67%

§11 frontier exploration

Goal

Frontier research (non-blocking): prototype the mission-fitness LLM fitness-function and reflexive self-governance constraints over the artifact fabric; MRVA-style fleet variant analysis; deliverable is decisions/ records + follow-on proposals.

Implementation Sketch

  • Mission-fitness prototype: missions.md §anchors ingested as nodes; graph assembles the anchor+ownership context; LLM judges mission fit for a changed surface; measured against known DESIGN:mission-misalignment precedents.
  • Self-governance: 3-5 STRUCT rules (operational-rule-leak, exit-reason parity, api-surface-leak) as standing concept/constraint query pairs over the fabric domain with LLM confirmation.
  • Variant analysis: one confirmed seam query run fleet-wide across the 12 repos; ledger DO-NOT-RE-TRY shapes evaluated as standing variant queries.
  • Output: decisions/ records; winners become follow-on proposals — never silently expanded scope here.

Spec References

  • .claude/rules/missions.md (anchors + conflict rules); prior-art record (jQAssistant concept/constraint, CodeQL MRVA, evolutionary-architecture fitness functions); .claude/rules/learning-ledger.md §6 (dead-end shapes)

Work Items

  • Prototype the mission-fitness LLM fitness-function (holistic+dynamic class): the mission-anchor+ownership substrate in the graph is currently incomplete (6/8 ## § top-level anchors captured, 0/19 #### ori_*/oric per-crate sub-anchors, and zero mission-to-crate ownership edges — MATCH (rs:RuleSection {file:’.claude/rules/missions.md’})-[r]->() returns 0 rows). Treat this as substrate-extension research: assemble the anchor+ownership context directly from .claude/rules/missions.md for the prototype’s LLM fit-judgment (do not block on a graph rebuild), calibrate the judge against DESIGN:mission-misalignment precedents sourced from .claude/rules/finding-categories.md + plan bodies + git history (no in-graph ground truth exists for this label), and capture the mission-model completion this prototype needs (anchor-parse fix incl. tier-4 sub-anchors + a mission-to-crate ownership edge registered through graph_contract_data.build_contracts) as a decisions/ record and follow-on proposal, never inline scope growth.
  • Prototype reflexive self-governance constraints (jQAssistant concept/constraint split) over the artifact fabric: codify 3-5 STRUCT rules as standing constraint queries with the LLM verification gate
  • MRVA-style variant analysis feasibility over the now-13-repo fleet corpus (the 12 named repos plus ori-platform, landed by §07’s platform-domain ingestion): confirm the fleet substrate is the per-repo :Symbol sets (a repo-parameterized seam query fanned out across all 13 and unioned) — NOT SIBLING_DIVERGENCE, which is 3012 edges confined to intra-ori and does not span the fleet; run one confirmed seam query fleet-wide; source ledger DO-NOT-RE-TRY dead-end shapes from plan.json:ledger[] / .claude/rules/learning-ledger.md §6 (a corpus read, hand-authored as cypher — :DesignDecision nodes are reference-repo issue decisions, not the wrapper ledger, and are never a graph-node source) as standing variant queries, never a repo-name or dead-end catalog hard-coded into a generic runtime; record findings.
  • Emit decisions/ records + follow-on proposal drafts for the winners; explicitly non-blocking for plan close (research deliverable is the record, not shipped tooling)

Fresh intel (regenerated)

This is NOT a build section — it is a research section whose SUBSTRATE has ONE decisive property: the artifact-fabric that §07 (Python estate) and §08 (artifact fabric) built has LANDED and is fresh (platform/artifact-fabric facets stamped 2026-07-15), so all three prototypes have live terrain to query — but every one of the three has a specific substrate HOLE the graph exposes here that the prototype must fill from OUTSIDE the graph.

The three holes, each a decision the package forces before any prototype is written:

  1. Mission-fitness — the mission model is materially incomplete. missions.md has 8 ## § anchors in source; only 6 are :RuleSection nodes (§Compiler and §AIMS are DROPPED), and all 19 per-crate #### ori_* ownership-bearing sub-anchors (§ori_lexer … §oric, §Tooling — this plan’s OWN parent anchor) are ABSENT. Worse, the 6 anchors that DO exist carry ZERO outgoing edges — there is no mission→crate owns_crates linkage. The prototype’s prescribed “graph assembles the anchor+ownership context” cannot run on the graph as-is. [JOIN] mission-anchor-parse-gap + zero-ownership-edge → the mission-fitness substrate is a two-part build (complete the anchor set incl. #### sub-anchors; add the mission→crate ownership edge) BEFORE any LLM fit-judgment, and this section is research-only so that build routes to a decisions/ record + follow-on proposal, never inline.

  2. Self-governance — the fabric edges are live but DESIGN:mission-misalignment has NO in-graph ground truth. The concept/constraint substrate exists (INCLUDES 298, CITES 453, DISPATCHES 182, EMITS/HANDLES_EXIT_REASON 6/15) and the 23 scripts/skill_audit/*.py filesystem scanners ARE the in-house concept/constraint prior art the STRUCT constraint-queries mirror. But the mission-fitness measurement target (DESIGN:mission-misalignment precedents) is not a node anywhere — precedents must come from finding-categories.md + plan bodies + git history, not a graph query.

  3. MRVA — the cross-repo fleet substrate is per-repo Symbol sets, NOT divergence edges, and the ledger dead-ends are not in the graph at all. SIBLING_DIVERGENCE is 3012 edges ALL intra-ori (zero cross-repo); the “fleet-wide across 12 repos” seam query runs over the 13 per-repo :Symbol sets (ori 36.6K, ori-platform 34.9K, rust 34.4K, … lean4 1.7K, elm 1.7K), never over SIBLING_DIVERGENCE. The “ledger DO-NOT-RE-TRY shapes as standing variant queries” have no graph source — dead-ends live in plan.json:ledger[] / learning-ledger.md §6, read from the corpus and hand-authored as cypher.

Cross-cutting verdict: all three prototypes are substrate-extension research, not consumption of a ready substrate — the graph is faithful about exactly where each capability’s inputs are missing. The section’s completed deliverable remains the decisions/ record; normal bug and tooling routing applies to any remaining valid work.

["EXIT_REASON_ROUTING"]

DIG DEEPER

scripts/intel-query.sh status
scripts/intel-query.sh plan-status intel-graph-deepening
scripts/intel-query.sh dag-ascii intel-graph-deepening

(full dossier: frontier-exploration—s-a07efcd2.intel.md)