The matrix is current for the governed architecture. The repository provides hermetic enforcement for exact target slots, complete TCB denial, signed cross-phase cost accounting, ordinal delivery effects, comment-only review, independent human-merge observation, deterministic closure, audit/metric/budget evidence, and durable grant-claim boundaries. No row is fully deployed under the readiness rule below because a secure App, state publisher, trusted executor/bridge, OIDC redeemer, key custody, authenticated conditional evidence store, serialization, administrator configuration, billing, platform protections, and security-owner review remain human/deployment prerequisites.
| ID | Threats | Preventive control | Detective / recovery control | Owner / plane | Implementation | Required evidence | Residual risk |
|---|---|---|---|---|---|---|---|
| C01 | T01, T02 | Untrusted content is data-only; model outputs use strict target-free schemas and executor output names only trusted logical slots. | Adapter logs rejection reason and content digest. | Kernel / model boundary | Repository implementation | Negative tests for targets, commands, unknown fields, hidden content, slot substitution, and oversize | Semantic injection can still influence permitted content. |
| C02 | T02, T05 | Trusted Binding derives immutable IDs and SHAs from event plus fresh API reads. | Pre-effect reread detects mismatch and discards stale attempt. | GitHub adapter | Repository implementation | Wrong-repo/install/issue/PR/branch/Project and TOCTOU tests | Platform/API compromise remains. |
| C03 | T03 | GitHub App only; no PAT fallback; operation-scoped short-lived installation token minted after policy gate. | Token identity/permission audit receipt; immediate disposal. | Credential broker | Repository implementation | Permission-diff tests and proof models/workers cannot read token | Installation permissions may still be misconfigured by admins. |
| C04 | T04, T19 | Effect idempotency plus a signed one-time runtime nonce and exact (workflow, run_id, run_attempt) redemption; every logical effect also carries a trusted ordinal and type. Effect claims bind claimant, plan, binding, and operation digests. |
Runtime authorization CAS consumes nonce/run attempt; Single Writer conditionally appends against an authenticated head and reconciles reads after writes. | Kernel / scheduler / redeemer | Repository implementation | Concurrent/sequential replay, run-attempt, nonce, effect-ordinal, CAS-conflict, reordered, rollback, rerun, collision, state-race, partial-effect, and competing-claimant tests | Cross-system side effects remain difficult to make atomic. |
| C05 | T05 | Compare-and-swap predicates on complete Trusted Binding targets, contract, receipt head, Project schema/mapping, base/head repository/ref/SHA, and current Project value. | Reconciler reconstructs live state; stale attempt retained as evidence. | Single Writer / reconciler | Repository implementation | Target-substitution and race tests at every mutating edge | Tiny platform-side race windows remain. |
| C06 | T06, T07 | Reviewer must be non-bot, independent where required, currently eligible by immutable identity/role/team, and approved on exact head SHA. Automated review is schema-fixed to COMMENT; orchestration stops at a signed awaiting-human checkpoint and cannot mint approval or merge evidence. |
A later invocation validates independently issued exact-head approval before a later fresh human-merge observation; head or permission change invalidates evidence. | Policy Gate | Repository implementation | Negative tests for missing/out-of-order approval, requester/bot/stale review/lost role/wrong team/wrong head and non-comment review events | Human collusion or compromised reviewer account. |
| C07 | T07 | Rulesets/CODEOWNERS/required checks configured by human admins; no automation bypass actors. | Drift monitor blocks readiness and opens an admin runbook item. | Human admin / adapter | Repository implementation | Configuration conformance test and drift fixture | Administrators can intentionally bypass or remove policy. |
| C08 | T08 | Every privileged route requires semantic threat result exactly success for exact input/output digests. The execution bridge job additionally requires both generated detector outputs (detection_success=true and detection_conclusion=success). Platform safe outputs stay staged, and implicit failure/missing/incomplete/noop issue handlers are disabled; only the trusted bridge can construct an App effect plan. |
Warning, absent, skipped, cancelled, stale, or mismatched result fails bridge validation; generated conclusion jobs have no issue-write permission. | Policy Gate | Repository implementation | Deterministic warning, digest-substitution, stale-evidence, generated-permission, fallback-tool, detector-output, and forged-authorization tests; deployed write-path tests remain | Detector false negatives; deployed alerting remains. |
| C09 | T09, T13 | Immutable capabilities and runtime bindings declare exact agents, primary/safety skills, local tools, GitHub MCP reads, staged Safe Outputs tools, commands, external MCP, network destinations, secrets, effects, and budgets. Workflow frontmatter is reconciled to policy; omitted, wildcard, or mismatched authority is rejected. | Source validation and invocation/effect audit; emergency capability disablement remains deployment work. | Capability Registry | Repository implementation | Registry semantics, workflow/agent/skill/tool parity, role-binding, wildcard, escalation, and eval-fixture validation | GitHub's built-in read gateway and approved dependencies may be compromised. |
| C10 | T10 | Runtime policy denies external network, external MCP, secrets, PAT fallback, and model access to App credentials. Framing/review workflow reads use the ephemeral GITHUB_TOKEN; execution has no GitHub tools or writable credentials. |
Generated-lock inspection, executor environment scrubbing, and future secret/log scanning. | Runtime / security | Repository implementation | Static no-fallback/no-network validation and credential-stripping tests; deployed canary/redaction tests remain | Platform logs and semantic secrets require deployed testing. |
| C11 | T11, T16 | Hash-chained receipts bind source event, contract, policy, provider/model, input/output digests, actor, target, and effect. GitHub effect evidence is signed by a deployment-held key, attributed to configured numeric App/bot identities, chain-linked to an authoritative conditional head, and binds exact workflow, revision, effect type, and logical ordinal. One signed execution bundle carries the authorization, actual applied Kernel result, current policy digests, exact-success threat evidence, and complete signed patch content. The post-upload OIDC invocation binds repository/workflow/run identity plus exact artifact ID, archive digest, and bundle digest. | The trusted service validates runtime policy before artifact use, derives one freshness window only from its authenticated limits.maxEvidenceAgeMs, rejects legacy caller limits, downloads raw archive bytes, recomputes their SHA-256, extracts exactly one bounded CRC-checked canonical bundle, and binds its embedded authorization to the authenticated workflow identity. It issues one opaque policy-bound freshness capability through a module-private issuer, registers it in a runtime identity set, binds policy/authorization/run/Kernel identity plus the same trusted clock and full runtime-evidence recheck, rejects structural/no-op/substitute-clock forgeries, passes it unchanged through concrete delivery, and revalidates bundle, authorization, Kernel, threat, patch, and effect evidence immediately before each Single Writer mutation. Canonical binding resolution and exact binding-store readback remain mandatory. |
Evidence Ledger | Repository implementation | Forgery, edit, refused/noop/stale/wrong-capability Kernel proof, wrong-run/artifact/archive/bundle identity, unpersisted binding, caller freshness enlargement, independent adapter clock, delayed concrete delivery, exact boundary, mismatched evidence ages, wrong-author/App, bad-signature, rollback, reorder, duplicate-head, wrong-subject, artifact substitution, and replay-equivalence tests | GitHub comment deletion is detectable only while the deployment retains an authoritative head; external audit may still be needed. |
| C12 | T12, T19 | Runtime entry is a role-restricted slash command whose executable source is verified against the current default-branch SHA; no workflow_dispatch ref is exposed. Review checks out only trusted runtime code and materializes an exact-SHA patch snapshot before the model starts. Execution uses an isolated Git repository and disables inherited configuration, hooks, credentials, prompts, external diff, and replacement refs. Workflow locks are generated only by pinned gh-aw v0.86.2. |
Source/lock/action-lock freshness, trigger, compiler/action-ref, default-branch, snapshot-only review, literal-path/exact-index, isolated-Git, and live-head checks. | Maintainers / supply chain | Repository implementation | Compiler metadata, generated-diff, trigger, action pin, source-SHA, no-untrusted-checkout, hostile-Git-config, and dependency-policy tests | Upstream or build infrastructure compromise. |
| C13 | T13 | Domain Packs and Depth Profiles can only narrow compiled policy; lattice ceiling is enterprise policy. | Compiler emits effective-policy digest and reason trace. | Accord Compiler | Repository implementation | Property tests proving monotonic narrowing | Misclassified data or risk can choose an insufficient ceiling. |
