I read the report and skimmed the three research documents it summarises. Short answer: I revise the recommendation in scope, and push back on the framing and on about half the gates. Where the report is right - Reducer contract is missing. There is no reducer(state, event) -> (state, effects) signature, no effect representation, no rule for unknown event types or enum values, and no wildcard precedence rule (does an explicit READY row beat Any non-terminal? what if two wildcards match?). These are real and cheap to close. - Run-outcome precedence is unstated. Though see below: with serialised admission and a global sequence, "simultaneous" terminal conditions do not exist at the event level; the missing rule is the evaluation order when several conditions are derivable at once, and the "On quiescence" list already gives an order for that case. It needs one paragraph, not a phase. - Minimum schema-evolution policy is absent. schema_version fields exist without a rule. For the MVP the rule can be strict: replay requires identical schema and policy versions; unknown event type or enum value is an admission failure; no migrations. - No LICENSE file despite SPDX-License-Identifier: AGPL-3.0-or-later headers in the scripts. That is a one-file fix that should not wait. - Harness-side input/output contracts are missing. The user-workflow research is correct that nothing says how a run is created, configured, observed, or its result retrieved. Where I push back 1. "Do not start implementation yet" contradicts the report's own best recommendation. It correctly says (line 116) that the table verifier proves structure, not behaviour, and (line 131) that the right first move is a narrow replayable vertical slice with crash injection. Behavioural properties (replay determinism, crash boundaries, guard implementability) can only be established by executable code. Five review rounds have shown that prose specifications of the reducer drift on every edit; the schemas plus a reducer skeleton with tests will pin down gates 3, 4, 8 and 9 faster and more reliably than another specification pass. Those gates should be deliverables of the next work item, not preconditions for it. 2. Gate 6 and the third-party stack are out of MVP scope by the project's own documents. roadmap.md fixes a single-process controller, a transactional append-only store, no distributed execution, no partitioned ordering. PostgreSQL, Temporal, OPA and Kubernetes are a production stack; a Temporal compatibility spike before the vertical slice exists would be deciding an integration boundary for a component that does not yet have a shape. The alternatives memo itself says its facts are unverified and it is truncated. Record it as informative, defer the decision to after the four-arm experiment, and use an in-process transactional store (SQLite is sufficient) for the slice. The only requirement worth stating now is the one the memo gets right: whatever dispatch substrate is eventually used must never hold normative state. 3. Gate 7 is already closed. control-plane.md:93 says mutation admission is serialised; roadmap.md:13 says single-process controller. It needs one explicit sentence ("the MVP has exactly one admission thread; leases and fencing are out of scope"), not a gate. 4. Gate 2 is disproportionate. Acceptance criteria, dependencies, tests, artefacts and exclusions for all 32 steps, plus a requirement-ID scheme and a traceability matrix, is heavyweight process for a project of this size and would add a third identifier namespace next to headings and event names. The conformance-gate list in experiment.md already enumerates the requirements; the right traceability is one test name per bullet. Write acceptance criteria for the next milestone only, and add the rest as milestones are reached. 5. Gate 1 (full user-workflow document) is premature. For the first experiment the user is the harness, by design: no human in the loop. What is needed now is the harness contract: run configuration (kernel, root contract, budgets, manifest reference), progress/audit output, and result and evidence retrieval. Those are schemas, and they belong in the schema task. The "existing project", intervention, and CLI workflows should wait until there is something to intervene in. 6. Gate 11's sandbox tests are sequenced wrongly. They must precede the first untrusted executor (the model adapter, roadmap step 14), not step 1. The vertical slice runs a trusted deterministic command executor. 7. Snapshots (gate 5) are a replay optimisation. The MVP replays from genesis; say so in non-goals and move on. Revised recommendation Keep the schema task as agreed, and fold the genuine gaps into it so they are closed by artefacts that the checker and tests enforce rather than by prose alone: 1. Specification closure, small and targeted (one commit, mostly control-plane.md and roadmap.md): reducer signature and effect list; unknown event/enum handling; wildcard precedence, with tables.py rejecting definitions where two rows match the same (state, event); the quiescence evaluation order restated as the outcome-resolution rule; strict MVP schema-evolution policy; explicit single-admission-thread statement; non-goals extended with snapshots, multi-worker, external dispatch substrate. Add LICENSE. 2. Schemas as previously scoped, extended with run configuration, project kernel, experiment manifest, progress/audit record, and run result. This is the harness contract and closes the legitimate part of the user-workflow gap. 3. Vertical slice per the report's item 10 and pre-implementation SHOULD 3: in-process store, reducer over the generated tables, one deterministic executor, replay equivalence, crash injection at append and dispatch boundaries. Acceptance criteria written for this milestone only. 4. Then decomposition, concurrency, model adapter (with sandbox tests immediately before it), and the experiment arms. Housekeeping: docs/report-research.md and docs/research/ should be added to the document map as informative, and user-workflow.md:52 should be corrected as the report notes. Complexity impact on the earlier estimate: step 1 adds a small amount; step 2 grows by roughly five schemas and their fixtures; nothing else changes. The Temporal spike, full workflow document, and 32-step traceability matrix should be explicitly deferred rather than left as open gates.