Epistemic Compiler — Research Protocol & Mission 01

V0 is a protocol-defined research IR, not a machine-enforced typed graph or an autonomous research factory. It combines provenance conventions, four human gates, manually executed work packages, and a real-time friction-log schema. Mathematical formalisms (sheaves, presheaves, descent, assume-guarantee contracts, relational joins) are candidates to be justified, not premises.

Current State

• Mission 01 executed WP-01..WP-07 and the pre-registered falsification suite against StrangeTcy/rl_eval_generator.
• Pre-registration lock: cbd4e4a081d0b5b1af8a062b6cdf5c4a4ac7d4e7.
• Initial combined target-repository test run: 17 passed (historical); a later repeat could not collect two harness modules because the resumed sandbox lacked torch, so no fresh target-suite pass is claimed. The 17 tests verify listed engineering behavior, not mathematical gluing proofs.
• Post-Gate-4 Mindcluster importer/retriever checks: 9 passed (tests/test_import_mindcluster_html.py + tests/test_retrieve_context.py), including graph integrity, provenance, uncertainty preservation, and candidate-link rendering.
• Mission 01 has human approval through Gate 4 as a bounded diagnostic. The user accepted accept_bounded_record at Gate 3 and approved C01–C03 at Gate 4; the broad local-to-global / agent-coordination question is classified inconclusive_result. See mission-01/gates/ and mission-01/claim_set.md.
• Preserve the locked spec and raw results. Do not infer a general sheaf theorem, cohomology class, real-agent contract benefit, or Astra-vs-swarm advantage from this run. Part B is a deterministic enumerated compatibility stress test, not a comparison of independent LLM agents.
• The Gate 2 decision-table update outputs are quarantined as exploratory, not calibrated Bayesian posteriors. Mission 01 coordination-efficiency metrics are not estimable from the incomplete real-time friction log.

Architecture Boundary & Pilot Lessons (Prospective)

The intended product boundary is a research-control layer above a replaceable agent runtime—not a hand-built swarm runtime. V0 provides the research-memory schema/retriever, mission design conventions, provenance, human gates, and validation/claim controls; it does not provide persistent agents, dispatch, a DAG runtime, automatic delegation, or a research-factory framework. Do not expand runtime machinery without a separately justified need.

A draft, first-class Strategy IR proposal now separates procedural problem transformations from declarative Mindcluster memory and from runtime execution. Its trigger-first taxonomy is explicitly unvalidated; there is no automatic strategy selector, application engine, or promotion mechanism yet. See research_protocol/strategy_ir.md and research_protocol/strategy_families.yaml.

Mission 01's manual execution is retained as a pilot, but its lessons are kept separate: (1) the human had to perform runtime-like context/coordination work, which is a qualitative design observation only because the friction log is incomplete and total burden/κ is not estimable; and (2) the initial mission framing had a conceptual/methodological defect, documented in mission-01/reviews/REV-01_methodological_review.md. A future registration may begin with local validity → declared boundary compatibility → global validity, treating sheaf/presheaf/descent formalisms as candidate models rather than premises. This is a prospective design direction, not a retroactive result or revision of Mission 01's frozen registration.

Repository Layout

.
├── README.md
├── research_protocol/
│   ├── protocol.md             # Ontological separation, provenance, Gates 1–4, canonical boundary contracts
│   ├── instrumentation.md      # Append-only friction schema, controlled vocabulary, coordination efficiency
│   ├── strategy_ir.md          # Draft problem-transformation IR; human reviewed, not implemented/validated
│   └── strategy_families.yaml  # Twelve proposed trigger families; no promoted strategy cards yet
├── knowledge/                  # Imported 118-node Mindcluster graph with source/edge provenance caveats
├── scripts/import_mindcluster_html.py # Safe literal-JSON importer for the uploaded HTML export
├── scripts/retrieve_context.py # Deterministic lexical + graph-hop context compiler (no vector DB)
├── tests/test_import_mindcluster_html.py
├── tests/test_retrieve_context.py
└── mission-01/
    ├── seed.yaml               # Historical pre-registration seed; inaccurate statements are documented in REV-01
    ├── council/                # Four completed role responses and original prompts
    ├── critiques/              # Cross-critique and preserved disagreements
    ├── spec/                   # Draft and Gate-2-frozen approved spec
    ├── work_packages/          # WP-01..WP-07; overlaps.yaml is retrospective only
    ├── results/                # Generated audit, benchmark, and summary artifacts
    ├── falsification/          # WP-07 falsification output
    ├── reviews/                # Post-run review, migration snapshots, and analysis errata
    └── gates/                  # Human Gate packets; Gate 4 approves the bounded claim set

Mindcluster Retrieval (Prospective)

The user's knowledge graph is designed as source records, claim-bearing nodes, and typed edges. Source-backed claims remain distinct from interpretive relations. scripts/retrieve_context.py compiles one shared evidence view plus role-specific packs and a provenance manifest using deterministic lexical matches and explicit graph hops. No embeddings, external searches, orchestration runtime, or automated truth validator are included.

The user-uploaded epistemic_research_graph(2).html was imported after Mission 01: knowledge/ now contains 118 nodes, 301 directed edges, and 77 source records (including the export itself). The importer records the HTML hash and Git provenance without executing scripts; its external links and annotations remain unverified, and uncertain edges are preserved. Mission 01's Council did not receive compiled context packs; do not describe it as mindcluster-informed. See knowledge/README.md for import scope, caveats, and usage.

Target Repository

• StrangeTcy/rl_eval_generator
• Mission branch: mission-01/sheaf-gluing-protocol
• Gate decisions and approved bounded claims: mission-01/gates/gate3_execution.md, mission-01/gates/gate4_claim.md, and mission-01/claim_set.md.
• The uploaded HTML arrived on remote main in commit 2e3847a; that commit was merged locally before importing it, preserving the user's upload in history.