| C14 | T14, T19 | Activation Lease and Work Accord cap calls, tokens, cost, duration, retries, loops, decomposition, patch size, and parallelism. Each workflow invocation reserves at most 500 AIC; the 1500 daily ceiling permits framing, execution, and review. The hermetic slice commits budget with a signed cost-ledger hold before any provider attempt exists; the attempt binds that hold with holdDigest, and release derives the open set from durable cost-ledger state rather than a caller-supplied unresolved list. After the final awaited head read it fresh-reads signed nonrevoked lease state and validates phase, capability, expiry, and all ceilings immediately before invocation. It reconciles and settles authoritative provider usage even after failure, lost acknowledgement, or disagreement with model-reported usage, holds unknown usage, and releases only the proven remaining balance. |
Signed lease/hold/attempt/reservation/usage/settlement/release records and CAS heads; compiler AIC steering, audit logs, crash recovery, and pause/expiry route. | Kernel / runtime / redeemer | Repository implementation | One-call and exact-token boundaries, token overflow, expiry during final read, per-call revocation, cost/token mismatch, full-cost, insufficient-credit, provider failure, unknown usage, lost acknowledgement, settlement/release forgery, replay, loop, repair, recursion, concurrency, and freshness tests | Provider reporting may lag; deployed reconciliation remains required. |
| C15 | T15 | One ordinal-bound effect per plan where practical; prerequisite reads and idempotent API operations. Closed-schema signed evidence validation rejects impossible or identity-substituted persisted states. Commit creation consumes the exact signed patch bundle and completion requires fresh typed commit/tree/blob/content/patch evidence. Draft-PR completion retains immutable creation-time base evidence; only re-observation of an already-completed draft-PR effect may accept a newer live base tip. That external replay still requires the exact current one-revision PR-bound canonical binding digest plus fresh repository numeric/node/full-name, Issue number/node, Project owner/node/item, and PR/base/head repository/ref/node/number/head-SHA/draft/open identities. Active verification, review, approval, merge observation, closure, pre-effect delivery, and pending reconciliation require exact bound-base SHA. Every effect family requires a canonical target-specific applied observation; wrong-target, stale, and stable no-op observations fail. The first invocation stops at a signed awaiting-human checkpoint. Awaiting-human, pre-release, closure, settlement, and release evidence remains valid until explicit signed expiry, while resume freshly validates mutable policy, lease, head, approval, and merge state. After independent approval and merge observation, a discoverable signed pre-release checkpoint with a stable release key is durably updated with the release receipt before a signed closure checkpoint permits Project, Issue, delivery, operations, and final Kernel convergence without rerunning model phases. Definite retryable rejections preserve an owned claim and the complete authenticated server minimum retry deadline; deadlines beyond automatic policy defer closed and ambiguous outcomes are never blindly retried. | Typed target/postcondition observations, retry-not-before/external-window/retryable/partial evidence, public binding lookup of awaiting/release checkpoints, authenticated closure resume, and deterministic read-after-write reconciliation; conditional state transitions never overwrite history. | Single Writer | Repository implementation | Signed-bundle commit round-trip, completed-effect base-tip advancement, active-phase base rejection, repository/Issue/Project/binding/PR aggregate substitution, wrong/no-op/tree/content/patch and every-effect target/postcondition tests, long-wait/explicit-expiry, missing/out-of-order human events, lost release acknowledgement, invalid-state and modified-state evidence, closure restart at every effect, HTTP 403/429 rate-limit, malformed/overflow deadline, timeout, ambiguous 5xx, full retry-delay boundary, retry exhaustion, CAS conflict, partial-effect, and read-after-write fault injection | Some platform effects are not reversible. |
| C16 | T17 | Capability denylist for publication, deployment, production/CRM/ERP/payment mutation, license/visibility change. Provenance policy gates reuse. | Release-candidate license/SBOM/secret/privacy/trademark/similarity review. | Kernel / legal / release | Repository implementation | Prohibited-effect tests and signed readiness checklist | Human/manual release can still violate policy. |
| C17 | T18 | Model policy pins one reviewed model with no fallback and forbids treating its detector or reviewer role as independent human review. | Workflow metadata and future Evidence Ledger record provider/model/context lineage; human review remains required. | Policy Gate | Repository implementation | Manual eval records reject same-model evaluator/subject identity; deployed lineage tests remain | Providers can share latent training/correlation. |
| C18 | T01-T25 | Fail closed on unknown schema/version/state/capability/identity/evidence, missing evidence authentication or conditional storage, and control unavailability. | Stable refusal taxonomy, authenticated evidence-chain diagnostics, alert, and human recovery route. | All deterministic planes | Repository implementation | Fault-injection matrix showing no success-shaped fallback or token mint before required controls | Availability decreases under ambiguous conditions. |
| C19 | T01, T02, T04, T06, T09, T10, T13, T14, T17, T21 | Canonical profile selection, immutable compiled authority, exact artifact-policy evidence, canonical request grants, authenticated usage, structured reviewer fields, DLP, claims/rights authority, role separation, human gates, and exact repository/ref/tree/diff checks keep Domain Packs policy-subordinate and repository-only. | Signed evidence is revalidated at every model, package, COMMENT, approval, merge, and closure boundary; ambiguity or drift fails closed. | Domain adapter / Policy Gate / Kernel / redeemer | Repository implementation | Domain schema/compiler/runtime and adversarial tests, including policy digest mutations, delimiter-like content, replay, identity substitution, role aliases, and zero external calls | Semantic classifier false negatives, human collusion, compromised source/approver/provider accounts, and local host compromise remain; authoritative services and live controls are undeployed. |
| C20 | T04, T23 | Repository effects require a configured durable store. The adapter snapshots and authenticates a challenged current store head, requiring exact genesis (0,null) or non-genesis (positive,Digest), then atomically claims the exact signed operation-grant redemption key using a distinct challenge and expected predecessor/sequence. It snapshots the returned claim before validating appended CAS status, store identity, transition, grant/request/repository/run/operation identities, and time. A canonical claim digest is required in the causally later package receipt. |
Replay, impossible chain state, split-read/accessor substitution, stale challenge, conflict, wrong-store, malformed, stale, forged, or causally misordered evidence blocks before ref CAS or ledger append; operators reconcile the signed claim head. | Trusted grant store / adapter | Repository implementation | Schema/runtime validation; valid genesis/non-genesis; impossible (0,Digest) and (positive,null); wrong-store, cached-challenge, expiry, wrong-purpose, invalid-signature, single-snapshot, and malformed-receipt refusals before claim/mutation; challenge/predecessor/sequence and causal-time checks; replay tests |
The store service, trusted challenge source, and independent key custody remain undeployed. |
| C21 | T14, T16, T22 | Closed hash-chained audit events omit raw identities and arbitrary text. Event creation, serialization, and deterministic fixed-label metric derivation use one canonical snapshot per caller event. Redaction removes unknown key names/values, retains only field-specific enum/digest/counter values and exact labels/usage/projectedUsage objects, preserves all canonical Budget Decision digests including nullable predecessor evidence, rejects anonymous roots, and caps UTC timestamps at 24 characters/millisecond precision. Budget evaluation snapshots its complete input before checking attempts, fanout, concurrency, tokens, tools, effects, wall clock, expiry, and exact run start. |
Chain discontinuity, split-read/accessor substitution, invalid schema, overflow, exhausted limits, malformed containers, anonymous values, unbounded or sub-millisecond timestamps, or redaction failure blocks emission or work; operators alert on fixed reason codes. | Trusted adapter / observability sink | Repository implementation | Canonical Audit and Budget Decision preservation, one-read mutable budget/event/metric regressions, millisecond timestamp schema/redaction, cardinality, invalid-container/digest, anonymous-root, secret-field, startedAt, and every-ceiling tests | Deployment transport, retention, alert thresholds, and sink availability require administrator validation. |
| C22 | T01, T05, T08, T12 | Privileged automated review checks out only trusted default-branch runtime code long enough to materialize one bounded exact-base/head comparison and recheck both identities. Before inference, trusted code buffers the exact reviewer profile and two required skills, deletes every workspace entry, recreates only those supported discovery files plus review-target/evidence.json, removes duplicate activation and prompt source/import material, and verifies the exact four-file tree. Source pins Copilot CLI 1.0.79 for review and threat detection, supplies --no-auto-update, disables bash, edit, CLI proxy, and ambient custom instructions, and supplies write/shell deny flags; the exact reviewer profile omits write/shell tools, so denial precedence is not authoritative. Generated execution has neither all-tools nor all-paths authority and exactly two add-dir roots. The model can read only declared evidence, profile, skill, and unavoidable pinned gh-aw runtime-control scopes, has no GitHub reads, and cannot select a target. Generated Safe Outputs stay staged with an empty permission map, so only the trusted bridge may construct a current-head-bound COMMENT effect. |
