Cigito v3 — Distill Data Curator (W2 milestone)

Standalone parallel-track POC. Not wired into Concinno main / Sancio.

Purpose

W2 milestone of the Cigito v3 plan: curate 5 000 (instruction, response, ZIQ trajectory, CBUA stage, FieldRead compression %) pairs from existing Concinno trajectories — at $0 CPU, pure stdlib heuristic, no GPU/API call.

The curated jsonl is the seed corpus for the W3 distill loop (DeepSeek V4 student + Cogito v2.1 671B teacher, IDA composite loss, USPTO patent skeleton — see W1 design doc).

Status

This W2 milestone is data curation only — it does not run the distill loop, does not start RunPod, does not invoke any LLM API. The output data/pairs_v0.jsonl is the input to W3.

PyPI publication note (0.0.1 first publish)

The 0.0.1 release intentionally has zero callers in the Concinno ecosystem. Wiring is gated on the W4 GO/NO-GO Gate per the Plan v3 parallel-track spec, not on this package's PyPI presence.

If you are a third-party adopter:

• Treat 0.0.x as Pre-Alpha — the Development Status :: 2 - Pre-Alpha classifier in pyproject.toml reflects exactly that. Public APIs (curator schema, output jsonl shape, classes under cigito_v3.distill) may change incompatibly between 0.0.x releases as Phase 1 POC findings come back from RunPod runs.
• The package is standalone and pure-stdlib by design — it does not import concinno, concinno-skills-*, or sancio-runtime. Installing it next to those packages does not affect their behaviour.
• A production wire-in (e.g. a concinno cigito-distill subcommand or a Concinno main pkg Distill skill) will land only after the W4 GO/NO-GO Gate evaluates the Phase 1 POC. Until then this package is a staged research artefact, not a turn-key feature.

Layout

cigito-v3/
├── pyproject.toml          (placeholder, version 0.0.1, AGPL)
├── src/cigito_v3/
│   ├── distill/
│   │   ├── curator.py        — 5k pair selector (heuristic, $0 CPU)
│   │   ├── trajectory.py     — ZIQ outcome bus replay → decision sequence
│   │   ├── stage_tagger.py   — CBUA stage transition labeler
│   │   └── compression.py    — FieldRead pre/post compression demo
│   └── data/
│       └── pairs_v0.jsonl    — curated output (or pointer)
└── tests/
    └── test_distill_v2.py    — ≥12 tests

CLI

python -m cigito_v3.distill.curator \
    --concinno-home ~/.concinno \
    --handoff-root _AI_BRAIN/06_Handoffs \
    --target-pairs 5000 \
    --output data/pairs_v0.jsonl

The curator is read-only — it never writes outside the explicit --output path. It refuses to call any external service (asserted in test_curator_no_external_api_call).

Output schema

Each line of pairs_v0.jsonl is a JSON object:

{
  "instruction": "<task framing extracted from the source session>",
  "response": "<canonical response derived from the trajectory>",
  "ziq_trajectory": [
    {"step": 0, "tunable": "...", "value": ..., "reward": 0.93,
     "sps_x_ftrl": 0.71, "source": "..."}
  ],
  "cbua_stage": ["C0", "C1", "C2", "B1", "U1", "A1", "A3"],
  "fieldread_compression_pct": 0.42,
  "source_session": "<session id, anonymized>"
}

Hard constraints

1. $0 CPU — pure stdlib + heuristic. No embedding model, no LLM call.
2. No personal paths — every input root is parameterized via CLI / env.
3. No GPU API call — test_curator_no_external_api_call asserts the relevant env vars are unset during curation.
4. Idempotent — same inputs deterministically produce the same output.
5. AGPL — per W1 design doc, parallel-track Cigito tracks AGPL like Concinno mainline.

Not a GO/NO-GO Gate

W2 ships a curated pairs jsonl. It is not a Phase 0 PASS milestone. Per the post-exec sediment warning, "Phase 0 PASS" framing is misleading — this milestone is one input to the W3 distill loop, the GO/NO-GO Gate lives at the W4 evaluation stage with held-out test pairs and capability deltas vs the DeepSeek V4 base.

The string Phase 0 PASS is deliberately absent from outputs (regression test: test_no_phase0_pass_string_in_outputs).

License

AGPL-3.0-only (per W1 design doc — Cigito follows Concinno mainline AGPL).