Omnis HelixGenome IDE
← Docs
Helix CLI docs
Browse Helix CLI docs

Governed Predictor Adapters

Omnis Helix should treat external and internal predictors as governed in-silico model calls, not as opaque authority. The product advantage is the audit layer: every model output that influences a design decision must be replayable, attributable, and linked to the evidence bundle.

Product Thesis

Helix is the governed Genome IDE for model-assisted edit review. Predictor quality matters, but the durable moat is the ability to:

  • run multiple model families under explicit assumptions,
  • preserve exact input and output hashes,
  • expose disagreement instead of hiding it,
  • attach model receipts to proof bundles,
  • verify the decision chain offline.

This lets Helix integrate stronger model families over time without turning the platform into a black box.

Adapter Contract

Each predictor adapter must emit a helix.predictor.receipt.v1 receipt with:

  • model identity: model id, version, adapter id, adapter version, digest, source reference,
  • execution contract: local/remote mode, determinism class, network requirement,
  • request envelope: task, candidate id, sequence window, guide, edit context, assumptions, provenance refs,
  • output envelope: normalized scores, labels, uncertainty, warnings,
  • digests: canonical model/request/output hashes,
  • safety scope: in_silico_only.

The receipt is the artifact. UI panels and evidence bundles should consume receipts, not raw model responses.

Near-Term Plan

  1. Native Baseline Ship a deterministic baseline adapter as the contract reference. It is intentionally a ranking proxy, not a frontier model.

  2. Frontier Adapter Shells Add disabled-by-default adapter shells for published predictor families. They should declare required files, model digests, supported tasks, and benchmark status before any score is trusted.

  3. Benchmark Harness Add replayable benchmark bundles that compare model families on held-out public datasets. Store benchmark receipts separately from run receipts.

  4. Decision Cockpit Surface model disagreement as a first-class review object:

    • native baseline score,
    • imported predictor score,
    • mechanistic simulation summary,
    • policy gate state,
    • calibration/benchmark badge.

  5. Evidence Bundle Integration Add predictor_receipts/ to proof bundles. Export must preserve exact receipts and fail verification if any linked receipt changes.

CLI Surface

List available governed predictor adapters:

helix-cli predictor list-models

Generate a native baseline predictor receipt:

helix-cli predictor score \
  --task prime_editability \
  --candidate-id candidate-001 \
  --sequence ACGTTGACCTGACTACGATC \
  --guide ACGTTGACCTGACTACGATC \
  --assumption scope=in_silico_modeling_only \
  --out predictor_receipt.json

Verify a standalone receipt:

helix-cli predictor verify-receipt \
  --receipt predictor_receipt.json

Compare receipts for review:

helix-cli predictor compare-receipts \
  --receipt model_a_receipt.json \
  --receipt model_b_receipt.json

Generate receipts from a JSONL candidate set:

helix-cli predictor score-batch \
  --task prime_editability \
  --input candidates.jsonl \
  --outdir predictor_receipts

Add --bundle path/to/bundle to attach generated receipts to an existing artifact bundle in the same run.

Attach one or more receipts to an artifact bundle:

helix-cli predictor attach-receipts \
  --bundle path/to/bundle \
  --receipt predictor_receipt.json

Attached receipts are written under predictor_receipts/ and added to manifest.json.

Summarize bundled receipts for review surfaces:

helix-cli predictor summarize-bundle \
  --bundle path/to/bundle

Compare bundled receipts only:

helix-cli predictor compare-bundle \
  --bundle path/to/bundle

Use --threshold on compare-receipts or summarize-bundle to tune score deltas that count as disagreement.

UI Direction

The Decision Cockpit should not say “the model is right.” It should say:

  • which models were run,
  • what each model assumed,
  • where they agree,
  • where they disagree,
  • which evidence or policy gates prevent stronger claims.

That is the trust infrastructure angle: transparent model comparison with reproducible receipts.

Boundary Rules

This system remains in-silico software. Adapters may score simulated candidates and compare model outputs. They must not produce real-world execution instructions or unstated claims about observed outcomes.