Materialization fails on head/base drift, unsafe workspace identity, missing/symlinked/oversized trusted files, purge error, residual entry, or evidence limits. The bridge checks current head again; validation rejects untrusted checkout, either exact-version installer mismatch, compat lookup, missing no-auto-update flag, effective-version mismatch, a non-exclusive tree, undeclared add-dir, post-guard setup-step drift, model GitHub/shell/edit/all-path authority, non-staged output, or generated output permissions. | Runtime / Policy Gate | Repository implementation | Official 1.0.79 archive/executable, effective no-auto-update version, and complete v0.86.2 setup-JS-tree digest probe; harness sibling/archive/effective-version mutation tests; generated/probe flag parity; fixed child-env allowlist; negative missing-agent test; 160-bit undisclosed evidence-head echo; live present mode-000 outside-root sentinel with post-launch path/descriptor dev/inode/type/mode/size/content checks, output/log secret scanning, shell-marker filesystem check, and safe descriptor cleanup; delete/truncate/rewrite/replace/chmod/symlink/error-preservation fixtures; byte-stable gh-aw validation; exact four-file shape and seven-step residual control-surface sequence; source/lock/tool/staged/no-permission assertions; exact-head tests; CodeQL |
Pinned gh-aw/CLI, its action/harness/tool/prompt/config/log control root, and runner isolation remain trusted dependencies. The denial challenge does not prove a model tool call occurred or make --add-dir a read sandbox. Patch APIs may omit binary/truncated content, so missing evidence and human review remain explicit. |
| C23 | T02, T03, T04, T05, T14, T15, T16, T24 | Closed installer config independently binds exact release source and enterprise/organization/repository numeric/node/full identity, installation, ref, target and trusted-precomputed result heads, current state, optional last stable recovery state, receipt CAS head, release and migration manifest digests, selected directional step checksums, operation, and version. Every untrusted input is snapshotted once and every path layer rejects absolute/traversal/empty/.git segments. Raw receipt arrays require a primitive nonnegative safe-integer length capped at 512 before element access or cloning, then are materialized once so structure and the non-empty authentication decision use the same snapshot; no chain reset/compaction protocol is implemented. The plan embeds its digest-bound selected configuration, release and migration manifests, complete expected pre/post states, and nullable last stable recovery state. CLI is plan/offline-validate-only; authenticated live validation requires the injected read-only trusted adapter, current authorization, fresh observation, and signed bound evidence. A new effect requires current explicit human authorization through an injected trusted adapter that alone owns App credentials and time; the caller rereads time immediately before apply and the adapter must atomically enforce expiry, expected state/head, and idempotency with the effect. Exact persisted receipts reconcile without a new effect after authorization expiry. One contiguous checksummed migration graph uses bounded arbitrary-precision ordering, ends at currentVersion, and declares preconditions, irreversibility, rollback, and evidence retention; uninstall requires a known matching installed manifest and removes only exact package-owned files. | Apply-side validation re-runs migration selection and re-derives exact actions, four preconditions, and the deterministic operation-evidence path from bound configuration/release/migration/current/result/recovery contracts before verifier or adapter access. It requires stable input outside recovery, requires partial input plus a stable recovery base for recovery, then independently enforces strictly sorted unique actions, create/update/delete digest semantics, operation/action compatibility, effect/result-head consistency, journal capacity, and disjoint action/retained-evidence paths. Canonical plan digest and complete effect/result/evidence idempotency digest, explicitly structural versus authenticated journal APIs, trusted verification of every non-empty journal receipt signature/target/stable-state/sequence/predecessor/terminal head over the same bounded snapshot, signed live-validation evidence, signed applied receipt, authoritative receipt-store reread after apply/recovery, exact result-head/state reconciliation, and independent partial-state backup binding fail closed before effects. Partial/ambiguous state or exhausted journal capacity permits only documented human recovery; backups and evidence are retained. | Packaging planner / trusted installation adapter / human admin | Repository implementation | Aggregate-bound/changing-length/coercible-length Proxy and overridden-array-method access, accessor/verifier mutation substitution, raw path/Git metadata, schema/planner/apply/release-path drift, source/target/result-head/time/state/recovery-base/release/migration/configuration substitution, self-rehashed arbitrary-customer action/idempotency/evidence/current-operation collision, numeric version overflow, awaited-check authorization expiry, offline/live distinction, exclusive authorization expiry/signature and expired idempotent recovery, CAS, replay, missing receipt persistence, receipt signature/target/state/head/gap/reorder, disconnected/skipped/unknown/downgrade/irreversible migration, lost-ack, plan-tamper reconciliation, repeat apply, sequence-zero/journaled partial recovery, unknown/mismatched uninstall, ownership, and evidence-preservation tests | Trusted adapter, credential broker, signer, durable store, authenticated checkpoint service, result-head precomputation, and customer admin configuration remain deployment prerequisites. Human admin effects are external. |
| C24 | T11, T12, T16, T17, T25 | Exact clean canonical Git top-level commit is packaged from aggregate-bounded regular blobs only with Git 2.46+, replacement objects, partial-clone lazy fetching, and repository fsmonitor disabled and output outside the source tree. The schema's registered release-path format delegates to the shared assertReleasePath semantic validator, which requires canonical NFC Unicode, rejects controls plus U+2028/U+2029, and requires UTF-8 representability using a ustar name of at most 100 bytes or slash-boundary prefix/name fields of at most 155/100 bytes. Byte-canonical deterministic ustar with strictly sorted unique paths, closed manifest, checksums, snapshot-validated SPDX SBOM, source/base/head provenance, unsigned in-toto statement, compatibility matrix, snapshot-validated open-source assessment, and no-go candidate checklist are generated without network, credentials, signing, or publication. Unknown platform/tool versions fail closed. | Central document validation, planning, writer, and verifier reuse the semantic path validator. Independent byte reproduction and verification reject traversal, malformed/control/line-separator/non-NFC Unicode, unrepresentable or noncanonically split paths, links/submodules, duplicate/unsorted/unexpected paths, noncanonical ustar header/terminator/type/mode/owner/time/padding drift, tampering, preallocation oversize, subdirectory-root/in-source output, stale source/head, replacement refs, missing local objects, unsupported Git, fsmonitor execution, dependency/license expression/notice drift, SPDX/readiness accessor substitution, evidence binding substitution, build-time trust flags, and any trusted-attestation claim over unsigned output. | Release tooling / human release, security, product, OSPO and legal owners | Repository implementation | Exact ASCII/multibyte name and prefix boundaries, schema semantic format refusals, U+2028/U+2029, BOM-preserving field round trips, Unicode normalization, canonical split, determinism, independent-root byte comparison, duplicate-path/ignored-header/surplus-terminator/aggregate archive, Git-top-level/in-source output, Git-version/no-lazy-fetch/fsmonitor controls, SPDX/readiness accessor, archive and replacement-ref adversarial fixtures, exact-tree/evidence/SPDX regeneration, invalid-license and unsigned-trust refusal, command-specific flag parsing, schema/docs/toolchain drift validation, and license/notices integrity hashes | Local host and complete local Git object database remain trusted for local evidence. Production signing, transparency, release service, publication, license, and readiness decisions are not implemented. |
| C25 | T02, T04, T05, T09, T12, T13, T26 | The autonomous demo foundation fixes one four-entry catalog, complete stage/control reservations, globally injective model agent/capability/workflow identities, explicit empty non-model bindings, content-addressed immutable contracts, and independent complete per-demo shard pairs. Trusted binding metadata—not workflow filenames or self-declaration—selects exact validation classes. Same-core journey receipts cannot change Kernel state or authority; cross-core advancement requires the exact applied Kernel result and receipt. Project fields are projections and Kernel Stage is converged last. | Strict duplicate-key JSON ingestion, schema and semantic closure, exact source-to-binding coverage, capability-phase checks, receipt-chain verification, applied-result/snapshot comparison, and deterministic refusal block substitution before inference or effects. | Foundation validators / Kernel-subordinate demo runtime | Autonomous demo Foundation | Positive and adversarial tests for canonical order, missing/unknown versions and digests, immutable snapshots, duplicate/reused identities, cross-demo/generic-agent substitution, zero/single-demo shard closure, model control fields, same-state misuse, Kernel receipt mismatch, and Stage-last projection | All four demo shards and runtime integrations are installed. Live runtime activation, App credentials, trust services, and paid inference remain undeployed human/deployment work. |
