@@ -64,16 +64,31 @@ heuristic), not more infrastructure.
6464 bridge lands (ADR-006, tracked in the `affinescript` repo).
6565* *Proofs* — PROOF-NEEDS.md governs the "lattice" claim, not the dogfood loop.
6666
67- == Path to production-ready (proposed order)
68-
69- 1. *Dogfood loop* (this document) — DONE: persistence + budgeted query.
70- 2. *Measurement* — the experiment above; publish numbers in this file.
71- 3. *Ingestion quality* — tree-sitter (or per-language) definition extraction;
72- `calls`/`imports` edges, not just containment + co-change.
73- 4. *Freshness* — record the HEAD commit in the lattice file; `query` warns
74- when the lattice is stale relative to the repo.
75- 5. *Agent surface* — a Claude Code skill (estate-wide) that fronts `reticulate
76- query`; optionally the REST API for non-CLI consumers.
67+ == Path to production-ready (owner-ratified 2026-07-07)
68+
69+ Order and shape reflect the owner's decisions D1–D4 (recorded in agent memory
70+ `project_git_reticulator_dogfood_loop_2026_07_07`).
71+
72+ 1. *Dogfood loop* (this document) — DONE: persistence + budgeted query (PR #62).
73+ 2. *Measurement (D2: run it first)* — the A/B experiment above, run against the
74+ current heuristic loop to get a baseline, then re-run after step 3. Publish
75+ the input-token deltas + files-found accuracy in this file. This is the gate:
76+ it proves (or kills) the premise before deeper investment.
77+ 3. *Ingestion quality (D3: tree-sitter now)* — replace the line-prefix extractor
78+ with tree-sitter per-language symbol extraction; add `calls`/`imports` edges,
79+ not just containment + co-change. This is the biggest single lever on pack
80+ quality and its own focused PR.
81+ 4. *Freshness (D4: warn, still answer)* — record the HEAD commit in the lattice
82+ file; `query` prints a staleness warning when the lattice was built at a
83+ different HEAD, but still returns the pack. Small, own PR (extends the
84+ FileStore envelope + a query-time check).
85+ 5. *Agent surface (D1: MCP server)* — expose `query` as an MCP tool
86+ (`reticulate_query(topic, level, budget) -> context pack JSON`) so any
87+ MCP-capable client can consume it, not just Claude Code. Own PR.
77886. *CRG B* — lint/fmt/doc-coverage targets per READINESS.md.
78- 7. *The neuro-symbolic stack* — embeddings, VeriSimDB, proof-carrying
79- retrieval — only after steps 1–5 prove the symbolic half pays for itself.
89+ 7. *The neuro-symbolic stack* — embeddings, VeriSimDB, proof-carrying retrieval
90+ — only after steps 2–5 prove the symbolic half pays for itself.
91+
92+ Each of steps 3–5 is a separate reviewable PR; do not bundle them. Step 2
93+ (measurement) precedes 3 so the tree-sitter investment is made against a known
94+ baseline, not on faith.
0 commit comments