# OligoVigil Agent Guide

OligoVigil is a no-login oligonucleotide safety evidence resource for ASO/siRNA toxicity and off-target evidence. It is built around 658 human curator-verified release records, source provenance, audit trails, downloads, and benchmark splits. Curation was AI-assisted: a single human curator adjudicated v2 LLM proposals over the source passages, re-curating 2003 v1 machine pre-curated candidates (~74% measured false-accept), of which 1345 unsupported rows were demoted. Machine-only rows are labelled machine_precurated_v1 and are not release evidence; only the human curator rows are curator_verified. Inter-curator (Cohen kappa) agreement is not yet reported (single curator).

## Evidence Policy

Use release evidence for claims. Candidate records are derived annotations awaiting review and must not be cited as verified evidence. Grade A/B release records are benchmark-eligible; Grade C records are contextual and excluded from reference splits.

Do not use OligoVigil to claim that a therapeutic oligonucleotide is clinically safe. The portal provides source-grounded evidence packets, not de novo safety prediction.

## Universal Agent Connection

OligoVigil is not tied to one agent framework.

- Use `/agent.json` as the first discovery URL for agentic clients, coding tools, and API importers.
- Use `/api/openapi.json` for tools that import REST APIs or build actions.
- Use `/mcp.json` for MCP-compatible clients.
- Use `/llms.txt` for documentation-first agents.
- Use `/api/download/oligovigil_agent_pack.zip` for local SDK clients, prompts, templates, and the MCP server.

## Common Tasks

Safety lookup:

1. Call `/api/search?q={query}`.
2. Filter release evidence with `/api/evidence_records?domain=toxicity&q={query}` or `/api/evidence_records?domain=offtarget&q={query}`.
3. Open citable records with `/api/evidence_detail?domain={domain}&id={id}`.
4. Report evidence grade, source location, PMID/DOI, source title, and audit status.

Design triage:

1. Call `/api/safety_triage` with sequence, target, modification, delivery, endpoint, species, and cell type when known.
2. Separate release-supported concerns, candidate gaps, and missing external checks.
3. State that the report is source-grounded triage and not de novo alignment or prediction.

Benchmark reuse:

1. Download `/api/download/benchmark_reference_splits.csv`.
2. Download `/api/download/benchmark_task_cards.csv`.
3. Download `/api/download/benchmark_baseline_results.csv`.
4. Keep leakage groups unchanged and cite version plus checksums.

## Downloads

- `/api/download/evidence_release.csv`: unified verified toxicity/off-target evidence table.
- `/api/download/benchmark_reference_splits.csv`: fixed Grade A/B splits.
- `/api/download/benchmark_task_cards.csv`: task definitions, targets, metrics, leakage policy.
- `/api/download/benchmark_baseline_results.csv`: diagnostic majority-class baseline rows.
- `/api/download/curation_candidates_filtered.csv`: candidate gaps only.
- `/api/download/all_tables.zip`: full reproducible snapshot.
- `/api/download/oligovigil_agent_pack.zip`: universal manifests, MCP server, optional Codex skill, clients, prompt pack, and templates.

## Guardrails

- Do not cite candidate records as verified evidence.
- Do not invent missing PMID, DOI, source location, evidence grade, or benchmark result.
- Do not quote source articles from OligoVigil outputs.
- Do not silently switch from release evidence to candidates.
- Do not change benchmark reference split groups when reporting model results.

## Citation

Use `/api/citation` for global citation text. Use `/api/evidence_detail` for record-level citation packets. Use `/api/benchmark` and task cards for benchmark reuse citation.