| C26 | T01, T02, T05, T14, T26, T27 | Four static issue forms bind by trusted filename to the exact Foundation demo/profile reference and declarative schema digest. Inputs are a closed bounded data object; repository text is a hint only. Deterministic preflight validates exact profile/catalog closure, consent, enabled/current activation, immutable submitter ID, allowed depth, nonzero fixed budget, resolved repository binding, and fresh digest-matched Project binding while keeping credential, reservation, inference, and issue-creation authority denied. The four Project schemas contain fifteen fields, structurally separate Requested Stage Agent input from fourteen trusted projections, preserve core Stage options, derive Journey Stage options from reservations, require an explicit supported display-only color for every single-select option, bind each projection source, and write Stage last. Option names and descriptions remain the accessible semantic source; colors are excluded from all authority decisions. | Missing or malformed input returns a closed reason code before privileged boundaries. One empty required field yields one deterministic DemoMissingInformationRequest bound to the issue node, form ID, and submission digest; no punctuation heuristic or automatic issue creation exists. Catalog and exact-target bootstrap planning/readback remain dry-run and human-admin-only. Target manifests bind each Project-schema digest, setup bindings retain exact option node IDs/names/colors/descriptions, and omitted, unsupported, or drifted color data blocks binding instead of defaulting to gray. | GitHub adapter preflight / human administrator | Project UX | Exact catalog/form/snapshot fixtures and adversarial tests for form/profile substitution, unchecked consent, disabled/expired activation, unauthorized submitter, unresolved/substituted repository, stale Project binding, malformed or oversized content, disallowed depth, missing budget, data-only repository hint, single typed clarification, no authority grant, dry-run flags, complete palette and cross-demo consistency, omitted/unsupported/mismatched color refusal, exact target/color/description/identity drift, and readback | The supported API can rename views and configure visible fields. Grouping and Journey Stage column ordering require exact manual UI steps. Runtime activation remains human/deployment work. |
| C27 | T01, T02, T04, T05, T08, T09, T12, T14, T26, T27 | Portfolio convergence requires the exact four complete demo contract sets, every Foundation reservation, 23 exclusive runtime candidates, exact source/lock pairs, fixed command catalogs, signed pack review evidence, Kernel-first receipts, and Stage-last projection. Runtime scripts select an opaque binding only from the validated catalog/profile/stage/candidate tuple. Review evidence derives from two stable trusted-reader observations and must match every canonical repository, work-item, and Project identity retained by fresh expiring signed target evidence. The privileged bridge runner is Git-only; non-Git commands require a separately isolated credentialless network-denied runner. COMMENT application atomically binds the complete fresh execution-state, plan, effect, binding, and expected-head digests. The Kernel recognizes the fixed pack predicate/evidence vocabulary and rejects unknown names. | validate:demos rejects omissions, partial pairs, duplicates, substitutions, fallback identities, workflow/action-lock drift, and missing recovery fixtures. simulate:demos invokes the actual Kernel for every cross-core happy-path boundary under distinct pack Work Accords, invokes a signed fake provider inside exact fences, continues the exact Human Review snapshot, uses typed consumers and expected refusal patterns for 96 directed substitutions, matches each recovery case to executed tests, emits canonical audit/metric chains, and validates closed zero-valued fixture-declared external-call assertions. | Portfolio runtime / Policy Gate / human administrator | Portfolio integration | tests/demo-integration.test.ts, hybrid-selection tests, Control Kernel requirement tests, bounded-runner isolation tests, exact-head and atomic-precondition Single Writer tests, repeated byte comparison, full repository matrix, LICENSE and action-lock checks | App installation credentials, billing, rulesets, protections, isolated command service, trust services, and a sandbox canary remain post-merge human-administrator gates. Projects remain non-authoritative after configuration. |
| C28 | T01-T28 | A closed non-authoritative hardening plan enumerates the complete all-demo adversarial matrix, every credential/claim/reservation/provider/persistence/effect/acknowledgement fault boundary, and the readiness classification. It can select only exact tests, never runtime authority. | validate:hardening rejects duplicate-key/open plans, omissions or reorder, failed/errored/skipped/todo/missing/duplicate tests, non-deterministic simulator bytes, incomplete cross-demo refusals, failed projection/review/runtime invariants, or a missing/renamed/extra/cross-demo/nonzero fixture-declared external-call assertion. Ambiguous effects require stable exact readback or remain blocked. | Portfolio validation / security owner / human administrator | Hardening | Closed schema and semantic tests; all configured actual test files content-bound to executed JUnit; Kernel/runtime/adapter/writer/project/delivery/recovery boundaries; two identical real simulations; per-demo simulator digests; fault-boundary results; four schema-valid exact-four-category zero assertion fixtures | Fixture assertions are not runtime telemetry or an OS sandbox. Credentialed live faults, deployed isolated-runner/service availability, platform drift, provider behavior, and a real canary remain human-admin sandbox gates. |
| C29 | T02, T04, T05, T09, T12, T14, T28 | StageAgentBindingSet@2.0.0 declares closed none, fixed, and user-selectable participation with no fallback. Enterprise and project policy, Work Accord capabilities, Phase Contract, Capability Registry, lease, budget, actor authorization, exact Project item, run/attempt/generation/receipt head, and current PR head intersect before one signed exact-agent grant is issued. Requested Stage Agent is untrusted input; dispatch and the run fence bind the accepted grant digest. Its blue discovery and purple implementation colors are display cues only and never select a candidate or stage. Static workflows never use dynamic engine.agent. | Typed selection refusal, stale/replay reconciliation, policy-generation invalidation, Stage-last trusted projection, exact-target bootstrap planning, explicit human confirmation, and complete readback. | Policy compiler / trusted adapter / scheduler / human Project administrator | Hybrid agent governance | tests/hybrid-agent-selection.test.ts, tests/adaptive-delivery-demo.test.ts, and tests/github-project-bootstrap.test.ts; wrong-stage, unauthorized actor/bot, missing selection, cross-target, stale generation/head, duplicate, conflicting replay, dynamic-agent, color-only drift, plan drift, and readback tests | A deployed trusted adapter must authenticate Project observations, persist accepted grants, inject the grant digest into the static workflow, and operate independent stores. GitHub Project view settings require manual UI configuration. |
| C30 | T02, T03, T13, T24, T29 | DeploymentTopologyPlan, GitHubAppRegistrationPlan, GitHubAppInstallationTargetBinding, and AdministratorPlan (ADR 0013) fix the exact eight independent trust-service identities (each with a fixed principalId plus unique signingKeyId/oidcAudience), four durable-store identities (each with a unique namespace/credentialId), and least-privilege permission union derived only from the reviewed GITHUB_PERMISSION_MANIFEST operation table; none is caller- or model-suppliable. Ruleset ref conditions accept only literal, fully-qualified ref names — no wildcard/glob metacharacters — since exact-set ref subtraction is unsound over globs this contract does not evaluate. A GitHubAppInstallationTargetBinding is a separate, human-approved, immutable owner/App/installation/repository-set identity, never derived from a readback. Every plan and readback carries an explicit nonAuthoritative marker and cannot install/transfer/authenticate an App, mint a token, apply a ruleset/required check/Actions policy/GHAS setting, or mutate a Project; a readback silently reporting false for a field the marker requires is compared field-by-field and flagged like any other drift. Administrator rulesets are modeled as a closed branch/tag discriminated union, since GitHub exposes pull-request/CODEOWNERS/current-head controls only for a branch target and restrict-creation/-update/-deletion controls only for a tag target — an impossible tag ruleset carrying branch-only controls is rejected at the schema, generator, and comparator layers. | validateDeploymentTopologyPlan and validateAdministratorPlan reject an ingested document that omits, duplicates, or adds a service, store, budget, monitoring signal, retention window, protection scope, ruleset bypass actor, required check, or that shares a signing key/audience/namespace/credential across the closed set, that declares a conflicting duplicate retention window for the same artifact kind, that declares a wildcard/glob ref, or whose ruleset ref conditions do not resolve to its declared protected refs. validateAdministratorPlan and compareAdministratorReadback also reject a duplicated rulesetId/environmentId/checkName/incident-contact-role before any lookup Map is built from that collection, so a conflicting duplicate cannot be silently discarded last-wins. validateGitHubAppRegistrationPlan independently re-derives the manifest union and rejects any ingested drift. compareGitHubAppPermissionReadback and compareAdministratorReadback bind planDigest/repository identity/target binding and a caller-supplied freshness window before any field comparison — and, for compareAdministratorReadback, validate the supplied plan itself