MDCP report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The current reports support a bounded account of host-interpreted documentation work, deterministic compilation and conditional mechanical checking, with a separate maintainer feedback plane. Both previous epistemic blockers are resolved. No material unsupported conclusion or unresolved cross-report contradiction remains.

I read all supplied members, correction answers, previous verification and report-check. The structural report-check says none, so there is no structural failure to route. The runtime, memory and workflow answer files are empty; the epistemic file answers both prior blockers as corrected. I judged those answers against the current reports and frozen source rather than treating them as proof. SRC-1's checkout HEAD matches the reviewed commit, its working tree is clean, and its top-level tracked tree agrees with the boundary coverage. No target, test, example, benchmark, model provider or evaluator was executed. Source text supplies implementation or doctrine/design, not observed compliance or causal benefit.

Record support and evidence strength

  • Host policy and feature delivery. RT-CMP-host-model, RT-OBJ-mdcp, RT-RTE-feature-work, RT-CLM-context-discipline and RT-BAP-skill-host agree with SRC-1, skills/mdcp/SKILL.md and skills/mdcp/references/workflows/feature-level.md. I traced task intake, workflow selection, branch creation, audience-specific shards/indexes, implementation, as-built reconciliation, validation and submission. Headless defaults make human approval conditional. The skill requests one shard and one workflow, but provides no enforced cumulative window budget, eviction operation or child-return schema. Its vendor-policy restriction does not turn arbitrary host writes into controlled admissions. The workflow account correctly distinguishes prescribed phase handoffs from implemented workers, and submission from acceptance, merge or deployment.
  • Compilation, persistence and later consumers. RT-OBJ-config, RT-RTE-compile-check, MEM-OBJ-compiled, MEM-OBJ-refs and MEM-OBJ-shard-cache are supported by SRC-1, packages/mdcp-cli/src/cli.ts, packages/mdcp-cli/src/compile-workspace.ts, packages/mdcp-core/src/config/load.ts, packages/mdcp-core/src/config/schema.ts, packages/mdcp-core/src/compile/assemble.ts, packages/mdcp-core/src/compile/output-documents.ts, packages/mdcp-core/src/compile/section-manifest.ts, packages/mdcp-core/src/compile/shard-cache.ts, packages/mdcp-core/src/compile/write-output.ts and packages/mdcp-core/src/refs/registry.ts. I traced manifest/fallback/transitive selection, separate output-location assembly, cache lifetime, output writes, refs generation/comparison and registry reload. Orphans reject before compilation; later checks can fail after outputs and refs are written. Standalone refs-check computes comparison data without regenerating the saved registry. Backup rename is not transactional rollback. SRC-1, packages/mdcp-site/scripts/sync-content.mjs independently establishes the shipped downstream consumer: registry/compiled source tags locate ownership, source files supply page bodies, and generation writes pages/navigation. This is wired deterministic reuse, not observed agent recall.
  • Alternate write paths. RT-RTE-hook-extension, RT-RTE-shard-import, MEM-RTE-peer-fix and EPI-RTE-peer-fix agree with SRC-1, packages/mdcp-core/src/index.ts, packages/mdcp-core/src/compile/hooks.ts, packages/mdcp-core/src/config/resolve-compile-hooks.ts, packages/mdcp-core/src/shard/orchestrator.ts and CLI dispatch. Public assembly/writer APIs are separate from the CLI gate. Callbacks can replace bodies, arrays replace default hooks, unresolved hook names are skipped, import removes destinations before rebuilding, and fix delegates mutations without consuming returned peer statuses. These are supported alternate routes, not a claim that malformed inputs defeat a nonexistent sandbox. None establishes per-content semantic approval.
  • Authorship, history and semantic maintenance. MEM-OBJ-shards, MEM-OBJ-research, MEM-RTE-shard-authoring and RT-RTE-doc-review match SRC-1, skills/mdcp/references/workflows/getting-started.md, skills/mdcp/references/workflows/doc-only.md, skills/mdcp/references/workflows/design-architecture.md, skills/mdcp/references/workflows/doc-review.md, skills/mdcp/references/review/lenses.md, skills/mdcp-arch-research/SKILL.md and packages/mdcp-core/src/validate/review.ts. Brownfield preservation, current-text withdrawal, superseded ADRs and immutable research observations apply to different material. Review code produces normalized mechanical proxies; instructions request code spot-checks, action rationales, a leave option and recurring-rule tuning. The semantic workflow's afforded status coexists with implemented diagnostics. The memory annotations correctly distinguish authoring from I/O, compiler selection from task selection, and rationale accessibility from actual rationale consultation.
  • Narrow epistemic operations. EPI-OBJ-evidence-links, EPI-OBJ-insert-body, EPI-OBJ-path-claims, EPI-RTE-code-evidence, EPI-RTE-inline-inserts and EPI-RTE-path-probe agree with SRC-1, packages/mdcp-core/src/compile/hooks/code-evidence.ts, packages/mdcp-core/src/compile/hooks/inline-inserts.ts and packages/mdcp-core/src/validate/path-probe.ts, together with their CLI consumer. The comment-before-declaration example is a valid conditional lexical deduction, not an execution report. Existing line fragments follow a preserving branch; they are not validated. Insert expansion copies content without accepting its truth. Path checks evaluate existence under a recognizer, roots and author exemptions; default-off and warn/error disposition constrain their force. EPI-CLM-structure-truth and EPI-CLM-research-warrant therefore neither erase these narrow checkable propositions nor promote them to general prose-truth acceptance.
  • Maintainer feedback and release. RT-CMP-claude, RT-RTE-skill-evaluation, EPI-OBJ-skill-description and RT-RTE-release-admission are supported by SRC-1, .agents/skills/skill-creator/scripts/run_eval.py, .agents/skills/skill-creator/scripts/run_loop.py, .agents/skills/skill-creator/scripts/improve_description.py, the result-application instruction in .agents/skills/skill-creator/SKILL.md, docs/developer/live-skill-evals.md, tests/skills/mdcp/evals/README.md, .github/workflows/ci.yml, .github/workflows/release.yml, scripts/release-build.mjs and package.json. I followed expected trigger labels through subprocess detection, scoring, proposal history, candidate reuse, stopping and best-candidate selection. Blinded proposal history does not make reused test-score selection an untouched final evaluation. Returned descriptions need separate frontmatter installation. The scripts do not pin provider weights or establish production self-modification. CI, release permissions and remote-only installation in .claude/hooks/session-start.sh establish local wiring, not deployed protection, successful release or an isolation envelope. The reports appropriately keep learning, criticism, reflection and exercised production self-improvement independent.

