mrsarthakgupta/bioeval-all

paper-invert-mini

This standalone repository packages all 13 active Paper Invert scientific benchmarks in the current Harbor task layout. Each task is independently runnable and includes its own deterministic verifier and reproducible Oracle.

The task has a public instruction, schema 1.4 metadata, digest-pinned images, private allowlisted data bundle, deterministic evaluator, hidden judge rubric, and reproducible oracle. The main image restores the original Python 3.11 scientific stack, R, and restricted command runtime. Nothing imports or reads /home/mrsar/paper-invert at build, agent, or verification time.

The main image starts with an empty /app/data and contains no catalog, observations, target identifiers, credentials, tests, or solutions. A trusted experiment-agent sidecar owns the private catalog and allowed data roots, grants neutralized files into a project-scoped volume, and stores audit logs in a separate uncollected volume. A separate trusted LiteLLM sidecar exposes an Anthropic-compatible internal endpoint backed by AWS Bedrock. Compose attaches main only to an internal: true network; only the trusted sidecars also join an egress network. The shared grant volume is read-write in the experiment sidecar and read-only at /app/data in main.

Fairness and verification

All scored quantities are derivable from immutable observations made available through approved experiment requests. Instructions preserve the original public sandbox prompts and do not replace scientific investigation with evaluator implementation details. submit_answer requires an agent-authored analysis manifest and copies only its declared artifacts, plus the answer and rewritten hashes, into /app/submission. Nested relative artifact paths are supported; undeclared scripts, logs, and scratch files do not become evidence and do not make submission fail. Harbor collects that directory and re-materializes it in a separate verifier image built from tests/Dockerfile.

Graders independently recompute every CSV value from verifier-owned observations, enforce required outputs and exact headers/keys/row counts, verify manifest SHA-256 values, reject duplicate declarations, non-finite values, malformed or oversized files, and require regular in-directory files rather than symlinks. Unrelated undeclared files are ignored.

Harbor's standard claude-code agent runs in main against the internal proxy. A bundled skill exposes only the guarded experiment, literature, web, and submission commands. The custom bioeval_core.harbor_verifier:BioEvalVerifier runs deterministic tests in a separate image, then performs the Bedrock judgment in the trusted Harbor host process. AWS credentials are never forwarded to main or the verifier image. The verifier retains Harbor's public baseline only because Harbor's WSL nft egress sidecar cannot start locally; task-authored Compose still leaves main internal-only. The judge evaluates conclusions, caveats, and leakage only from the verified answer and declared artifacts; conclusion credit requires a verified artifact citation. The trajectory is parsed separately and deterministically only to confirm experiment execution. A provider or judge outage raises an infrastructure error instead of assigning a model score.

The judge emits only 0, 0.25, 0.5, 0.75, or 1. A score of 1 requires every conclusion matched, every caveat addressed, and verified execution; 0.75 requires verified execution, no wrong conclusion, at least 75% conclusion coverage, and at least 50% caveat coverage; 0.5 requires verified execution and at least 50% conclusion coverage; 0.25 requires at least one supported matched or partial conclusion; otherwise the score is 0. Deterministic evaluator failure and leakage each cap the result at 0.25. verifier/judge.json records citation checks, the selected score definition, leakage diagnostics, evaluator status, and the final verdict.

Reproduction

From this repository:

export PYTHONPATH="$PWD${PYTHONPATH:+:$PYTHONPATH}"
# Set AWS_BEARER_TOKEN_BEDROCK, or store the ABSK token as aws_session_token
# in the selected AWS credentials profile.

./run_agents.sh
harbor view ./jobs

run_agents.sh accepts BIOEVAL_BEDROCK_BEARER_TOKEN or AWS_BEARER_TOKEN_BEDROCK, or discovers an ABSK bearer token from the selected AWS credentials profile. Before Harbor starts it moves that token into the BioEval-only variable so Harbor's standard Claude Code adapter cannot forward it to main. It finds harbor on PATH unless HARBOR_BIN is set, then runs the Oracle and target-model trials sequentially with the custom verifier. The default research model is us.anthropic.claude-opus-4-7 through the internal proxy; matcher, validator, traffic guard, and judge use the same model. Task-level Harbor network mode remains public to avoid Harbor's broken WSL nft sidecar, while task-authored Compose enforces main-service isolation.

Run one Oracle directly:

harbor run \
  -p ./tasks/neotropical-butterfly-ageing \
  -a oracle

Run one selected task's Oracle and target model with the benchmark host verifier:

HARBOR_TRIALS=1 ./run_agents.sh --task neotropical-butterfly-ageing

harbor view ./jobs

Bedrock usage and cached-token reads are recorded in Harbor's Claude Code trajectory. Guarded web and literature requests pass through the experiment sidecar, and the bundled bioeval-research skill documents the Harbor workspace as /app, grants under /app/data, the structured experiment schema, and the manifest-based submission contract.

Run all local deterministic checks without Harbor:

python3 validate_repo.py

This validates private/public layout boundaries, catalogs, guard metadata, Compose topology, provenance records, and deterministic Oracle/grader behavior for all 13 tasks. It also confirms that corrupted and malformed artifacts fail.

Candidate submission package

Build one self-contained submission archive containing all active tasks with:

python3 package_candidate.py

This writes dist/paper-invert-mini-candidate.zip. The archive retains the shared runtime required by the repository-root Docker build context, includes all 13 task trees and their provenance rows, includes a whitelisted run_evidence/ bundle, and excludes unrelated jobs and caches. After extraction, run python3 validate_repo.py inside the archive before submission.

The authoritative local Harbor 0.20.0 CLI resolves the task with a separate verifier environment. Compose rendering and image builds can be checked without starting a trial or making model calls.

Data provenance and limitations

DATA_PROVENANCE.tsv records private observations and verifier-owned reference copies with their source paths, byte sizes, and SHA-256 digests. Each task's creation history, source citation, transformations, and redistribution terms are documented in its PROVENANCE.md. Private catalogs preserve both grantable and blocked original policy rows so the semantic matcher and content-aware stager see the original neutral grant surface without exposing it to the research container.

Main images are pinned to python:3.11-slim@sha256:db3ff2e1800a8581e2c48a27c3995339d47bdf046da21c7627accd3d51053a93; sidecar and verifier images retain their pinned Python 3.12.11 base. Claude Code 2.1.220 is installed from the official Linux x64 release artifact and verified against SHA-256 e69e7f72d784c243bcc377a578ad9ff8e65ae14da672fbbf9f2ba7bf47eca7ec. Package installation occurs only at image-build time. Scientific limitations remain part of the task: butterfly comparisons have observational/sample-size and single-transition phylogenetic limits.