first — then reject a duplicated observed permission key (or rulesetId/environmentId/checkName/incident-contact-role) before any per-key comparison, then fail closed on a denied, elevated, or downscoped permission, an extra or missing permission, a stale manifest version, a mismatched installation target, a stale/future observation, a ruleset that is inactive, unobserved, ref-drifted, source/target-mismatched, or (for a tag ruleset) missing a restrict-creation/-update/-deletion control, an unplanned extra ruleset/environment/required check absent from the plan, an Actions policy that allows all Actions, grants review approval, or disables fork-PR approval/SHA pinning, a disabled GHAS control, an environment with no protection rules, Project mutation-confirmation drift, a substituted incident contact, or a false nonAuthoritative marker. A readback's own driftFound is checked only for coherence by checkReadbackDriftCoherence, never as an override. | Deployment planner / GitHub App broker / human administrator | Pre-App contracts | tests/deployment-topology.test.ts, tests/app-registration-plan.test.ts, tests/administrator-plan.test.ts, tests/pre-app-api-surface.test.ts; omitted/duplicated/extra service-store-check-scope, shared-identity, principal/kind-mismatch, conflicting-duplicate-retention-window, narrowed monitoring-signal, over-ceiling journal bound, ingested-manifest-drift, denied/elevated/downscoped/extra/missing permission, duplicated-permission-key, target-binding mismatch/expiry, observation staleness/future-dating, stale-manifest, bypass-actor, inactive-ruleset, ruleset source/target mismatch, branch/tag discriminated-control drift, unplanned-extra-ruleset/environment/check, false-nonAuthoritative-marker, wildcard/glob-ref refusal (including a wildcard exclude that would otherwise misstate a literal include as protected), ref-condition drift, duplicate-key refusal with conflicting values for rulesets/environments/checks/contacts, unvalidated-plan refusal, Project-binding/incident-contact drift, driftFound-coherence, and public-export tests; npm run validate:schemas additionally exercises the six synthetic examples/pre-app/ fixtures including a target binding and one intentionally drifted readback | The eight trust services, the App installation, and every ruleset/check/Actions/GHAS/Project configuration remain undeployed human-administrator work; an authenticated readback still depends on a separately deployed trusted adapter and a separately human-approved target binding this repository does not provide; the freshness check trusts a caller-supplied clock; a ruleset requiring a real glob/fnmatch ref pattern cannot be represented and must be decomposed into literal-ref rulesets. |
| C31 | T04, T23, T30 | The local durable trust substrate (ADR 0014) is mechanism, never authority: it stores exact caller-produced canonical bytes and returns them unchanged, signing nothing, minting nothing, reading no clock, and repairing no record — every store port's existing caller in src/ still recomputes and re-validates the returned receipt head, so the substrate cannot introduce contract semantics. All fifteen ports reduce to an idempotent appendOnce and an exact compareAndSwap, both returning appended/existing/conflict, where existing requires byte-identical canonical bytes so a replay carrying a mutated body can never satisfy a caller's idempotency check. Writes run inside BEGIN IMMEDIATE … COMMIT under WAL with synchronous = FULL, re-checking the expected head inside the transaction so a moved head yields conflict rather than a lost update; because DatabaseSync is synchronous and signers are asynchronous, no await may appear inside that window and callers sign beforehand. Store identity, backend namespace, credential identity, isolation, and journal bound are consumed from a validated DeploymentTopologyPlan (ADR 0013) rather than from a second descriptor, and binding refuses an omitted/duplicated/unknown store, a shared namespace or credential, a store shared with a model runner, or two stores pointed at one file. Opening requires both a supported Node major taken from the repository compatibility matrix and a live capability probe (DatabaseSync, backup, WAL, synchronous=FULL, user_version, integrity_check, conditional-update hit-then-miss); there is no filesystem fallback, no degraded mode, and no default path. node:sqlite is confined to one backend module so the release-candidate surface can be replaced without touching an adapter. Concurrent first-open provisioning of the same absent file is serialized behind the identical BEGIN IMMEDIATE write lock rather than the prior lock-free check-then-act: an optimistic read still skips the lock for an already-provisioned store, but an uninitialized read re-checks user_version inside the lock before creating schema and identity, so every racing creator converges on exactly one identity — including a genuine identity/namespace/bound conflict, where the loser takes the same STORE_BINDING_INVALID path an ordinary mismatched reopen would — with no raw UNIQUE constraint failed: durable_meta.id driver error, blind mutation retry, partial metadata, lock leak, wrong-identity acceptance, or weakened open verification (first-open race hardening). | Every read re-derives the stored digest from the stored bytes and refuses non-canonical, truncated, non-UTF-8, or rewritten content; PRAGMA integrity_check runs at open and a user_version mismatch refuses instead of silently upgrading, and a store whose identity row is missing after schema creation refuses STORE_CORRUPT rather than accepting an absent or asserted identity. Chain checks reject the impossible states (0, digest) and (positive, null), sequence gaps, and broken predecessor links. A full journal raises CAPACITY_EXHAUSTED and never evicts, so replay evidence cannot be destroyed to make room. An unknown write outcome raises a distinct DurableAmbiguousAcknowledgementError that is never collapsed into success or failure, and the substrate guarantees a stable reread so the caller's double-read reconciliation resolves it; ambiguity concerns durable content rather than connection state, so the commit path rolls back best-effort before raising it, releasing the cross-process write lock and keeping later operations inside the typed taxonomy instead of leaking a raw cannot start a transaction within a transaction, and reconciliation is defined as reading through fresh handles so a same-handle read cannot observe uncommitted rows; a lock timeout is reported separately as STORE_UNAVAILABLE because the transaction never opened. A commit that fails after being attempted during first-open provisioning raises the identical ambiguity a write's commit does, after the same best-effort rollback that frees the lock; unlike a write, a failed open has no caller-visible key to reread, but retrying the open is always safe because the retried transaction re-derives its outcome from whatever is actually durable. A produced backup manifest and journal record each validate against their own published schema, with the manifest's namespace list bounded by the store-wide journal ceiling rather than a smaller cap that backup() could exceed. Backups carry a closed manifest binding store id, format version, and per-namespace head, and a restored copy re-verifies its chain. | Durable store substrate (nonproduction reference) | Pre-App durable substrate | tests/durable-substrate.test.ts, tests/durable-multiprocess.test.ts, tests/durable-recovery.test.ts, tests/durable-api-surface.test.ts; multi-process proofs by real forked processes that racing appends and compare-and-swaps yield exactly one appended with no lost update, byte-identical-replay-existing, mutated-replay-conflict, stale-head-conflict, tampered-body/non-canonical/truncated/non-UTF-8 refusal, integrity_check and user_version refusal, impossible-chain-state and sequence-gap refusal, capacity refusal without eviction, ambiguous-commit rollback proven by fault injection with lock-release, typed-subsequent-write, fresh-handle reconciliation and actual-durable-outcome assertions, produced-manifest and produced-record schema conformance at the store's exact 512-namespace upper bound with the 513th refused, restart continuity including a held fence that stays unwinnable, backup/restore and disabled-state recovery, unsupported-runtime and capability-probe refusal, absolute-path-only enforcement, and static source proof of no environment/network/credential/ambient-clock access, single-module node:sqlite confinement, no await inside a transaction (asserted independently for the write and first-open provisioning transactions), and absence from the public barrel; concurrent first-open provisioning by real forked processes converging on exactly one identity for a same-identity race and on exactly one deterministic winner (with the losing identity refusing STORE_BINDING_INVALID, never accepted) for a conflicting-identity race, plus single-process proofs of a lock timeout during creation refusing STORE_UNAVAILABLE, a failing commit during creation raising ambiguity with the lock released and a safe retry, a store missing its identity row refusing STORE_CORRUPT, and deterministic fault-injected proof (independent of process-fork scheduling) that a same-identity or conflicting-identity concurrent creator landing in the exact window before the write lock is observed and converged upon rather than raced against; npm run validate:schemas additionally validates the two synthetic examples/durable-stores/ fixtures and re-derives the journal record's chain link | This is a nonproduction local reference store, not a deployed service, and carries no SLA claim; the deployed durable store of prerequisite item 5 remains undeployed human/deployment work. node:sqlite is a release-candidate built-in, mitigated by single-module confinement and an open-time capability probe but still subject to upstream change. Filesystem-level protection of the store path, and the operator response to a CAPACITY_EXHAUSTED or STORE_UNAVAILABLE refusal, remain outside this repository. |
