DOS benchmarks

Just want to run one? → BENCHMARKS.md. The standardized runner gives all six a single surface: python -m benchmark._run list | preflight <b> | run <b> [--arm A] | status. It shells the existing entrypoints (never reimplements them), resolves a named arm to its DOS_* env via the shared _arms.py vocabulary, loads the Gemini key from .env for paid arms, and stamps each run with the kernel SHA so status can flag stale numbers. The full run-surface inventory behind it is _BENCH_MAP.md.

This directory holds six independent research programs, each measuring a different DOS claim against a real corpus. They share no measurement code path (only the run surface above); pick the one that matches your question. The four tabled below are the load-bearing programs; fleetforge/ (newest fleet-collision sim) and the two offline detector replays (agenthallu/, agentprocessbench/) are in list. The numbered design notes that interpret these runs live in ../docs/ (see ../docs/README.md's research-arc tables and ../docs/ENTERPRISEOPS_ARC.md).

Benchmark	Research question	Read first	Numbers	Vocabulary
enterpriseops/	Does the intervention ladder (OBSERVE/WARN/DEFER/BLOCK) and arg_provenance lift a model on ServiceNow ITSM tasks?	THEORY_LADDER.md (how the tiers run) → HANDOFF_next_agent.md (what's next)	RESULTS.md	—
fleet_horizon/	Open-loop vs closed-loop fleet integrity — how many lies/overwrites does adjudication catch, at what review cost?	README.md	RESULTS.txt	—
toolathlon/	Do byte-clean in-flight detectors (dangling_intent / tool_stream / terminal_error) fire on a third-party-scored trajectory corpus, and do they convert to a FIX?	EXPLAINER.md (first-time) → AB_RUN_RECIPE.md (run the A/B) → HANDOFF.md (state)	(in _results/)	GLOSSARY.md — byte-author / byte-clean / net-new / additivity / SSOT
iot_tier/	Across a model-size ladder (frontier→mid→small→IoT-class), where does DOS's recoverable-failure fraction PEAK and where does it COLLAPSE — is the proof point the weakest model or the middle one?	README.md	RESULTS.md	calibrated-sim / recoverable-fraction / the gemini-null self-test

(Many more $0 programs are in python -m benchmark._run list — the table above is the load-bearing subset, not the full inventory; the inventory is _BENCH_MAP.md.)

Reading order by audience

• Kernel developer tuning intervention costs → enterpriseops/THEORY_LADDER.md then RESULTS.md.
• Fleet operator validating orchestrator trust → fleet_horizon/README.md then RESULTS.txt.
• Detector researcher measuring DETECT→FIX → toolathlon/EXPLAINER.md, then GLOSSARY.md for the vocabulary, then AB_RUN_RECIPE.md to reproduce.
• "Does DOS help small/edge (IoT-class) models?" → iot_tier/README.md (the rise-then-collapse curve + the honesty contract) then iot_tier/RESULTS.md.

Notes

• Live-run outputs are scratch, not source. The live_results*/, _results/, and the cloned enterpriseops-gym/ are .gitignored (large, re-derivable). The harness code and the scored summaries (the docs above + the numbered docs/NN_*.md) are tracked; the raw trajectory dumps are not. See the root .gitignore.
• enterpriseops/_PAIRED_RESULT.md is superseded — a small (n=20) rewind A/B whose favorable draw was refuted by a larger run. Read ../docs/172_the-rewindable-fix-loop-experiment.md §3.5 / §9 for the settled result, not that file in isolation.
• These benchmarks are consumers of the kernel, never part of it — the same one-way arrow as the MCP server and the release tooling (see ../CLAUDE.md).