Skip to content

SKILL.md

Latest commit

 

History

History
131 lines (109 loc) · 5.35 KB

SKILL.md

File metadata and controls

131 lines (109 loc) · 5.35 KB
Raw

OpenAMP Foundry Agent Skill

Overview

Start from docs/README.md, then use the relevant domain hub. For scientific work, read the current truth source docs/evidence/METRICS_CURRENT.md, then docs/research/ROADMAP.md, then AGENTS.md. Treat computational scores as triage only. Do not describe candidates as active, safe, therapeutic, or validated without qualified lab evidence.

Key Components

  • src/openamp_foundry/pipeline.py: rank candidates and emit evidence.
  • src/openamp_foundry/scoring/: activity, safety, hemolysis, novelty, synthesis, expert, and rich selectivity scorers.
  • src/openamp_foundry/selection/: ensemble/expert ranking, diversity, pilot panel selection, and structural-class floors.
  • src/openamp_foundry/benchmark/ plus scripts/benchmarks/: benchmark honesty checks and regression entrypoints.
  • src/openamp_foundry/calibration/: lab-result intake, gate, and proposal-only recalibration.

Diagrams

Flowchart

flowchart TD
  A["Candidate sequences"] --> B["Validate + feature extraction"]
  B --> C["Independent scorers"]
  C --> D["Ranking + diversity selection"]
  D --> E["Evidence certificates"]
  E --> F["Human review"]
  F --> G["Qualified external assay"]
  G --> H["Calibration intake + gate"]
Loading

Component Diagram

flowchart LR
  Data["data"] --> Pipeline["pipeline"]
  Features["features"] --> Scoring["scoring"]
  Scoring --> Selection["selection"]
  Selection --> Evidence["evidence"]
  Evidence --> Reports["reports"]
  Reports --> Calibration["calibration"]
  Benchmark["benchmark"] --> Scoring
  Benchmark --> Selection
Loading

Sequence Diagram

sequenceDiagram
  participant Agent
  participant CLI
  participant Scoring
  participant Selection
  participant Evidence
  Agent->>CLI: rank / pilot-panel
  CLI->>Scoring: compute raw scores
  CLI->>Selection: rank, diversify, optional class floor
  Selection-->>CLI: selected candidates
  CLI->>Evidence: write certificates
  Evidence-->>Agent: reviewable artifacts
Loading

Start Here

  1. Read AGENTS.md, CLAUDE.md, MISSION.md, and docs/evidence/METRICS_CURRENT.md.
  2. Treat docs/evidence/METRICS_CURRENT.md plus outputs/metrics_snapshot.json as the current benchmark truth when docs disagree.
  3. For the required disconfirming pass, use docs/evidence/DISCONFIRMING_TEST_RECORD_GUIDE.md when recording a challenge.
  4. Preserve the safety boundary: dry-lab scoring and evidence only. No wet-lab protocols, pathogen enablement, toxicity-maximizing objectives, or biological proof claims.

High-Leverage Benchmark Checks

  • make bench-500: broad AMP-vs-decoy discrimination.

Disconfirming-evidence gate

Phase AC records are two-layered: disconfirming_test_record.py stores one attempt to break a claim, while phase_ac_disconfirming_gate.py aggregates those records. A refuted record requires an explicit claim downgrade and an inconclusive record requires investigation. An empty aggregate is not_established, not a pass. Run openamp-foundry phase-ac-disconfirming-gate-check --entry-json ... or make phase-ac-disconfirming-gate-check to exercise the aggregate workflow; the command exits nonzero until the aggregate is verified. Neither artifact is biological validation.

The Phase AA reproducibility gate is also available as a normal review-loop command: openamp-foundry phase-aa-reproducibility-gate-check --entry-json ... or make phase-aa-reproducibility-gate-check. It returns success only when RMC, DCR, CFP, and SBW artifact IDs are all present. This is a structural provenance check, not proof that the underlying run is scientifically correct or biologically valid.

External-result intake is also fail-closed at the review boundary. Use the structured loader/report fields invalid_lab_result_files and input_validation_status to preserve schema-invalid returns; the calibration-intake command exits nonzero and the recalibration gate refuses to proceed while any invalid file is excluded. Missing or non-directory result paths return an input error before a report is written. An existing empty directory is the only valid zero-result state. Duplicate result IDs and duplicate panel candidate IDs are also preserved as input_integrity_issues and block clean intake. These controls catch incomplete or ambiguous input, not assay-quality or biological-validity problems. Control-failed assay observations remain visible for audit but are excluded from per-assay actual predicates, cohort metrics, and interpretable per-candidate outcome flags. Raw outcome fields and failed-result IDs remain available for audit; failed controls still block recalibration.

  • make bench-easy-baseline: trivial length/charge baselines.
  • make bench-charge-matched: adversarial check that removes charge-density separation before comparing ensemble vs charge alone.
  • make bench-per-family: structural-class blind spots.
  • make bench-selectivity: hemolytic-vs-selective AMP ranking.
  • make metrics-snapshot: regenerate the machine-readable source of truth.

Current v0.5.38 Note

The charge-matched decoy benchmark is an honesty check, not a win claim. Correct pH-7.4 charge-density matching remains imperfect with the current decoy pool (mean_abs_charge_density_delta=0.1296), and charge density still beats the ensemble (0.8166 vs 0.7792). Treat raw AMP-vs-decoy AUROC as charge-inflated until a better charge-balanced negative set exists.