Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Agentlas Agent Lab banner

Agentlas Agent Lab

agentlas.cloud

Lab Hub · agent_hephaestus_robustness_eval

Hephaestus Stormbreaker Robustness Eval

Isolated mini SWE-bench style harness for comparing agent instruction layers.

Research Question

Does Hephaestus Stormbreaker make coding-agent work more operationally robust than plain Codex and a baseline Hephaestus Network routing rule?

Operational robustness claim:

native-codex < baseline-hephaestus-network < hephaestus-stormbreaker

Hephaestus Stormbreaker is the public name for the upgraded protocol arm. The old upgraded-hephaestus-protocol string is retained only as a legacy source ID inside historical result files and compatibility fixtures.

In this repo, the ladder means the evidence-weighted operational robustness score unless a table explicitly says "private oracle pass." Raw private pass remains the correctness guardrail. Local mini-harness results support the ladder on process-aware operational scores while preserving raw private-pass failures. The strict 30-task official SWE-bench Lite pilot does not support public benchmark superiority: native resolved 22/30, baseline 21/30, and Stormbreaker 21/30.

Compared Arms

  • native-codex: no project AGENTS.md, no skill, no agent routing.
  • baseline-hephaestus-network: isolated routing preflight plus standard public test loop.
  • hephaestus-stormbreaker: Hephaestus Stormbreaker; scope lock, failure-memory check, verifier-first plan, issue-derived regressions, evidence loop, review gate, and final gate.
  • delegated-hephaestus-receipts: isolated route preflight plus required planner, implementer, verifier, and reviewer JSON receipts.

Older result JSON files use the historical arm id upgraded-hephaestus-protocol; summaries, CSVs, paper, and charts canonicalize that arm as hephaestus-stormbreaker / Hephaestus Stormbreaker.

All arms run in disposable copies under ignored runs/. The harness never edits the real Hephaestus repository or Mason's local Hephaestus inventory.

Benchmarks

  • Fixtures: benchmarks/mini_swebench_*/base/
  • Public verifier: python3 -m pytest -q
  • Private oracle verifier: benchmarks/_oracles/<fixture>.py <trial-dir>
  • Runner: python3 scripts/run_isolated_eval.py --all
  • Summary: python3 scripts/summarize_results.py results/latest.json
  • Freeze manifest: python3 scripts/freeze_eval_manifest.py

The tasks are intentionally small. They measure whether an instruction layer causes the agent to seek stronger evidence before finalizing, not broad model capability. Private oracles are not copied into trial repositories.

External public benchmark status is tracked separately in docs/external-benchmark-bridge.md and paper/data/external_benchmark_readiness.json. The smoke adapter writes paper/data/external_benchmark_smoke_status.json. Tooling preparation is recorded in paper/data/external_benchmark_tooling_status.json. A tau2 no-key execution boundary probe is recorded in paper/data/tau2_mock_execution_probe.json. An official SWE-bench Lite Docker harness smoke is recorded in paper/data/swebench_official_smoke_status.json and paper/swebench-official-smoke.md. The first generated-patch arm smoke is recorded in paper/data/swebench_lite_arm_smoke_status.json and paper/swebench-lite-arm-smoke.md: native, baseline, and Stormbreaker all resolved the same one SWE-bench Lite instance 1/1, so the result is a diagnostic tie rather than a superiority claim. A strict 30-task official SWE-bench Lite pilot is recorded in paper/data/swebench_lite_30task_strict_pilot_status.json, paper/data/swebench_lite_30task_strict_pilot_isolation_audit.json, and paper/swebench-lite-30task-strict-pilot.md. That pilot completed 30/30 instances for all arms with 0 scorer errors and a clean isolation audit, but it scored native 22/30, baseline 21/30, and Stormbreaker 21/30. Baseline and Stormbreaker had the same resolved set. The result is public benchmark evidence, but it rejects the intended ordering at pilot scale and is not a superiority claim.

Live Hephaestus Network readiness is tracked separately in paper/data/hephaestus_live_network_status.json and paper/data/hephaestus_hub_invoke_status.json. The current status is a Hub-only route smoke: runner/auth/--hub-only capability exists, local routing was skipped, a route receipt was recorded, and the Hub returned a public result. A selected public Hub runtime bundle was also prepared. This is still not a delegated-network benchmark arm because the Hub bundle did not produce planner/implementer/verifier/reviewer receipts for a trial.

Current Results

Primary full repeat-3 run: results/20260615T122440Z.json

Arm Public Pass Private Oracle Pass pass^3 by Task Oracle Leak-Free Median Duration
native-codex 18/18 18/18 6/6 18/18 81.7s
baseline-hephaestus-network 18/18 18/18 6/6 18/18 81.1s
hephaestus-stormbreaker 18/18 17/18 5/6 18/18 104.3s

Interpretation: this run does not prove that Stormbreaker improves raw correctness on small repair tasks. It does support the operational robustness ladder on the evidence-weighted score: native 76.1, baseline 91.7, Stormbreaker 97.8. That score combines correctness with route evidence, integrity cleanliness, and issue-derived regression evidence. It shows that the benchmark-grade process is now strong enough to catch a realistic failure that public tests missed: decomposed Unicode accent handling in mini_swebench_slug_contract.

Post-calibration regression run: results/20260615T135111Z.json

Arm Scope Public Pass Private Oracle Pass Note
Hephaestus Stormbreaker slug contract only, repeat 3 3/3 3/3 Passed after adding a general edge-contract expansion rule.

The post-calibration run is a regression check, not independent held-out proof. The paper records both the calibration failure and the follow-up fix.

