Sciev v0.2.0 — aligned scientific decision pipeline

Sciev v0.2.0 — aligned scientific decision pipeline

Typed decision heads (choice/noul/score) over frozen LLaDA-8B, trained
and evaluated under the systemone-v2 contract: group-disjoint splits,
explicit overflow handling (max_ctx=960, no silent truncation), shared
R2 preprocessing across train/calibrate/eval/serve, dev-only temperature
fitting, and provenance manifests for every file.

Highlights

• Matched frozen/adapted comparison (identical recipe + seed + data):
  mean±sd over seeds 7331–7333. Domain adaptation robustly improves
  out-of-distribution verification (SciFact noul 0.709 vs 0.483, +22.6pt)
  at a noisy cost to in-distribution rubric scoring — task-dependent, not
  a uniform win/loss.
• Evidence controls (empty/shuffle, agreement-with-reference): expose
  option-side leakage — choice is ~80–87% solvable without reading the
  passage; score is the task whose accuracy most directly reflects
  evidence use.
• Hard-distractor battery (--hard-choice): strict same-passage +
  numeric-perturb pool reduces leakage ~6–8pt but does not remove it.

Contents

• heads/release/ — fr_matched_s7331 heads (recommended; frozen backbone)
• heads/experimental/ — da_matched_s7331 heads (DAPT arm; documented caveats)
• evals/ — 32 evaluation JSONs (arms × tasks × benchmarks × controls)
• manifests/ — dataset manifests incl. exclusion provenance and input hashes
• sciev-v0.2-paper.pdf, MODEL_CARD.md, LICENSE, SHA256SUMS,
  release-manifest.json

Checkpoints are self-describing (inference config, training-data
contract, input fingerprints, recipe, seed). Controls are scored as
agreement-with-reference, never relabelled.