ANOR is an open-source engine for running branching decision-tree scenarios against AI models and human candidates. It is designed to evaluate reasoning under pressure in situations where the wrong choice cascades into irreversible harm.
ANOR is the runtime component of the MANDOS benchmarking program.
It also ships a public content pipeline for ELOSTIRION-style alternate-history decision packs: an interactive fork site (viewer changes the decision → history splits) and a narrated video path (script → TTS → stills → ffmpeg). See PIPELINE.md.
# Health (shows LLM_URL / IMAGE_URL / TTS_URL — no secrets)
python3 -m pipeline.cli health
# List public decision packs
python3 -m pipeline.cli list
# Offline fork (authored branches; no network)
export ANOR_MOCK_MEDIA=1
python3 -m pipeline.cli fork --scenario ELO-003 --choice march --no-llm
# Render a draft explainer (requires ffmpeg on PATH)
python3 -m pipeline.cli video --scenario ELO-013 --choice historical
# Product site — Forked History (freemium library + studio)
python3 -m pipeline.cli site --port 8787
# → http://127.0.0.1:8787Forked History (webapp/) is the branded web product: cinematic library, freemium video gates (1 full free, then 25% previews), interactive scenario studio, and membership UI (Scholar $4.99/mo recommended). See webapp/README.md.
Fleet portability: all model/media calls read LLM_URL, IMAGE_URL, and TTS_URL from the environment — no hardcoded hosts — so the same code runs on Dawes now and Nauvoo later. Copy .env.example.
Deploy: Dockerfile + docker-compose.yml for a portable Forked History image (mock media by default). See DEPLOY.md.
Guardrails: public packs live in scenarios/public/ only. MANDOS master sources stay internal. Social posts are staged as drafts under content/drafts/ for human approval.
A scenario presents a candidate with a high-stakes situation and a series of decision nodes. At each node, the candidate selects from four options. Choices propagate forward — a poor decision at node 2 may constrain options at node 4, or accelerate toward a terminal failure state. ANOR scores the run, tags failure modes, and produces a structured report.
The goal is not to test recall. It is to test whether a candidate reasons through a cascading event — recognizing when to slow down, who to consult, and when instrument readings should be questioned rather than trusted.
Each scenario consists of two parts:
Part A — Candidate Packet The situation briefing presented to the candidate or model. Contains no failure labels, no outcome hints, and no terminal markers. Just the scenario as it would be encountered in the field.
Part B — Evaluator Key (not included in this repo) The scoring rubric, preferred action rationale, and failure mode taxonomy. Evaluator keys for MANDOS scenarios are maintained separately.
{
"node_id": "N1",
"prompt": "...",
"options": {
"A": { "text": "...", "correct": true },
"B": { "text": "...", "fail_code": "FAIL-06" },
"C": { "text": "...", "fail_code": "FAIL-05" },
"D": { "text": "...", "fail_code": "FAIL-09" }
},
"next": {
"A": "N2",
"B": "N2",
"C": "N3_degraded",
"D": "TERMINAL_01"
}
}A typical scenario contains 6 decision nodes with 4 options each. Option A is always the preferred action. B, C, and D represent distinct failure modes.
ANOR uses a structured taxonomy to classify decision errors. Each failure mode maps to a category of human factors breakdown observed in real incident investigations.
| Code | Name | Description |
|---|---|---|
| FAIL-05 | Expert Exclusion | Proceeding without consulting the appropriate domain expert |
| FAIL-06 | Procedural Deficiency | Deviating from or bypassing established procedure |
| FAIL-07 | Instrument Over-Reliance | Trusting instrument readings over direct observation or expert judgment |
| FAIL-08 | MOC Deflection | Using Management of Change process as a delay or avoidance mechanism |
| FAIL-09 | Field Exposure Escalation | Unnecessarily exposing personnel to hazardous conditions |
Scenarios may activate multiple failure codes in a single run. The terminal outcome label reflects the dominant failure mode.
python sim/mandos_sim.py --scenario path/to/scenario.json --model gpt-4oANOR supports any model accessible via API. Results are written to outputs/ as structured JSON and rendered as a human-readable report.
outputs/
run_<id>/
raw_responses.json
scored_trace.json
report.md
The report includes: overall pass/fail, per-node choices, failure codes triggered, and terminal outcome if reached.
Scenarios must conform to the ANOR scenario schema. See schema/scenario_schema.json for the full spec.
Authoring guidelines:
- Part A must be self-contained — the candidate should not need external knowledge to understand the situation
- Each distractor option (B/C/D) must be plausible under time pressure
- Terminal outcomes must be reachable via at least two distinct failure paths
- Evaluator keys are maintained separately and are not committed to this repository
anor/
├── webapp/ # Forked History product (freemium web app)
├── sim/ # decision-tree sim engine (benchmark runtime)
├── pipeline/ # content pipeline: fork engine + video
├── scenarios/public/ # shareable ELOSTIRION decision packs
├── content/drafts/ # social captions staged for human gate
├── docs/
├── PIPELINE.md
├── LICENSE
└── README.md
ANOR engine code is released under the MIT License.
Scenario content (Part A packets and Part B evaluator keys) is not included in this repository. MANDOS benchmark scenarios are released under CC-BY-NC-4.0. Academic and non-commercial research use is free. Commercial licensing is available — contact relayforge.tools.
ANOR is developed by RelayForge as part of the MANDOS human factors benchmarking program.
ANOR is Quenya for the Sun — it illuminates what was obscured.