Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents
Paper • 2609.00065 • Published • 1
role string | source_revision string | command string | working_directory string | python_version string | exit_code int64 | tests_passed int64 | tests list | source_hashes dict | bounds string |
|---|---|---|---|---|---|---|---|---|---|
Local inert Python fixture-contract verification; not Go runtime execution | 7d7b8f1da2b1516a9d834e251d1795f0a59931a4 | python -m unittest -v benchmarks.scripts.test_benchmark_actionfacts_chain_fixtures | pinned source checkout (private local location omitted) | 3.12.14 | 0 | 4 | [
{
"name": "test_checked_in_fixture_validates",
"status": "ok"
},
{
"name": "test_checked_in_fixture_is_reproducible",
"status": "ok"
},
{
"name": "test_complete_proof_rejects_identity_drift",
"status": "ok"
},
{
"name": "test_only_over_bound_control_exceeds_eight_events",
... | {
"benchmarks/scripts/test_benchmark_actionfacts_chain_fixtures.py": "2e2d6d92c0221e21a5c31936cdb972fcf412cbc1451b1bd0e4c77e14140a701d",
"benchmarks/scripts/benchmark_actionfacts_chain_fixtures.py": "420a77ed4e048674fd35db0e713a9b5f61836a816ca0a2dd609f8a7f2fc8469f",
"benchmarks/fixtures/actionfacts-chain-identity... | Four checked-in Python construction and identity tests; no claim of end-to-end enforcement, complete proof coverage, or full-benchmark rerun. |
Runtime Guardrail Research Artifacts
This collection releases derived aggregate evidence, original figure data and plots, offline analysis code, and reproducibility manifests. The current research portfolio has 4 consolidated empirical manuscript directions: scanner configuration/gate comparisons, small-model score interfaces and source transfer, runtime cascade action composition, and deterministic-rule maintenance/evidence contracts. Current authoring/delivery state: Four consolidated empirical manuscripts and standalone supporting supplements; delivery verification is recorded in separate source-bound local receipts. This card does not certify acceptance, deployment validity, or completion of manuscript production.
paper06/reproduction: Advisory Decisions and Sequential Exits in Tool-Call Guardrail Cascades. How do confirms, terminal decisions, stage ordering and exit bands change final disposition flags and judge workload?
The paperXX paths also preserve thirteen historical analysis partitions. These are stable evidence identifiers, not thirteen active papers or independent replications. In particular, the current paper03/reproduction companion consolidates static and runtime rule evidence; earlier paper13 remains historical runtime-rule evidence. Small-model and runtime-cascade analyses share two evaluation populations, and repeated revisions or model outputs are not independent security trials. Each central result has one manuscript owner.
Archived measurements, fresh aggregate arithmetic, and exploratory reanalysis are labeled separately. Source-defined positive and block tags are not independently adjudicated authorization truth. A detection flag, advisory confirm, emitted block decision, and actually prevented unauthorized effect are different outcomes. Refitted score cutoffs are descriptive; frozen transfer is a separate analysis. Existing multi-tier guards and deterministic/semantic policies are relevant prior art, so the portfolio claims empirical observations rather than invention of these broad designs.
paper_catalog.csv records each module's relationship to the current portfolio. Protected source payloads, prompts, credentials, original private repository identities, individual score vectors, and manuscript prose are withheld. Version digests bind bytes; they do not validate scientific labels. AI-assisted owner-side working assessments are not external human peer review or editorial decisions.
Reproduction
Use the module-specific reproduction instructions with environment-requirements.txt. Aggregate arithmetic and figure replay require no provider calls or protected inputs. Full row-level or inference replay can require authorized upstream inputs and historically matched harness/tokenizer/configuration dependencies. Recovered current source does not by itself establish the exact historical inference implementation. Public release replay does not imply new inference, training, live tool execution, or human adjudication.
recompute_aggregate_metrics.py checks released confusion-matrix arithmetic; recomputed-confusion-metrics.csv records results. Nominal independent-binomial intervals, where present in legacy modules, are illustrative diagnostics and do not replace dependence-aware or selection-aware analysis. Use the current companion's stated estimands and assumptions. SHA256SUMS.json inventories released bytes. Preserve negative findings and legacy evidence even when their standalone manuscript is retired.
Rights and workflow tools
Derived numerical data and original figures use CC BY 4.0. Original aggregate analysis code is released under the MIT terms in LICENSE-CODE.txt. Third-party sources retain their own terms; this release grants no rights to withheld upstream inputs.
Installed academic skills support writing, citation verification, visualization, statistical analysis, hypothesis development, evidence critique, review and venue formatting. Installed revisions: scientific-agent-skills 154988403bb5a18e9d3c0ce4e6d5e2e4b184a298 and LitReviewSkill a53cd419352e4dd05958f67340fde3642d84abc3. Attribution: Kassis et al., Scientific Agent Skills (2026), workflow repository, and LitReviewSkill. Skills do not certify validity or novelty.
- Downloads last month
- 272