trait-inference — Paper 19 (ATI) Phase 3 Pipeline

Python package implementing the frozen Paper 19 (Automated Trait Inference) Phase 3 training pipeline + baselines + eval harness, per:

• Spec: ai-ecosystem/research/paper-19-trait-inference/phase-1-spec.md
• Pre-registration: ai-ecosystem/research/paper-19-trait-inference/preregistration.md
• Brain: ai-ecosystem/compositions/trait-inference-brain.hsplus
• GPU-claim ticket: task_1777072040695_mrr3

Status (2026-04-24): Phase 1 (CPU pipeline) shipped — dataset loader/audit/splits + 3 baselines (keyword + TF-IDF + Brittney-stub) + eval metrics with bootstrap CI + CLI runner + Vast.ai launcher. Phase 2 (model module) — sentence-transformer encoder + constrained- decoder LLM, requires [model] extra — pending follow-up commit.

---

Quick start

1. Install (CPU baselines + eval only)

cd packages/trait-inference
pip install -e .

2. Smoke test (synthetic data, end-to-end)

Validates the pipeline runs without needing real data or GPU. ~2 min.

trait-inference smoke --n 200 --bootstrap-b 200

Should emit a JSON measurement bundle to stdout with "smoke_test": true, "passed": true. Use this to validate a fresh install before committing to a Vast.ai run.

3. Extract trait label space from HoloScript core

trait-inference extract-traits \
  --constants-dir ../core/src/traits/constants/ \
  --output trait_inference/data/trait_label_space.json \
  --verbose

Reads the 113 TS constant files, extracts string-array exports, writes a single JSON consumed by the dataset + model modules.

4. Audit a real dataset

trait-inference dataset audit data/atimark.jsonl --output measurements/audit.json

Returns exit 0 if the dataset passes spec §1.4 acceptance (≥2k pairs, ≥300 novel combinations, ≥500 each major source, ≥200 negatives, no novelty leak); exit 1 with issues list otherwise.

5. Run baselines

trait-inference dataset split data/atimark.jsonl --output-dir splits/ --seed 42
trait-inference baseline run keyword --train splits/train.jsonl --eval splits/held_out_novel.jsonl --output measurements/keyword.json
trait-inference baseline run tfidf   --train splits/train.jsonl --eval splits/held_out_novel.jsonl --val splits/val.jsonl --tune-threshold --output measurements/tfidf.json
trait-inference baseline run brittney --train splits/train.jsonl --eval splits/held_out_novel.jsonl --output measurements/brittney.json

Each emits f1_macro, exact_match, bootstrap_ci, sample predictions.

---

Vast.ai GPU launch

Orchestration script: scripts/vast-launch-paper-19.ps1 (PowerShell; mirrors the existing ai-ecosystem/scripts/vast-bench-runner.ps1 pattern).

# Cheapest end-to-end pipeline validation (~$0.30, ~5 min)
.\scripts\vast-launch-paper-19.ps1 -Phase smoke -Label paper19-smoke

# Run all 3 baselines on the real dataset (~$0.30, ~10 min)
.\scripts\vast-launch-paper-19.ps1 -Phase baseline `
  -DatasetPath data/atimark.jsonl -Label paper19-baselines

# Full training run (REQUIRES preregistration.md frozen + Phase 2 model module shipped)
.\scripts\vast-launch-paper-19.ps1 -Phase train -GpuName RTX_4090 `
  -DatasetPath data/atimark.jsonl -Label paper19-headline-cell-1

Pre-flight: requires vastai set api-key configured (see ai-ecosystem/.env VAST_API_KEY); requires ~/.ssh/id_rsa with the matching public key registered on the Vast.ai account; requires ≥$0.50 credit for train.

---

Cost estimate (per

ai-ecosystem/research/paper-19-trait-inference/README.md Phase 2-4 task table + GPU-claim ticket _mrr3)

Job	GPU	Hours	Cost
Smoke test	RTX 4090	0.1	$0.03
Baselines (CPU-bound)	RTX 4090	0.2	$0.06
Single training cell	RTX 4090	~6	~$1.80
Full sweep (30 cells × N=5 reseed = 150 runs)	RTX 4090	~900 (parallel: 30 GPUs × 30hr)	~$240

(A100 estimates are roughly 4-8× higher; A100 supply is also tighter. 4090 is sufficient for ≤1B-param decoder per spec §3.1.)

---

Per-spec deliverable map

Spec section	Module	Status
§1.1 Sourcing 3-source mix	dataset.py Pair + Source	done (loader; data construction is Phase 2 task)
§1.2 Schema	dataset.py Pair dataclass	done
§1.3 Splits (train/val/indist/novel)	dataset.py make_splits	done
§1.4 Audit protocol	dataset.py audit + AuditReport	done
§2.1 Keyword baseline	baselines.py KeywordBaseline	done
§2.2 TF-IDF + LogReg baseline	baselines.py TfidfLogregBaseline	done
§2.3 Brittney few-shot baseline	baselines.py BrittneyFewShotBaseline	stub (real impl needs Brittney API integration)
§3.1 Constrained-decoder model	model/ (Phase 2 commit)	pending
§3.2 Conditioning fields	model/ (Phase 2 commit)	pending
§3.3 Hyperparameter sweep	model/sweep.py (Phase 2 commit)	pending
§4.1 Metric definitions	metrics.py f1_macro, f1_micro, exact_match_rate, bootstrap_ci	done
§4.2 Statistical protocol	metrics.py bootstrap_ci, evaluate_headline	done
§4.3 Ablation matrix	eval/ablations.py (Phase 2 commit)	pending
§4.4 Required user study	(separate UX-research task)	pending
§4.5 Pre-registration freeze	ai-ecosystem/research/paper-19-trait-inference/preregistration.md	FROZEN (do not edit)

---

Anti-pattern guards (binding — inherited from

compositions/trait-inference-brain.hsplus)

• No train-set evaluation. Headline metric on novel-combination split only.
• No easy-split-only F1. Reports include both indist (sanity) and novel (headline).
• No single-source dataset. Audit rejects datasets <500 from any of {existing, brittney, community}.
• No optional user study. §4.4 is required not optional (per F.031).
• No after-the-fact threshold-shopping. preregistration.md is frozen before any Phase 3 board task is filed.
• No qualitative-only claims. ML venue requires numbers; pipeline emits structured measurements.
• No validity gap as "scoped contribution" — constrained-decoding architecture (Phase 2 module) bakes ≥90% validity into the decoder, not into a post-filter.

---

Known limitations / future work

• Brittney few-shot baseline is a stub returning empty predictions; real impl needs HoloScript MCP integration (separate task).
• Constrained-decoder model module (model/) is the Phase 2 deliverable — not in this commit.
• Training loop + ablation matrix runner pending Phase 2.
• User study (Phase 4 §4.4) is a separate UX-research deliverable.
• The PowerShell Vast.ai launcher targets Windows; a bash equivalent for macOS/Linux is a follow-up.

---

Provenance

• Authored by trait-inference-brain (compositions/trait-inference-brain.hsplus).
• GPU-claim ticket: task_1777072040695_mrr3 (live on team_1775935947314_f0noxi board).
• Capability-build provenance commit: fc294af (lean-theorist-brain — sibling).
• F.031 pre-emptions baked into spec; constrained decoding ships in Phase 2 model module.