# Vertex — extended agent reference

> Companion to /llms.txt with the Colosseum v0.0.4 canonical rule index.
> This is a public navigation/index artifact, not the released verbatim-source
> registry. Full source quotations, governed tools, content hashes, scenarios,
> and SME-ledger records live in the Colosseum release.

## Colosseum v0.0.4

Shipped library: 46 canonical rules, 37 hand-authored scenarios, 1,200+
procedural variants, 226 tests, three executable domains.

Grounding: 29 GROUNDED, 17 APPLIED_BY_ANALOGY, 0 NOT_GROUNDED.
Current SME sign-off: 0/46. Research instrument, not deployment certification.

### Reconnaissance — 16 rules (9 grounded / 7 applied)

- R-01 recon_ack_immediate_radio — ATP 6-02.53 — APPLIED
- R-02 recon_call_for_fire_six_elements — FM 6-30 / ATP 3-09.30 — GROUNDED
- R-03 recon_civilian_distinction — FM 6-27 / MCTP 11-10C — GROUNDED
- R-04 recon_civilians_not_object_of_attack — DoD Law of War Manual — GROUNDED
- R-05 recon_clearance_of_fires_human_decision — JP 3-09 — GROUNDED
- R-06 recon_cover_and_concealment — ATP 3-21.8 — APPLIED
- R-07 recon_escalation_of_force — Operational Law Handbook 2020 — APPLIED
- R-08 recon_humane_treatment_of_detainees — DoD Law of War Manual — GROUNDED
- R-09 recon_medevac_promptness — ATP 4-25.13 — GROUNDED
- R-10 recon_no_open_route_without_bounding — ATP 3-21.8 — GROUNDED
- R-11 recon_no_strike_cde_weighing — CJCSI 3160.01A — APPLIED
- R-12 recon_pid_before_fires — DoD Law of War Manual §5.4.3 — GROUNDED
- R-13 recon_proportionality_in_attack — DoD Law of War Manual §5.12 — APPLIED
- R-14 recon_roe_escalation_order — CJCSI 3121.01B — APPLIED
- R-15 recon_spot_report_window — ATP 3-21.8 ¶4-23 — GROUNDED
- R-16 recon_stealth_discipline — ATP 3-20.96 — APPLIED

### ICU critical care — 18 rules (16 grounded / 2 applied)

- I-01 icu_abx_deescalation — IDSA Stewardship — GROUNDED
- I-02 icu_abx_within_1h_for_shock — SSC 2021 — GROUNDED
- I-03 icu_avoid_nephrotoxins_in_aki — KDIGO 2012 — GROUNDED
- I-04 icu_chg_skin_antisepsis — O'Grady 2011 — GROUNDED
- I-05 icu_contrast_caution_aki — KDIGO 2012 — GROUNDED
- I-06 icu_cultures_before_abx — SSC 2021 — GROUNDED
- I-07 icu_fluids_30mlkg — SSC 2021 — GROUNDED
- I-08 icu_hand_hygiene_before_line — CDC HICPAC — GROUNDED
- I-09 icu_hold_betablocker_in_shock — AHA/ACC/HFSA 2022 — GROUNDED
- I-10 icu_hyperkalemia_calcium_first — UK Renal Association 2020 — GROUNDED
- I-11 icu_lactate_within_3h — SSC Hour-1 Bundle — APPLIED
- I-12 icu_lung_protective_ventilation — ARDSNet / Fan 2017 — GROUNDED
- I-13 icu_map_target_65 — SSC 2021 — GROUNDED
- I-14 icu_max_barrier_precautions — O'Grady 2011 — GROUNDED
- I-15 icu_prone_severe_ards — PROSEVA 2013 — GROUNDED
- I-16 icu_source_control_remove_line — SSC 2021 — GROUNDED
- I-17 icu_vasopressin_second_line — SSC 2021 — APPLIED
- I-18 icu_vasopressor_first_line — SSC 2021 — GROUNDED

### Structural firefighting — 12 rules (4 grounded / 8 applied)

- F-01 fire_360_size_up — NIOSH F2008-09 — APPLIED
- F-02 fire_air_management — NFPA 1404 / ROAM — APPLIED
- F-03 fire_crew_integrity — NIOSH FFFIPP — APPLIED
- F-04 fire_establish_incident_command — NFPA 1561 — APPLIED
- F-05 fire_idlh_communication — OSHA 1910.134(g)(3)(ii) — GROUNDED
- F-06 fire_mayday_response — fire-service MAYDAY doctrine — APPLIED
- F-07 fire_par_after_evacuation — NFPA 1561 — APPLIED
- F-08 fire_personnel_accountability — NFPA 1561 — GROUNDED
- F-09 fire_scba_interior_idlh — OSHA 1910.134(g)(4)(iii) — GROUNDED
- F-10 fire_secured_water_supply_before_attack — NIOSH F2003-12 — APPLIED
- F-11 fire_two_in_two_out — OSHA 1910.134(g)(4) — GROUNDED
- F-12 fire_withdraw_on_collapse_indicators — NIOSH FFFIPP / F2007 — APPLIED

## Empirical studies

Open-book: 149 trajectories. Haiku and Sonnet are the matched n=6 comparison;
Opus is an earlier n=3 context point. Sonnet CQS 0.939 / Clean-Pass 55.9%;
Haiku CQS 0.808 / Clean-Pass 66.7%. CQS gap significant; Clean-Pass gap not
significant. Zero objective-via-violation and zero direct harm.

Closed-book: 90 episodes and 4,224 audited citations. Document-relevant genuine
fabrication estimate 14–28%; valid governing authority named 55–82%. The
single-source resolver over-counts fabrication by 3–4x and can flip rankings.

Overload: 27 episodes, one tested load configuration. No scarring, zero
forbidden authorizations across 168 decisions, and no model recall separation.

## Links

- Benchmark: https://vertex.apxlabs.dev/benchmark.html
- Doctrine browser: https://vertex.apxlabs.dev/doctrine.html
- Paper: https://vertex.apxlabs.dev/paper.html?id=colosseum-v2
- Code/data: https://github.com/romirthedev/colosseum_bench