The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
FIRE artifact
This artifact accompanies FIRE: Failure-Informed Runtime Engineering for Reliable Language-Model Agents. It contains the frozen policies, trajectory-free attempt outcomes, analysis code, and derived tables used in the paper. It does not contain prompts, agent transcripts, tool calls, or task workspace contents.
Start here
Python 3.10 or newer is the only requirement for analysis. From this folder:
python3 analysis/verify_results.py
The command validates the attempt populations and task digests, regenerates
the statistical tables from attempt-level inputs, compares them with the
published tables, and verifies every file against MANIFEST.sha256. It does
not rewrite the artifact or its manifest.
To inspect regenerated tables separately:
python3 analysis/reproduce.py --out /tmp/fire-reproduced
See REPRODUCING.md for the verification levels and experiment requirements,
and SCHEMAS.md for field definitions.
Contents
policies/: frozen Luna, Terra, Sol, and Terra control policy source.configs/: model routes, harness settings, run-selection and replacement ledgers, exclusions, and inference settings.results/complete-suite/attempts.csv: 1,044 sanitized scalar outcome rows, covering six cells, 87 tasks, and two attempts per task.results/five-arm-panel/attempt-funnel.csv: 300 selected scalar outcome rows.results/five-arm-panel/lineage-audit.csv: selection and replacement provenance for the panel. This is metadata, not trajectory content.analysis/reproduce.py: analysis implementation for all derivable released CSV tables, including bootstrap intervals and permutation tests.analysis/verify_results.py: non-mutating cross-table and checksum validator.POLICY-PROVENANCE.md: relationship between executed and paper-layout policy copies.
Paper map
| Paper item | Artifact source |
|---|---|
Fig. 2 and Tables 1, 3 (fig:controls, tab:contrasts, tab:funnel) |
results/five-arm-panel/ |
Table 4 (tab:suite) |
results/complete-suite/primary.csv, comparisons.csv |
Table 5 (tab:transitions) |
results/complete-suite/task-transitions.csv |
Table 6 (tab:terminations) |
results/complete-suite/terminations.csv |
Fig. 4 and Table 7 (fig:scatter, tab:tiers) |
results/complete-suite/cost-latency.csv, primary.csv |
| Policy-family listings | policies/, configs/policy_profiles.json |
| Population and deviations | configs/final-analysis-run-ledger.yaml, results/complete-suite/protocol-deviations.csv |
Population and outcomes
Terminal-Bench 2.1 contains 89 tasks. qemu-alpine-ssh and qemu-startup
were uniformly excluded because their pinned Bullseye images could not install
the agent from the live Debian security mirror. Every complete-suite cell
therefore contains 87 tasks and 174 attempts.
The primary outcome is the verifier reward. A successful verifier can coexist
with a later timeout lifecycle label. This occurs in exactly three Sol rows:
- baseline,
extract-moves-from-video - policy,
extract-moves-from-video - policy,
make-doom-for-mips
This explains why Sol verifier passes are 131 and 140 while passed
termination labels are 130 and 138. No rows are dropped or relabeled.
Scope
The released scalar records are sufficient to regenerate the reported tables, confidence intervals, permutation tests, transitions, and aggregate resource statistics. Full model reruns require Terminal-Bench, Harbor, Codex CLI, FailproofAI runtime integration, paid model access, and credentials. The raw run directories are not part of this package, so this is an analysis-reproducible artifact rather than a self-contained model-execution archive.
Licenses
Original software and policy source are Apache-2.0 licensed; see LICENSE.
Original manifests, annotations, derived data, and scalar outcome records are
CC BY 4.0 licensed; see LICENSE-DATA. Third-party provenance is in NOTICE.
- Downloads last month
- 166