibility evidence. It is intentionally excluded from the canonical main branch.

Compatibility. The project is conceptually compatible with the guide’s hook model. The current host cannot expose every required interception point, so the branch is not presented as an equivalent or complete implementation.

Trust and operational effects. Adding the framework introduces another runtime, policy representation, and adapter surface without removing the need for GitHub-native checks and acceptance.

Rollback. Delete or abandon the experimental branch. Canonical operation does not depend on it.

Evidence. docs/AGENT-HOOKS-COMPATIBILITY.md and Northstar branch reference/ai-engineering-system-agent-hooks.

EXT-017 - Guide conformance and extension ledger

Guide gap. The guide defines the architecture but does not prescribe how a separate framework repository records the exact guide version audited or prevents undocumented implementation additions.

Implementation. The framework makes the guide non-negotiable in AGENTS.md, records the audited guide and Northstar revisions in architecture-lock.json, maintains docs/GUIDE-CONFORMANCE.md, and requires every non-guide mechanism to have a stable entry in this register with compatibility, trust, rollback, and evidence. Its machine-readable extension coverage manifest maps each extension to named mechanisms and evidence paths. The local tools/verify-architecture.ps1 check detects guide drift, registry and coverage mismatches, unknown extension references, active framework controls, reference-snapshot drift, and blocked reference changes. It permits only the existing ordinary docs.yml publisher, not enforcement or agentic workflows, and enumerates tracked/unignored Markdown through Git instead of scanning installed dependencies or generated pages.

Compatibility. The ledger cannot override the guide. It exists only to make provenance and gap-filling decisions inspectable.

Trust and operational effects. A guide change invalidates the previous conformance conclusion until both repositories are re-audited. An unregistered implementation is nonconformant until classified. The verifier checks ledger integrity, not semantic completeness; reviewers must identify new mechanisms during every architecture change.

Rollback. Restore the previous guide and conformance record or complete a new full audit. Do not silently preserve a conclusion derived from a different guide version.

Evidence. This is non-normative framework bookkeeping rather than a runtime Northstar feature. Seven tests in tools/verify-architecture.test.mjs cover positive synthetic fixtures and forbidden active paths; they do not certify the real release. The reference verifier matches 117 files at exact b65c2de5c8224342c72c37eeed7ef9f965ad8a2c. The required accepted-reference condition remains unmet: this audited revision is known defective, so real architecture verification must still fail. See docs/GUIDE-CONFORMANCE.md.

EXT-018 - Single-source agent instructions

Guide gap. The guide requires consistent repository instructions and warns against context drift, but different GitHub and IDE surfaces do not all read the same instruction-file convention.

Implementation. Northstar authors repository-wide instructions once in AGENTS.md, generates .github/copilot-instructions.md from that source, and fails validation when the generated compatibility copy is stale.

Compatibility. The mechanism preserves one source of truth while making the same constraints visible to surfaces that do not yet read AGENTS.md.

Trust and operational effects. AGENTS.md is authoritative. The generated file must never be edited by hand or allowed to drift into a second policy.

Rollback. Restore both files from the last synchronized revision or remove the compatibility file only after every supported surface reads AGENTS.md.

Northstar evidence. AGENTS.md, .github/copilot-instructions.md, scripts/sync-agent-instructions.mjs, and the instructions:check validation command.

EXT-019 - Controlled bootstrap and deferred activation

Guide gap. The guide defines pre-action constraints and high-risk governance but does not prescribe how to install those controls in a repository that does not yet have a trusted task contract, customized paths, or validated commands.

Implementation. Adoption uses a human-owned bootstrap branch, an inert .ai-engineering-bootstrap/ staging directory excluded through the local Git exclude file, and a deferred hook activation step. The native hook is moved into .github/hooks/ only after its task, plan, path, and command policies pass negative and positive tests. The bootstrap is a one-time installation path, not a reusable bypass.

Compatibility. The process establishes the guide’s constraints before asking them to govern work. It does not relax the final hook, workflow, review, or hosted-control boundaries.

Trust and operational effects. Until activation, an accountable human owns the bootstrap diff and GitHub configuration. The staging directory is not durable system state and must never be committed. The selected framework and reference revisions must be immutable and verified before files are copied.

Rollback. Revert the audited bootstrap commit and remove only the named staging directory. Do not preserve a hidden disable switch or rename active controls as a normal development technique.

Evidence. docs/FULL-INSTALLATION.md preserves the integrated recipe; docs/QUICKSTART.md is the feature-selection entry point. The installed Northstar control plane exists, but a successful cross-surface installation canary is not established. Release remains blocked.

EXT-020 - Documentation generation and publishing

Guide gap. The guide does not prescribe a publishing stack or URL layout for the separate framework’s learning documents.

Implementation. Non-normative framework bookkeeping: Astro/Starlight generates pages from canonical docs/ Markdown; the existing ordinary docs.yml workflow publishes the site. Relative document paths determine routes; generated output is separate from authored sources. Synchronization rejects overlapping roots and route collisions before removing stale generated Markdown, and reports watcher failures explicitly.

Compatibility. Publishing neither governs this repository through the system under design nor changes Northstar behavior. The modular quickstart is navigation over guide practices, not independent activation of coupled scripts.

Trust and operational effects. Rendered pages are not a second source of authority, and a successful site build does not establish architecture conformance. Repository/base URL, branding, and source/output paths are marked CUSTOMIZE; trust boundaries and collision checks are not optional settings.

Rollback. Revert generator/configuration changes and rebuild from the unchanged canonical Markdown. Do not edit generated copies as authority.

Evidence. Four tests in docs-site/scripts/sync-docs.test.mjs, the production Astro build, and the framework verifier tests. Reference identity: b65c2de5c8224342c72c37eeed7ef9f965ad8a2c, which is not an accepted release. The accepted-reference requirement for a conformant framework remains outstanding; this bookkeeping does not supply substitute evidence.