| C32 | T11, T12, T16, T17, T25, T31 | A CustomerStarterSelection is a closed, human-authored list of included/excluded release-path prefixes bound to an exact sourceHeadSha and a resolvedClosureDigest that binds the selection's own resolution at that exact commit (excluding the selection document's own self-referential entry, detected structurally by kind/profileId); an extension additionally binds extendsProfileId and a baseSelectionDigest to an exact base. buildCustomerStarterBundle/verifyCustomerStarterBundle (ADR 0015) independently re-resolve resolvedClosureDigest at both the reviewed sourceHeadSha and the current build head — ancestry of sourceHeadSha is not by itself sufficient evidence that no matched file was added, removed, mode-changed, or content-changed since review. A closed, trusted profile catalog (src/customer-starter-catalog.ts) is the only place a profileId's exact selection-document path, extendsProfileId, advertised scripts, and the two shared scan-denylist paths may be declared; buildCustomerStarterBundle/verifyCustomerStarterBundle accept only profileId -- no parameter accepts a selection object, a base selection, a base manifest, a knownSelectionDocumentPaths exemption list, a denylist, an advertised-script list, or the catalog itself, closing both a caller-suppliable-catalog exemption attack and a subsequent gap where an exported "test-fixture-only" function accepting a catalog was itself compiled into the shipped package and deep-importable regardless of its name. src/customer-starter.ts resolves its own module-private, recursively-frozen catalog reference (built independently from, and sharing no object graph with, src/customer-starter-catalog.ts's own exported, frozen, detached inspection copy), which is never exported and therefore unreachable by any deep import; mutating the exported inspection copy throws (ES modules are always strict mode) and cannot influence the engine's own reference regardless. Every selection's exact bytes are read directly from the exact reviewed Git tree rather than a caller-supplied object or the working directory. The module-import closure fails closed on any captured module specifier (static import/from, dynamic import(), require(), or a no-substitution template literal, in any quote/backtick delimiter) containing a backslash escape sequence, since a sequence such as \x2e decodes to a leading . at runtime without containing a literal . in its raw source form and this checker has no reviewed string-escape decoder. The shared output/verify path-safety helpers in src/release-support.ts (safeOutputPath, assertSafeOutputRoot, canonicalDirectory, writeExclusive, used identically by src/release.ts) require a build's output parent, and a directory being read back (a verify bundleRoot or the source repositoryRoot), to be a non-symlink directory owned by the current process user and free of group/other write bits, re-validating identity after every realpathSync resolution or directory creation and creating the output directory exclusively with a restrictive mode, closing a check-then-use (TOCTOU) symlink-swap window under a shared writable parent (e.g. world-writable /tmp) -- not a claim of protection against an attacker who already has code-execution as the same user/process identity. An extension's base scope is always independently re-resolved from the reviewed base CustomerStarterSelection (itself loaded through the same catalog) against the exact tree, and its manifest is derived entirely internally by the engine — there is no caller-suppliable base manifest value left to widen or narrow. Every included prefix must match a real path; every excluded prefix must actually carve something out of an included one; an extension's own files must be disjoint from its base; and the shared Git-tree listing (src/release-support.ts, used by both this tool and src/release.ts) rejects any two tree paths that collide under NFC normalization and case-fold on portable extraction. Content is fetched only for the resolved subset, then eight deterministic checks run before any bundle file is written: secret, internal-reference, and customer-data scans; generated-workflow lock/Markdown-source, JSON Schema $ref, module-import (TS/MJS/CJS/JS static and dynamic import, re-export, and require, resolved over a comment- and string-literal-aware source redaction rather than a bare regex), Markdown-link (resolved via markdown-it's own token tree, covering reference-style links/images and their definitions), and profile-scoped package-script closure. CustomerStarterPreflightReport is fixed decision: "no-go", authoritative: false, selfApproved: false, and embeds the live OpenSourceReadinessAssessment by exact digest with its unchanged 9-category gate. | Any scan hit or closure violation throws before any bundle file is written, so a produced report's eight scans can only ever read "clean" — there is no "flagged but shipped" state. checkPackageScriptClosure validates exactly a profile's declared advertisedScripts, not every package.json entry, so a demo-only script cannot force a core-only bundle to fail or silently ship unusable tooling. Byte-identical rebuild, resolved-closure drift refusal, catalog-bound profileId/selection-document resolution (rejecting an unknown profileId and a post-review file added under a reviewed prefix with no exemption parameter available to hide it), path-collision refusal, and embedded-readiness-digest binding are independently re-verified. | Customer-starter engine / human legal-OSPO-security-product reviewers | Customer starter and open-source preflight | tests/customer-starter.test.ts, tests/packaging.test.ts; selection/denylist validation, all eight closure/scan checks individually (positive and negative, including reference-style Markdown links, .mjs/.cjs/require(), and comment/string-literal false-positive refusal), byte-identical build/verify round trip, out-of-tree selection refusal, resolved-closure drift refusal after the reviewed commit, catalog-sealed profileId/selection-document resolution including a post-review evil-file attack, a compiled-output test-fixture-only-export refusal, and a frozen-catalog runtime-mutation-cannot-influence-output test, plus escape-sequence and TOCTOU symlink/writable-parent/owner/mode regressions, portable-extraction path-collision refusal, and base/extension composition including the disjoint-overlap refusal; a repeatable validate:customer-starter-extraction script independently extracts each real profile with no Git history and runs npm ci plus every advertised script inside it, retaining a JSON evidence record | scripts/installer.ts and tests/packaging.test.ts require a real repository .git history (git rev-parse --show-toplevel, exact base/head matching) and are therefore not part of either profile's shipped, standalone-tested surface; this is a pre-existing property of the installer's own exact-head model (ADR 0009), not a gap this control introduces. validate:workflows and validate:gh-aw shell out to gh aw compile, which itself requires running inside a real Git repository, so neither is advertised as standalone-runnable for demo-portfolio even though both scripts remain shipped; validate:demos is excluded from the advertised list for the same underlying reason, since it unconditionally invokes validate-workflows.js as one of its own steps, and validate:hardening is excluded because its own closed hardening plan mandates tests/packaging.test.ts among its exact test files. validate:review-agent-runtime is excluded because it unconditionally requires COPILOT_CLI_ARCHIVE_PATH/GH_AW_HARNESS_PATH environment variables pointing at external artifacts this profile does not ship or fetch, and eval:behavioral is excluded because it unconditionally requires --responses-dir=<reviewed-response-records> pointing at externally-reviewed response records this profile does not ship. The mechanical closure checks are completeness checks over a given selection, not a proof of global minimality; the exact enumerated selection remains reviewed, human-authored data. No license, publication, visibility, App/Project, or release decision is made or implied by any generated evidence. |
| C33 | T14, T19, T32 | EngineeringCostLedger.hold() writes a signed EngineeringCostHold to the engineering.cost-ledger namespace in runtime-state-store by compare-and-swap before EngineeringProviderUsageLedger.begin() can append any provider attempt to receipt-journal. The attempt and settlement both carry the hold's holdDigest; settlement chains directly to its hold in one ordered EngineeringCostLineageEntry sequence, and release derives open holds from that durable cost lineage rather than from a caller-owned array. expectedOpenHoldDigests is only a caller view checked as a subset, so omission cannot release budget and fabrication is refused. Holds and releases contend on the same cost-ledger head, so either the hold lands and release re-derives it, or the release lands and any later hold is refused because the reservation is closed. | validateCostHold, validateCostSettlement, and release validation enforce the reservation→hold→settlement→release chain, exact holdDigest binding, settled-phase coverage, whole signed unresolvedHolds, and reconciliationRequired when any open hold lacks proven settlement. A hold with no proven settlement is never released from absence of attempt or usage evidence. Stored engineering checkpoints at schemaVersion: "1.0.0" are refused rather than reinterpreted, so operators drain in-flight slices before upgrade instead of synthesizing holds that never existed. | Engineering cost ledger / provider usage ledger | Repository implementation | tests/engineering-slice.test.ts; hold-before-begin ordering, omitted unresolved attempt, crash/no-attempt hold, fabricated expected hold digest, concurrent hold/release, hold-after-release refusal, settlement-chain binding, and checkpoint-version refusal cases | Provider usage can still require human reconciliation when proof is absent. This remains a nonproduction reference slice with no paid inference, App credential handling, live GitHub effect, production SLA, or automatic orphan collection. |
