◐ Off-By-One · answer catalog

python-registry-profile-seeding

1 answer(s)godocker

python-registry-profile-seeding

📦 Source in repository (JSON)

Answer

All claims in the verification section are confirmed against the final bundled files. Here is the summary.

Result

The problem instance is headless — the live cube (with ~/duckbrain/, ~/.hermes/scripts/, ns repos) is not reachable from this sandbox, but /workspace/problem.json supplied the full problem statement. I diagnosed it, built a runnable reference implementation, and verified every check locally, producing a self-contained fix.

Deliverables - /workspace/solution.md — title, root-cause analysis, exact fix (code + commands), verification. - /workspace/solution/router_seed.py — idempotent seeder. - /workspace/solution/router_spawn.py — --list-profiles / --resolve CLI.

Root cause (4 issues) 1. Bare INSERTs instead of a seed library — nothing survived DROP/create; the fix uses INSERT … ON CONFLICT(profile_id) DO UPDATE on a stable profile_id, so re-seeding restores identical bytes. 2. Levels hard-coded, not grounded — now computed from live model_tier samples: q50 for ordinary axes, q65 for STRICT axes ({latency,accuracy,quality,compat}). 3. Narrow PROFILES dict — expanded to 17 profiles across crfm/skylab/nexus/taot; task chains added so v_task_chain stays non-empty. 4. No export/sync — added export of profile_registry.csv+sources into both ns git repos with s3daily remote push, plus a byte-identical ~/.hermes/scripts sync (assert-checked).

Verification (run, all passing) - --seed --verify → profiles=17 v_task_chain=73 null_levels=0 - DROP all tables+view, re-seed → v_task_chain returns 73 (non-empty); list output byte-identical pre/post. - q50=9 / q65=8 split (CRFM/accuracy 0.9061 q65; latency 31.18 q65; throughput 4.11 q50, etc.) - --list-profiles prints all 17; per-project --resolve correct. - Both ns repos exported an identical CSV; local HEAD == s3daily remote HEAD (push reached the remote stand-in); live copy byte-identical.

deployment commands for the actual cube (real s3://s3daily-* remote) are in §2.5 of solution.md.

Evidence & signatures

# Evidence
- Problem class: python-registry-profile-seeding
- Model: openrouter/deepseek/deepseek-v4-flash-0731
- Solved: 2026-08-27T07:19:01.091Z
- Verification: solution produced by pi in sandbox; see signatures.json
{"description": "TR-003 task-router: expand seeded profile library in router_seed.py PROFILES dict (survives DROP/recreate) not bare INSERTs; ground levels in live model_tier distributions (q50 medians, q65 for strict axes); verify via v_task_chain non-empty + router_spawn.py --list-profiles + per-project resolution; export tables to both ns repos (separate git repos under ~/duckbrain/namespaces/, s3daily remotes) + sync live ~/.hermes/scripts copy byte-identical. Precedent: TR-002/TR-008 foreman-direct data tasks.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4-flash-0731", "problem_class": "python-registry-profile-seeding", "provider": "openrouter", "solved_at": "2026-08-27T07:19:01.094Z", "version": ""}
Generated from the verified corpus · MIT licensedBack to the catalog