All declarations, source anchors, evidence layers and statuses were assessed in these groups and the correction checks below. No evidenced-absence records are declared. Source quotation occurrence and deterministic cross-member agreement remain code checks, not semantic proof.

Corrections, containment and reconciliation

The first epistemic answer is supported. The current inventory and dependent feature transformation/retention and peer-fix rows use MEM-OBJ-shards for project content, not RT-OBJ-mdcp. EPI-OBJ-document-body distinguishes transformed compile strings from selecting configuration; candidate descriptions are distinct from installed policy. Genuine configuration/policy roles remain intact. No dependent referent substitution from the prior blocker remains.

The second answer is also supported. I directly inspected SRC-1, docs/features/protocol/benefit-claims-and-evidence.md, docs/features/protocol/context-size-dogfood.csv and scripts/bench-context-size.mjs. EPI-CLM-context-size, EPI-OBJ-context-size and EPI-RTE-context-size now retain the conditional public claim, reported dataset and measurement implementation. The CSV gives 864 median feature-shard estimated tokens versus 61941 monolith tokens, with rounded ratios 1.4% tokens and 1.5% characters. The script counts decoded string length/tokenizer output, selects the stated corpus and computes ratios. The comparison table and bounded conclusion no longer assert absence of a size comparison. They correctly withhold observed adherence, complete prompt savings, recall and causal benefit; the dated CSV does not establish a benchmark run on the frozen commit.

I checked all eight epistemic Part of: relations: five object parts within compile/check or skill evaluation, two concrete hook routes within the general callback route, and path probing within compile/check. Cross-kind containment is explained and supported. These parts do not invalidate their parent containers or require split supersession.

The reconciliation correctly identifies the two peer-fix declarations as the same CLI route and supersedes the epistemic ID by the memory ID without transferring or discarding specialist findings. Its other distinct-identity decisions preserve policy, mutable shards, configuration, transient body state, persisted derivatives, metadata, cache, research artifacts and claims. It accounts for all memory integration issues. Its removal of the former object-boundary disagreement is justified by the current declarations and ledger, not merely by the correction answer. No current unresolved conflict requires an additional blocker. The reconciler's statement about answers not supplied describes its own input basis; it does not negate the answers supplied to this verification.

Consequential omission questions and stopping

The frozen tree, public exports and producer/consumer dependencies guided three selected questions beyond accepting listed anchors:

  1. Could alternate supported paths defeat the stated controls? CLI dispatch, config schema, compile workspace, peer resolver and public core exports establish reachable disabled-link, warn, skip-Vale, optional-peer and non-strict-coverage alternatives, plus direct API, import and fix paths. SRC-1, packages/mdcp-core/src/peers/resolve.ts and the CLI/config/core paths named above support the reports' local guarantees. I checked both guarded branches and alternate consumers; no universal approval or correctness guarantee survives in the submitted wording.
  2. Could other shipped instructions or consumers change the account of retention and reliance? Besides bootstrap, review, research, installation and site consumers, I inspected SRC-1, skills/mdcp-arch-oss-library/SKILL.md, skills/mdcp-arch-product-docs-site/SKILL.md and skills/mdcp-arch-gtm/SKILL.md. They add audience/layout and downstream-consumer guidance, not an implemented inference/window manager or automatic semantic acceptance gate. The reports explicitly leave their detailed artifact inventories, UX/tutorial phase behavior, evaluator-helper products, exact backup/site safety and host context/permission behavior incomplete. No system-complete negative or identical-workflow claim depends on those uninspected details. Positive protocol/toolchain and selected consumer findings remain supported; exhaustive route and effectiveness comparisons remain unavailable.
  3. Could context-discipline claims carry omitted evidence? The benefit-policy shard, CSV and measurement script now have records and a bounded assessment as checked above. The previous unsupported negative is withdrawn without weakening the supported distinction between static size evidence and behavioral/causal evidence.

Further optional expansion stopped because these questions have supported dispositions, and no identified material challenge remains without disposition. More available files alone do not require an exhaustive audit. Remaining uncertainty is already scoped beside the affected findings; it does not justify false absence, observed-operation claims or broader comparison values. No additional publication limit is needed to preserve this bounded account.

Blockers

none

Limits

none