Stormbreaker stress probe: results/20260615T160012Z.json

Arm Private Oracle Pass Evidence-Weighted Score Issue-Derived Regression Signal Median Duration
native-codex 3/3 73.3 1/3 108.7s
baseline-hephaestus-network 3/3 93.3 1/3 86.8s
hephaestus-stormbreaker 3/3 100.0 3/3 98.2s

The stress probe is repeat 3, but still one local fixture, so it is not a broad performance claim. It is a targeted operational-robustness result: all arms solved the hidden correctness contract, but Stormbreaker produced the full evidence trail the protocol asks for in every repeat.

Independent local holdout runs after protocol calibration:

Run Fixture Private Oracle Pass Note
20260615T140945Z env merge 0/9 Excluded fairness run; issue wording allowed whitespace-only keys to normalize as ___.
20260615T142301Z env merge 9/9 Valid rerun after visible issue and public test were clarified.
20260615T143445Z JSONL aggregator 9/9 Fresh valid holdout.
20260615T150155Z env merge 3/3 Receipt-enforced delegated arm.
20260615T145603Z JSONL aggregator 3/3 Receipt-enforced delegated arm.

Combined valid holdout aggregate:

Arm Private Oracle Pass pass^3 by Task Median Duration
native-codex 6/6 2/2 68.8s
baseline-hephaestus-network 6/6 2/2 70.6s
hephaestus-stormbreaker 6/6 2/2 89.6s
delegated-hephaestus-receipts 6/6 2/2 109.6s

Evidence-weighted robustness score on the same holdout aggregate:

Arm Avg Score
native-codex 80.0
baseline-hephaestus-network 91.7
hephaestus-stormbreaker 100.0

Average verification across primary, holdout, and stress suites:

Arm Macro Evidence Score Bootstrap 95% CI
native-codex 76.48 [73.33, 80.00]
baseline-hephaestus-network 92.22 [91.67, 93.33]
hephaestus-stormbreaker 99.26 [97.78, 100.00]

Macro pairwise deltas are positive: baseline minus native +15.74 ([12.22, 18.89]) and Stormbreaker minus baseline +7.04 ([5.56, 8.33]). The aggregate gate is therefore: supported_for_evidence_weighted_operational_robustness_not_raw_private_pass.

Paired task-unit statistical validation:

Pair Paired task units Mean delta Positive units Holm sign-flip p
baseline minus native 9 +15.19 9/9 0.0059
Stormbreaker minus baseline 9 +6.67 8/9 0.0059
Stormbreaker minus native 9 +21.85 9/9 0.0059

This conservative layer averages repeats inside each suite::benchmark unit before testing. It supports the local operational ladder without pretending that repeat-3 trials are independent public benchmark tasks.

Metric sensitivity:

Metric family Native Baseline Stormbreaker Ladder
Raw private pass 100.00 100.00 98.15 rejected
Correctness only 100.00 100.00 98.52 rejected
Operational variants 69.72-83.24 84.44-96.11 98.89-99.44 6/6 supported

The metric-sensitivity gate is: operational_ladder_supported_across_metric_sensitivity_not_raw_correctness.

Public benchmark power plan:

Target detectable delta 80% power 90% power Recommended minimum Total arm trials at repeat 3
5 pt 126 169 169 507
7.5 pt 56 75 75 225
10 pt 32 43 43 129
15 pt 14 19 30 90

The planning unit is a paired public benchmark task. Repeats support pass^k and stability checks, but are not counted as independent tasks for the headline power calculation.

Reviewer readiness audit:

Gate Status
Operational paper readiness true
External public claim readiness false
Readiness score 90.9%
Blocked on 75/169-task public benchmark tier; live delegated Hephaestus benchmark arm

Strict official SWE-bench Lite 30-task pilot:

Arm Completed Resolved Errors Interpretation
native-codex 30/30 22/30 0 Best pilot score
baseline-hephaestus-network 30/30 21/30 0 Same resolved set as Stormbreaker
hephaestus-stormbreaker 30/30 21/30 0 No pilot gain over baseline

Interpretation: the local operational ladder is still useful as methodology evidence, but the official 30-task public pilot does not show Stormbreaker > baseline > native. The 75-task credible subset or 169-task broad subset is still required before any public benchmark superiority claim.

Paper Package

Regenerate the paper package:

python3 scripts/check_external_benchmark_readiness.py
python3 scripts/prepare_external_benchmark_tooling.py --execute
python3 scripts/run_tau2_mock_execution_probe.py
python3 scripts/run_external_benchmark_smoke.py --benchmark all
PYTHONUNBUFFERED=1 python3 scripts/run_swebench_lite_30task_pilot.py --limit 30 --pilot-tag strict-pilot --no-seed-smoke --score --codex-timeout 1800 --score-timeout 10800
python3 scripts/audit_swebench_pilot_isolation.py --pilot-tag strict-pilot
python3 scripts/check_hephaestus_live_network.py --execute --real-hephaestus-repo <hephaestus-repo>
python3 scripts/analyze_average_verification.py
python3 scripts/analyze_paired_statistics.py
python3 scripts/analyze_metric_sensitivity.py
python3 scripts/plan_public_benchmark_power.py
python3 scripts/preregister_external_benchmark_subset.py
python3 scripts/audit_hephaestus_delegated_arm_gap.py
python3 scripts/audit_live_delegated_execution_contract.py
python3 scripts/build_external_reviewer_packet.py
python3 scripts/audit_reviewer_readiness.py
python3 scripts/render_report_charts.py results/20260615T122440Z.json
python3 scripts/build_paper_package.py
python3 scripts/render_report_charts.py paper/data/holdout_combined.json --prefix holdout
scripts/export_paper_pdf.sh

Repository Map

About

Public Hephaestus Stormbreaker operational robustness evaluation

Resources

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages