ML experiments: scaffold 5 neuro-symbolic + inverse-SBI tracks (A100-ready)

[mivertowski]

Summary

Runnable PyTorch scaffolding + per-track specs under experiments/ml/ to close the behavioral-fidelity gap by learning the generator's proposal distributions while leaving the symbolic constraint layer (debits=credits, A=L+E, document chains, IC matching) untouched.

Principle: the NN emits structure/shape; the existing Rust engine projects it onto the feasible manifold. Coherence stays a hard guarantee — no model ever emits a final balance.

This is scaffold only — no model trained yet. Built to run on an A100 when it frees up; the dev box OOMs on orchestrator-scale work, so verification is by syntax-check + design review here, then training on the GPU box.

Tracks

Dir	Architecture	BF metrics targeted
gnn/	GraphSAGE encoder + GAE inner-product decoder	P3 ClusteringGap, TriangleLogRatio
sequence/	causal transformer over JE event-token streams	P1 IETD + Autocorr, P2 JELineBurst, P4 MeanGap
flow/	conditional neural-spline flow (zuko) + round-number atom mixture	Benford / multimodal amount tails
surrogate/	MLP eval-surrogate + CMA-ES knob search	calibration-loop speed (zero coherence risk)

common/ — shared corpus→tensor exporter + BF-eval bridge (canonical Rust scorer + fast Python approximation for in-loop scoring).

Privacy / legal

• Corpus path via DATASYNTH_CORPUS_DIR env — never hard-coded, never logged.
• data/, weights/, runs/, path-bearing configs all gitignored — only code + specs tracked (verified: 0 corpus artifacts staged).
• Per-track memorization review required before any weight leaves the private box (k-anonymity / DP-SGD path documented in gnn/SPEC.md — the GNN has the highest memorization risk).
• No corpus content committed; schema uses DataSynth's own canonical field names, corpus column mapping via a gitignored local YAML.

Test plan

• All 18 Python modules pass py_compile (syntax).
• .gitignore confirmed to exclude *.pt / *.parquet / data/ / weights/.
• Per track, on the A100: data_export → train → bf_bridge.score_canonical lift vs the v5.26 baseline (success criteria in each SPEC.md).

Independent of SP6 (#202) — disjoint paths, branched off main.

🤖 Generated with Claude Code

[mivertowski]

Added track 5 — inverse / simulation-based inference (experiments/ml/inverse/): an amortized neural posterior q(θ|GL) that runs the generator backward to recover the latent process parameters a GL was distilled from, with SBC + credible-interval coverage validated on synthetic (where ground truth is known). Reuses the flow/ NSF + the forward simulator's free labels; many-to-one forward map ⇒ posterior, not point estimate.