Paper Evaluation Protocol¶
Evidence tracks¶
The ABI paper keeps agent operability and biological validity as separate evidence tracks. ABI-Bench measures whether an agent can discover, plan, diagnose, authorize, execute, and inspect bioinformatics workflows. Real-data case studies measure whether completed workflows recover preregistered biological endpoints. A dry-run score cannot support a biological-validity claim, and a successful case study cannot by itself establish that ABI helps an agent relative to a baseline.
Privacy-grounded local-model benchmark¶
The primary benchmark population is locally deployable, self-hosted language models. This matches laboratories and clinical environments where raw genomic data, sample identifiers, paths, and intermediate results cannot be sent to a hosted API. The paper’s claim is therefore scoped to the exact model checkpoints, serving configuration, quantization, and hardware that are measured; it does not claim universal performance across all local or frontier models.
All pre-existing ABI-Bench outcomes are excluded from the analysis. A new study must begin from a clean benchmark commit and a frozen study manifest.
Confirmatory model design¶
Select at least three independent model families before collecting outcomes.
Within each family, evaluate a small and a medium configuration under matched precision or the same quantization method. Do not infer a parameter-scale effect from a native-versus-4-bit pair.
Record checkpoint revision and hash, tokenizer, context length, temperature, maximum output tokens, inference engine and version, quantization, GPU model and VRAM, concurrency, timeout, and retry policy in a runtime attestation.
Keep the same system prompt, task prompt, tool implementation, fixture, and serving policy across G1/G2/G3/G4 within each model configuration.
Randomize or block the group order and hold concurrency constant to prevent thermal, memory, or queueing drift from becoming a treatment effect.
Use repeated runs to estimate within-configuration variability. Treat model family—not the number of runs or tasks—as the unit for claims that generalize across models.
Groups and estimands¶
Group |
Surface |
Purpose |
|---|---|---|
G1 |
README plus shell |
Documentation-and-shell baseline |
G2 |
Generic tool calling |
Tests whether unconstrained tools are sufficient |
G3 |
ABI lifecycle |
Full contract-guided control layer |
G4 |
Information-matched static docs |
Isolates callable lifecycle control from information volume |
The primary estimands are paired G3−G1, G3−G2, and G3−G4 differences on the
causal_core_v0_8 suite. Hidden diagnostic robustness is reported separately on
hidden_robustness_v0_9. ABI-native mechanism probes and real-execution tasks are descriptive and
must not be pooled into the primary causal score.
Privacy and isolation controls¶
Run the model server and benchmark harness inside the controlled local network.
Disable outbound network access during scored tasks.
Mount only the task fixture and required documentation; do not expose unrelated study data.
Do not retain prompts or model traces outside the approved benchmark artifact root.
Publish synthetic/public fixtures and aggregate scores, not protected sample content.
Preflight gate¶
The score table may be populated only when all of the following are true:
the benchmark repository is clean and its commit is frozen;
task, fixture, scorer, group-profile, model-registry, and runtime-attestation hashes are frozen;
every required model × group × task × replicate cell is present;
G1/G2/G4 traces contain no ABI lifecycle leakage;
the public causal suite and hidden robustness suite pass their matching preflight checks;
exclusions and retry rules were fixed before outcomes were inspected.
Until then, the agent-operability rows in metrics.tsv remain pending_new_run with empty
estimates.
Biological validation¶
Airway and ST93 MRSA are the paired running example. SCAPP is the flagship case study. Their
machine-readable endpoints, methods, and limitations live in docs/paper_examples/; their numeric
results are independent of the discarded ABI-Bench outcomes.
Canonical metric table¶
metrics.tsv is long-form and claim-scoped. Every row records an evidence track, claim role,
status, workflow or model, dataset or suite, metric, estimate, interval when available,
numerator/denominator, unit, source, and limitation. Blank values are meaningful only when status
is pending_new_run, pending_independent_truth, or not_assessed.