Skip to content

ROADMAP_EXECUTION.md

Latest commit

 

History

History
80 lines (60 loc) · 5.92 KB

ROADMAP_EXECUTION.md

File metadata and controls

80 lines (60 loc) · 5.92 KB
Raw

Execution queue (operator priority)

Effective: 2026-06-06 (Phase II lattice program added)
North-star: VISION_LIQUIFIED_LATTICE.md — self-annealing RAG/DAG with steerable shims and disk-scale pools
Policy: SHIM substrate rows (SHIM-CD-01, 02, 06, 08, 09) are on hold / last during Phase I. Core engine and research-validity work runs one track at a time until each step meets its exit criteria.

Block flag

Merge gate: CLEAR for this sequence (docs/next-session.md). SHIM debts stay OPEN but non-blocking until the queue below finishes.

Sequence (do not parallelize)

Step Track Exit criteria
1 ML correctness — projection DimensionProjection trains (no erroneous .detach() on teacher path); regression test fails if reverted
2 ML correctness — InfoNCE SedimentationInfoNCELoss does not use inter-doc false negatives; tests green
3 Infra hygiene isolated_adapter_state() leak fixed; *.pt backup pattern gitignored; no stray backup files
4 Sweep / CI cost run_large_sweep.py O(N²) JSON path fixed or bounded; CI torch cache documented
5 Packaging Missing py-modules in pyproject.toml added; pip install -e . smoke passes
6 Docs truth CLAUDE.md + CHANGELOG.md reflect current entrypoints and smoke paths
7 Model-Scope integration Shadow steering pilot on Qwen3.5-9B (one policy, provenance persist) per architecture-2026-05-01 phases 4–6 slice
8 E2E learning loop Single test: ingest → sedate → measurable metric delta on fixture corpus
Last SHIM substrate Resume SHIM-CD-01/02/06/08/09 only after step 8; restore Blocking=YES when re-entering shim program

Execution posture

  • Track work one step at a time, but do not leave long checks idle: if any test command in a step runs longer than ~30 seconds, launch at least one independent, lower-cost validation in parallel (state checks, docs check, packaging import smoke, or debt table audits).
  • Keep the operator queue explicit: no step is considered complete until the long-running validation and the parallel companion check both pass.
  • Use this automation to enforce the policy during execution:
    • bash scripts/run_step_with_checks.sh --always --threshold 30 --companion "python -m unittest tests.test_check_block_flag -v" -- python -m unittest -v tests/test_model_scope_runtime.py
    • Replace the primary command for each queue step; add step-specific companions as needed (for example, a docs audit or focused evidence smoke) and keep each check independent of the long-running step.

Automation

  • Phase loop prefers CORE-SLICE-* handlers (low priority number) before SHIM-SLICE-* (priority ≥ 200).
  • Ten-minute loops: bash scripts/loop_core_10m.sh (core track) instead of shim BHS loop unless explicitly requested.

What we are not doing (while on this queue)

  • No new doc-only shim cycles or 10-agent backlog slices.
  • No claiming SHIM-CD CLOSED without full substrate DoD.
  • No Model-Scope + full shim + EGGROLL at full throttle in parallel.
  • No GNN layer or disk-pool integration until Phase II steps 12–14 complete (schema + drift experiment first).

Phase II — Liquified Lattice program

Starts after: Phase I step 8 exit criteria met (and step 7 Model-Scope pilot closed or explicitly carried with TTL).
Vision doc: VISION_LIQUIFIED_LATTICE.md

Same one-track rule: finish each step before starting the next. Phase II does not reopen parallel doc-only shim cycles.

Step Track Exit criteria
9 Model-Scope shadow (close Phase I #7) One bounded steering policy on Qwen3.5-9B fixture; persist_records / load_records round-trip; bridge max_total_interventions exercised in test
10 SHIM substrate DoD SHIM-CD-01/02/06 production seams wired without env-only guards; rollback test; restore Blocking=YES on open SHIM rows when re-entering
11 Annealing controller Single module owns temperature schedule (explore ↔ stabilize); wired to sedimentation + online_updater + ES entrypoint; unit test proves high-T increases perturbation, low-T reduces it
12 Evidence DAG schema Typed graph contract over build_attribution_pool.py output (nodes: query/cluster/actuator; edges: retrieval/intervention links); JSON schema + validator; no GNN required yet
13 Disintegration loop isomer_detector / convergence_monitor triggers prune of low-fitness graph edges or pool entries; re-anneal path records fitness before/after in artifact
14 Concept-drift experiment Injected drift fixture (extend chelatedai-synthetic-collapse or road-course window); measurable recovery within N anneal cycles documented in CHANGELOG.md
15 GNN prototype Lightweight GNN over evidence DAG (PyG or DGL); only after steps 12–14 green; must beat flat-pool baseline on drift fixture or fail closed
16 Quant-aware shim routing adapter_router.py + QuantizationPromotionGate integrated as steering plane; promotion requires quant survival + retrieval fitness
17 Disk pool slice One precomputed pool shard readable via computational_storage_poc/block_graph.py with host parity check; documented in storage track docs

Phase II dependencies (do not skip)

9 (Model-Scope) → 10 (SHIM DoD)
10 → 11 (annealing controller needs stable steering seams)
11 → 13 (disintegration needs anneal schedule)
12 → 15 (GNN needs DAG schema)
13 + 14 → 15 (GNN needs drift benchmark)
16 depends on 10 + 11
17 depends on 12 + block-graph parity (can run parallel to 15–16 only after 12)

Phase II success metrics

See VISION_LIQUIFIED_LATTICE.md#success-metrics-program-level. Promotions require artifact evidence, not prose.