Skip to content

MANIFEST.md

Latest commit

 

History

History
63 lines (52 loc) · 4.9 KB

MANIFEST.md

File metadata and controls

63 lines (52 loc) · 4.9 KB
Raw

Aporia Engine — Repo Manifest (task 20; proposed 2026-07-04, awaiting Aaron's approval)

Every top-level path, classified. Rule 1: nothing enters git except by this manifest, explicitly. git add . is banned.

TRACKED — the system's code

Path What it is
orchestrator.py, orchestrator_test.py, shakedown.py the owned spine + its tests + shakedown runner
ladder.py, ladder_test.py, learn_null.py the acquisition ladder, its tests, the original null loop
new_brain_store.py / _write.py / _read.py / _extract.py / _server.py / _eval.py the fact store: storage, write path, recall, extractor, MCP server, eval harness
new_brain_test_lanes.py / _narrative.py / _read.py / _registry.py / _verify_loop.py fact-store test suite
new_brain_schema.sql, new_brain_seed.sql the fact store's schema + seed
checker.py, runtime_verify.py, verify_server.py, ue_verify_server.py, server.py verifiers + MCP seams
corpus.py, embed.py, cartridge_contract.py, stub_cartridge.py, build_ue_cart.py, warm_ue_cache.py the CART BUILDER pipeline (code only — cart data itself is gitignored)
index_cache.py, prewarm_cache.py, install_hooks.py, freshness_calls_test.py cart freshness spine (2026-07-19): content-keyed parse cache, detached prewarm, bird-clone git hooks, test suite
aporia_scout.py task 35: weekly zero-cloud-token discovery (GitHub/HN/watchlist + local qwen triage); its data outputs are gitignored
longmemeval_bench.py, longmemeval_store_leg.py task 34: LongMemEval retrieval benchmark — substrate harness + in-store pgvector leg (dataset + results in bench_data/, gitignored; DB secret read from host registration at runtime, never stored)
longmemeval_ladder.py task 36: the local ladder — staged hybrid/rerank/aimed-shot climbs over the task-34 baseline, zero cloud
typed_layer_runner.py, experiment-start.bat, experiment-stop.bat, experiment-status.bat task 37: operator-controlled typed-layer experiment — checkpointed nightly ingest, stop-in-seconds, scheduler ships disabled; all data outputs in bench_data/ (gitignored)
typed_layer_score.py task 37 scoring: floor/facts/fused three-column attribution, ready-questions-only gating
LAWS.md, claude_laws.py, claude_gate.py task 38: standing laws injected every turn + the claim gate — Claude's output checked against New Brain via the store seam, contradictions block the turn; gate_log.jsonl (receipts) gitignored
session_log.py task 39: session day-logs (MickyB's scratch-and-regenerate, adapted — transcript as scratch); session_logs/ gitignored
claims.py, extract_claims.py, generate_claims.py claim extraction/generation used by verification
jepa_core.py, fact_grammar_experiment.py, density_experiment.py, ablate.py, confident_wrong.py, plugin_experiment.py, synthetic_sanity.py, ue_experiment.py the research estate: JEPA-era experiments — each carries a DEPRECATED/provenance header (Aaron's condition, 2026-07-04): where Aporia started, clearly marked not-in-use
test_loop.py, test_verify_loop.py, ue_test_verify_loop.py loop tests
migrate_ob_to_documents.py the strangler-fig migration one-shot (credential scrubbed to env 2026-07-04, secrets gate)

TRACKED — docs & specs

Path Note
README.md, BRIEFING.md current contents predate the Aporia naming — task 21 replaces README; task 27 sweeps BRIEFING
return_type_matching_spec.md verifier spec
MANIFEST.md, .gitignore this file + the exclusion law

TRACKED — operational data (the receipts trail, rule 4)

Path Note
dispatch/ (all task records) the queue — task records are receipts
dispatch_log.jsonl outcome log (per-provider cost, acceptance) — verified free of proprietary symbols
roster.json the model roster (data, zero-code-alteration)
l2_allowlist.json the L2 source allowlist

IGNORED (see .gitignore for the per-rule reasons)

.venv/, pycache/ — regenerable environments .embed_cache*, .npy — THE CART DATA (embeddings of UE/RiftSuite source) .index_cache_, .corpus_dirty_*, .test_corpus/ — parse caches / freshness marker / test scratch (parsed UE/RiftSuite symbols — same rule as the embed caches) field_log.jsonl, ue_field_log.jsonl, learned_facts.jsonl, extract_results.json — cart-derived, carry Epic/RiftSuite symbols new_brain_eval_results.json, nb_eval_rerun.json — quote Aaron's personal captures external/ — vendored Apache-2.0 eb_jepa (version pinned in docs instead) *.log — run noise

Verification performed

  • Secrets scan: one hit found (hardcoded DB password in the migration one-shot), scrubbed to env var per Aaron's option-A ruling, re-scan clean.
  • Cart-derived and personal-content files identified by READING contents (extract_results opens with RiftSuite symbols; nb eval files quote real captures), not by filename guessing.
  • Sampled dispatch_log.jsonl: task/canary records with cost fields only — clean.