| C34 | T04, T09, T10, T12, T14, T15, T16, T23, T30, T32, T33 | npm run canary:synthetic is an optionless trusted entrypoint that validates the existing DeploymentTopologyPlan, invokes only fixed compiled hardening/durable test files, supplies an exact credentialless child-environment allowlist, and preloads a Node guard denying HTTP(S), socket-level TCP/TLS, UDP creation, DNS, fetch, WebSocket, and EventSource. It reuses the lifecycle/Work Accord/policy/Kernel/adapter/Single Writer and durable-store contracts without adding a route, capability, target, store descriptor, credential source, or effect. A deterministic Ed25519 key derived in memory from a public synthetic label signs target-free evidence and is never persisted. | Two hardening executions must emit byte-identical evidence; two durable executions must contain the same exact passing test set. Every named hardening and durable fault must appear exactly once, every compiled evidence file is pinned to its exact reviewed content digest, topology/plan/runner/simulator/guard inputs are digest-bound, and a failed/errored/skipped/missing/duplicate/content-drifted test, nonzero external-call assertion, unsupported Node major, invalid topology, changed readiness claim, invalid signature, secret-pattern match, or network attempt fails closed. Output records draft-only and COMMENT-only Human Review, no approval/ready/merge/deploy/publish/App/billing/inference/administration effect, and live activation as gated. | Synthetic canary runner / security owner / human exact-head reviewer | Repository implementation | tests/synthetic-sandbox-canary.test.ts, the closed hardening matrix, and the durable deployment/substrate/adapter/composition suites; deterministic/tamper-evident synthetic signature, pre-environment live-option refusal, fetch and raw-socket deny probes, static no-token/no-curl/no-live-effect assertions, all-boundary restart equality, real multi-process races, backup/restore, disabled recovery, corruption, replay/conflict/revocation/stale/wrong-target/budget/provider-unknown/kill-switch/ambiguous/lost-ack proofs, ADR 0016 hold ordering, double-run canonical output, and Node 24/26 full validation | The guard is process-level, not an OS sandbox; selected reviewed tests and the local host remain trusted. Fixture-declared counters are not live telemetry. App installation, deployed trust services, live platform/provider/billing/Project readback, rulesets, protections, and a separately authorized live canary remain human-administrator gates. |
| C35 | T02, T03, T04, T05, T07, T10, T15, T16, T17, T24, T29-T34 | AdministratorHandoffPlan fixes the authority order and complete 27-control catalog while binding canonical deployment/App/admin/fifteen-adapter/canary/customer-starter/open-source/LICENSE evidence. Every control fixes its responsible owner, trusted target source, exact state keys/value domains, readback procedure, rollback mode, and prohibited effects. Every mutation-capable control requires a separate AdministratorApplyPlan with exact applicable immutable target IDs (including transfer source/destination and Project item/field), counted sorted duplicate-free expected and desired values, attempt ID, derived idempotency key, one attempt, durable claim/receipt requirements, and no ambiguous retry. A separate human confirmation binds the canonical plan digest. The library returns structural validation only and has no effect adapter, credential reader, network client, target discovery, clock read, or fallback. | validateAdministratorApplyGate validates the plan, confirmation, expiry, fresh exact-target pre-readback, expected counts/values, confirmation digest, and attempt identity before reporting only readiness for trusted-adapter authentication and durable claim. validateAdministratorPostApplyReadback requires one attempt and binds confirmation, pre-readback, attempt ID, trusted receipt, ordered timestamps, and complete desired-state equality; ambiguous acknowledgement remains reconciliation-required. validateAdministratorHandoffReport re-derives every embedded document digest, gap, gate/readback result, classification, and whole-report digest. AdministratorHandoffReadback requires exact control coverage and rejects synthetic App-backed readiness. The repository handoff uses a synthetic-unconfigured readback; customer live readbacks remain protected external evidence. | Human administrator / trusted adapter / security and release owners | Repository implementation | tests/administrator-handoff.test.ts, tests/pre-app-api-surface.test.ts, npm run handoff:administrator, schema validation, repeated byte comparison, customer-starter closure/extraction, open-source no-go, internal-reference/credential scans, and LICENSE byte comparison | Live target reads, human identity/signature verification, durable attempt consumption, App/key/secret/billing custody, administrative apply, service health, OS runner isolation, customer/legal/OSPO decisions, and a real App-backed sandbox canary remain external human gates. |
| C36 | T02, T03, T04, T05, T07, T12, T15, T16, T35 | The upstream-only IssueTaxonomy contract fixes the authoritative repository full name/numeric ID/default branch, exact three label definitions, explicit #1-#18 mappings, recognized title prefixes, and closed pagination bounds. A least-privileged Actions job runs only after a push to main or issues.opened, checks out the exact event SHA with credentials unpersisted, gives the token only to the optionless reconciler step, and has only contents: read plus issues: write. No pull-request event, PAT fallback, model, arbitrary route, close operation, or customer-starter copy exists. | The reconciler validates event repository/ref/SHA/action/issue identity, fresh-reads all targets before mutation, rejects an unrecognized issue before any write, reconciles only exact taxonomy labels, preserves unrelated labels, and requires exact post-write readback. Concurrency is repository-ID scoped with cancellation disabled; reruns reconcile idempotently after a partial acknowledgement. | Repository maintainers / bounded Actions job | Repository implementation after reviewed merge | tests/issue-taxonomy.test.ts; schema closure, exact historical/prefix classification, wrong target/event/ref/identity, stale issue, no-write unknown title, conflict removal, unrelated-label preservation, workflow permissions/action pins/triggers, and customer-starter exclusion | A compromised GitHub Actions runner, repository administrator, default-branch protection, dependency, or GitHub API can still misuse the scoped token or alter metadata; issue labels remain display-only and are not lifecycle authority. |
| C37 | T02, T11, T16, T24, T25, T26, T31, T36 | ADR 0019 retains one closed technical epoch: Hyperfinite is the product name, while required CompatibilityMatrix.technicalIdentity constants fix agentic-framework/v1alpha1, package/archive names, API/schema origin, reusable Project schema name, Capability Registry publisher, taxonomy User-Agent, synthetic canary seed, synthetic OIDC audience prefix, and domain stem. Separate closed inventory evidence fixes counts and digests the complete content of every identity-bearing file plus its classified occurrences for the authoritative repository and both starter scopes. Existing wire, package, signature, digest, marker, format, fixture, and generated-evidence values stay in that epoch; only human-visible product wording uses Hyperfinite. Runtime, model-output, and migration contracts contain no identifier-epoch selector. | validate:technical-identity scans every bounded regular repository file; enclosing semantic changes, replacements, or removals change reviewed full-file evidence even when the edited line no longer contains the retained stem. The two starter selections normalize only their separately recomputed exact-head/closure/base evidence fields to break the repin self-reference; every selection-semantic field remains full-file bound. The validator rejects stale spaced/capitalized product spelling, Hyperfinite-shaped protocol, User-Agent, cryptographic, package, API, publisher, or domain identifiers, unclassified occurrences, and nonconstant package/lock/Project/API/schema/publisher drift. It parses decoded JSON keys and walks exact schema paths, canonicalizes schema URLs, normalizes ASCII Unicode escapes in code, verifies the TypeScript epoch field's brace ancestry, permits the epoch only at the exact compatibility config/schema/type declarations, and rejects a selector elsewhere in the shared packaging schema or any runtime/model/migration surface. validate:packaging independently binds package metadata to the same compatibility object, and exact-head customer-starter closure is regenerated through the clean-commit repin workflow. | Maintainers / schema and release validation | Repository implementation | tests/packaging.test.ts, tests/synthetic-sandbox-canary.test.ts, npm run validate:technical-identity, schema validation, package/lock/registry/Project/protocol checks, semantic-context, selection-normalization, removed/replaced occurrence, unclassified/stale/migrated/escaped/same-file epoch negative tests, exact-head customer-starter repin, and full packaging validation | A future legitimate migration still requires a new reviewed identifier epoch, versioned compatibility and stored-evidence treatment, regenerated artifacts, and human approval; protected-branch or release-tool compromise remains. |
| C38 | T02, T03, T05, T11, T16, T24, T25, T31, T37 | ADR 0020 and the required closed CompatibilityMatrix.productBoundary permit only exact-head authoritative repository clones for maintainers/local evaluators, verified customer-starter profiles for their documented profile commands, and reviewed full file-only copies for a complete customer sandbox. The package remains private, exports metadata only, and has no SDK entry metadata, direct or directories.bin executable, implicit server.js start, implicit binding.gyp install, or install/prepare/dependencies/package/publication/deployment/service lifecycle script; repository npm scripts are context-bound rather than a packaged CLI. Implicit filenames are ASCII-case-folded and dependency installation uses --ignore-scripts. Hosted/deployable service, registry package, TypeScript SDK, bundled administration, and live-effect distribution claims are unsupported. Independent human administration and deployed trust services remain prerequisites. | Schema/type/config parity, tests/packaging.test.ts, and validate:packaging reject omitted, widened, or contradictory boundary values, case-variant npm implicit entries, and lifecycle fallbacks; public docs, source/tool maps, and the customer evaluation guide are required to state the same supported and unsupported entry points and distinguish starter-profile scope from the full-copy matrix. Customer-starter closure and script-free clean-extraction validation prove the copied profiles retain their documented command surface without inventing a registry, binary, native build, service, credential, or effect fallback. Extraction evidence records actual outcome, exit status, signal, and measured duration, including the expected deep-import refusal. | Maintainers / packaging and product boundary | Repository implementation | Closed compatibility-schema negative test, private/exports/no-bin/no-SDK/directories.bin/server.js/binding.gyp/case-variant/lifecycle-script assertions, script-free dependency installation, complete extraction step evidence, required documentation statements and ADR supersession, exact-head customer-starter build/verify and clean-extraction validation | Source remains available for unsupported use. A future SDK, CLI, hosted/deployable service, publication, or production-support commitment requires separate reviewed product, compatibility, release, security, and operations work. |
| C39 | T02, T03, T05, T07, T10, T11, T16, T24, T25, T31, T38 | ADR 0021 and the upstream-only RepositoryMetadataContract fix the immutable numeric/node/full repository identity, exact README-aligned description, explicitly unset homepage, closed eight-topic set, human-admin-only apply mode, unavailable built-in-token apply path, required fresh pre/post reads, and nine false authority grants. The optionless planner accepts only closed stdin readback for the literal repository and has no network client, credential reader, target selector, apply flag, mutation adapter, PAT, App token, or success fallback. | Schema plus independent constant validation reject any widened or substituted contract. The planner rejects wrong identity, unknown fields, malformed/duplicate/oversized topics, and unrecognized permissions; it emits canonical contract/readback digests and complete description/homepage/topic drift. A drifted non-admin read is blocked. After reviewed merge, a human repository administrator confirms the fresh plan, applies through GitHub settings, and repeats the exact direct read; only complete zero drift is acceptance. | Repository maintainers / human repository administrator | Repository contract and read-only planner; live setting remains a human gate | tests/repository-metadata.test.ts, schema closure, accepted README wording, issue-taxonomy identity parity, current-state drift fixture, exact-match readback, arbitrary-repository/unknown-topic/non-admin/duplicate/extra-field refusals, no-mutation source assertion, customer-starter exclusion, and the maintainer metadata checklist | A repository administrator or compromised GitHub account/platform can drift About fields after readback. The observation is point-in-time evidence; display metadata grants no lifecycle, repository, target, Project, capability, credential, transition, release, or effect authority. |
| C40 | T02, T05, T07, T11, T16, T27, T28, T39 | ADR 0022's existing-Project repair uses a separate closed GitHubProjectDisplaySnapshot and content-addressed target manifest. The snapshot accepts exact user or organization ownership and current public/private, open/closed, populated display state, but rejects installation, credential, binding, item, and effect data. Targets are copied only from four fresh read-only snapshots in fixed catalog order, every current merged Project-schema digest is bound, title alone is insufficient identity, and an independent human must confirm the manifest digest before planning. Generic GitHubProjectLive, admin/bootstrap, organization/private/empty/View 1, installation, activation, and runtime-binding requirements are unchanged. | The closed plan can express only set-single-select-option-color with exact Project/field/option node IDs, before/after supported colors, accessible name/description context, and constant human-admin/display-only/non-authoritative markers. Unexpected field/option identity, name, type, order, description, target state, duplicate ID, timestamp, schema digest, manifest digest, action kind, or plan digest fails closed. View layout and API-visible field order are retained as observation but cannot authorize or block color repair. Post-apply readback binds both confirmed digests and a strictly later fresh snapshot, requires exact target/schema/custom-field/option state and zero drift, and records runtimeBindingProduced: false; apply/execute are rejected before file reads and the tool has no network or authentication path. | Repository maintainers / independent confirmer / human Project administrator | Repository display-only contract and offline CLI; live apply remains external | tests/github-project-bootstrap.test.ts; user/org owners, missing-installation separation, exact 158-action synthetic shape, deterministic bytes, target/title/digest/freshness substitution, duplicate/missing/extra/reordered/renamed field and option state, unsupported/tampered color and action, observation-only view order, exact zero-drift readback, changed identity refusal, generic planner rejection, CLI pre-read apply/execute denial, full validation, customer-starter extraction, and exact-head code/security review | A human administrator or compromised GitHub account/platform can apply a different color or drift after readback. External apply authentication and durable evidence storage are not implemented; issue closure remains a separate human gate. |
Customer-copy controls additionally retain a valid source CODEOWNER while
customer:configure rewrites every rule for the customer owner before the
initial commit. validate:customer-readiness allows the source maintainer only
at that exact source-repository CODEOWNERS path or when the CODEOWNER namespace
equals the destination repository owner, and exercises every deny rule with
positive and near-miss tests. Release, starter, provenance, and installation
plans bind canonical server plus owner/repository, preventing same-name
repositories on different GitHub Enterprise Cloud hosts from colliding.
C35's snapshot digest covers the complete readback body except the digest itself, including observation time and handoff-plan digest. A satisfied control must also carry the exact planned evidence owner/source/readback metadata, every required target digest, and satisfied prerequisites. The repository command checks clean HEAD/status before evaluation and again immediately before output; a synthetic apply proof must bind the same handoff-plan digest as its containing report. The repository command creates a plan-bound synthetic-unconfigured snapshot; only a current trusted-adapter readback may satisfy a live control. Customer profiles include only synthetic Project target examples, while the internal-reference scanner rejects non-synthetic Project node identities. Trusted Project bootstrap requires a separately confirmed target-manifest digest; raw customer Project/view IDs remain in protected customer repository-only manifest, and a self-rehashed substitute still fails the fixed digest comparison. The exported simulator uses synthetic Project and owner identities only. The npm package dry-run uses the same private-evidence exclusions; publication remains prohibited.
The nonproduction composition root in src/durable-store-composition.ts
validates the exact DeploymentTopologyPlan before opening storage, opens each
of its four isolated stores once, and supplies every one of the fifteen adapter
ports only the store fixed by ADR 0014. It accepts no ambient path, clock,
credential, signer, provider result, genesis state, compatibility list, or
budget. Caller-owned topology, path, dispatch, and genesis data is
canonical-snapshotted and privately frozen before open, preventing post-open
rebinding. tests/durable-store-composition.test.ts pins the closed mapping,
compares uninterrupted progression/refusal with a close/reopen after every
durable boundary, verifies all four backup manifests after restore, proves that
disabled-state recreation starts only at genesis, and covers corruption,
stale/replayed/revoked evidence, budget exhaustion, ambiguous commits, and
landed lost acknowledgements. Backup destinations are canonicalized and checked
against every live database and sidecar, then created under one private,
inode-pinned mode-0700 root. The trusted guard must return the same signed
writer-disabled generation/checkpoint before and after every copy, current under
an injected trusted clock. Restore reads
the exact chain-verified namespace inventory and rejects extra namespaces,
unsigned or mixed store-manifest sets, or a changed topology under the signed
closed composition backupSetId. The proof does not change C31's nonproduction
delivery or residual-risk classification.
A row moves from planned to implemented only when:
- enforcement code is merged;
- positive and negative tests cover every named path;
- evidence is emitted and replayable;
- failure and recovery behavior is documented;
- security owner reviews the residual risk.
Documentation, workflow completion, or a model verdict alone never changes control status.