Structured Dissent Protocol — 9 cognitive frames + 1 mandatory steel-man dissenter, exposed as an MCP server.
Imagine you have an important decision. Instead of asking one person, you ask nine advisors simultaneously — each with a different way of thinking:
- One looks at numbers and evidence
- One finds historical precedents — what happened the last 3–5 times
- One reasons from scratch, no assumptions
- One finds analogies from other domains (biology, military, finance)
- One maps the full system and its chain effects
- One surfaces the ethical tensions
- One pushes back in good faith
- One imagines what unlocks if everything goes 10×
- One assumes failure in 12 months and describes how
All nine respond in parallel. Their agreement is measured. High convergence means consensus. High dispersion means the decision is more fragile than it looks.
There is a tenth advisor required to disagree — not arbitrarily, but constructively. Their job: build the strongest possible argument against the consensus, without having read what the other nine said. They reason independently, anticipate where the nine will converge, and attack that conclusion with the best case they can build.
Why blind? If they read the nine first, they'd end up criticizing details. Reasoning in parallel, their dissent is genuine — it comes from a different direction, not as a reaction.
This is called steel-man: attacking the strongest version of the opposing argument, not the weakest. The opposite of a straw man.
The tenth is an Anthropic model. The eleventh is an OpenAI model — different company, different training, different reasoning patterns. Its job: read the tenth and sort their dissent into three piles:
- What holds → survived contrast with the nine; real signal
- What's revised → right direction, needs refinement
- What's discarded → model-specific bias, not a finding
A brilliant dissenter, plus an editor from a different school who separates valuable ideas from personal quirks.
Before the nine activate, an auditor reviews the question itself. It checks: is this decision reversible? how urgent is it? is the question well-formed? If poorly framed, it stops the process and proposes a reformulation — saving ~$1.00/run on questions that don't merit a full analysis.
A single opinion, however good, has blind spots. Nine diverse opinions reduce individual blind spots — but can share a common error (experts from the same field tend to fail the same way). The tenth is designed to attack that common error. The eleventh, coming from a different reasoning "family," keeps the tenth from just replacing one bias with another.
The result is not a single answer. It's a tension map: where there is solid consensus, where there are cracks, what risks you might be overlooking. The final decision is still yours — but you make it with much clearer eyes.
Demo — job offer · Demo — AI services · Paper · Limits · Methodology · CFI spec · Manifesto · Developer
Paste this prompt into Claude Code and it self-installs by running a deterministic shell script — no LLM step-following, no drift:
Install Henge from https://github.com/ChrisPiz/Henge. Idempotent flow:
1. Clone shallow (or pull if already there):
git clone --single-branch --depth 1 https://github.com/ChrisPiz/Henge-MCP.git ~/Henge \
|| (cd ~/Henge && git pull --ff-only)
2. cd ~/Henge && cp -n .env.example .env
3. Ask me for ANTHROPIC_API_KEY and OPENAI_API_KEY one at a time. When I paste each one, update the matching line in ~/Henge/.env in-place. Confirm only the LENGTH back to me ("got it, 108 chars") — never echo the value to the chat or any other tool.
4. Run the setup script — it handles Python ≥3.11 (with a 15-minute pyenv install fallback if missing), the venv, the editable install, the cross-cwd sanity check, key validation, MCP registration for every host installed (Claude Code, Claude Desktop, Cursor), and the /decide slash command:
cd ~/Henge && ./setup
5. When the script prints "✓ Henge installed.", tell me to fully quit Claude Code (close ALL terminals running `claude`) and reopen, then try `/decide should I take the new job?`.
Restart Claude Code fully when it's done, then try:
/decide should I take the new job?
Note: the
/decideslash command is Claude Code only. In Claude Desktop and Cursor, MCP tools don't appear as slash commands — you invoke Henge by writing your question normally ("Should I quit my job to start a company?") and Claude picks up thedecidetool from its description, or you can mention it explicitly ("use the decide tool to analyze ...").
For Claude Desktop, Cursor or any other MCP host, see the Manual install section in DEVELOPER.md.
Reads as one of three pre-registered states (full bin definition in
docs/cfi-spec.md):
- aligned-stable — the nine cluster tightly and CFI < 0.33
- aligned-fragile — the nine cluster tightly but CFI ≥ 0.33
- divided —
σacross the nine ≥ 0.03, no real consensus to attack
Claim (v0.6.0). Given a pre-registered protocol with temperature=0,
the same question produces the same CFI bin across runs, modulo numerical
noise from MDS init. v0.6 splits the run across two labs (Anthropic +
OpenAI) by design, so the synthesizer's hallucinations get caught by the
auditor instead of laundered into consensus. Validation against decision
quality (Henge-50 benchmark) remains in flight.
Nine consensus frames + one mandatory dissenter. Same prompt set across every
run; SHA-256 prefix exposed as henge.agents.PROMPTS_HASH and persisted in
every report.
| # | Frame | Lens | Model |
|---|---|---|---|
| 1 | empirical | quantification, base rates, [assumption] markers | gpt-5 |
| 2 | historical | precedents — what happened the last 3–5 times | Opus 4.7 |
| 3 | first-principles | reduce to physical / economic / logical atoms | gpt-5 |
| 4 | analogical | cross-domain mappings (biology, military, finance) | Sonnet 4.6 |
| 5 | systemic | feedback loops, second- and third-order effects | gpt-5 |
| 6 | ethical | deontological + consequentialist tension | Sonnet 4.6 |
| 7 | soft-contrarian | surgical reframe of the loaded silent assumption | gpt-5 |
| 8 | radical-optimist | what unlocks if it goes 10× better | gpt-5 |
| 9 | pre-mortem | assume it failed in 12 months — describe how | gpt-5 |
| 10a | Tenth Man — blind | pure steel-man · no view of the 9 | Opus 4.7 |
| 10b | Tenth Man — informed | sees 9 + blind · returns holds/revised/discarded | gpt-5 (cross-lab) |
Routing lives in henge/config/frame_assignment.py and is cross-lab by
design: synthesis and pure dissent stay in Anthropic; audit roles cross to
OpenAI. Override via FRAME_MODEL_MAP if you want the legacy single-model
configuration.
All frames respond in the language of the question (auto-detected). Force a
single locale with HENGE_LOCALE=en or HENGE_LOCALE=es.
| Stage | Lab | Model | Why |
|---|---|---|---|
| Scoping (base) | Anthropic | Haiku 4.5 | fast, cheap, ~3–5 s per call |
| Scoping (adversarial) | OpenAI | gpt-5 | cross-lab — challenges hidden assumptions |
| Meta-frame audit | OpenAI | gpt-5 | classify question; short-circuit if exploration/proxy |
| Canonical context | Anthropic | Opus 4.7 | tight summary of user answers, flag inconsistencies |
| 9 cognitive frames | mixed | gpt-5 ×6 / Sonnet ×2 / Opus | quality reasoning in parallel, cross-lab spread |
| Consensus synthesis | Anthropic | Haiku 4.5 | summarisation, structured output |
| Tenth-man — blind | Anthropic | Opus 4.7 | hardest reasoning, no view of the 9 |
| Tenth-man — informed | OpenAI | gpt-5 | cross-lab reconciliation, hallucination filter |
| Claim extraction | Anthropic | Sonnet 4.6 | falsifiable claim list from consensus |
| Claim verification | OpenAI | gpt-5 | rate each claim against the 9 frame outputs |
| Embeddings | OpenAI | text-embedding-3-large | ~15–25% better Spanish recall than -small |
Cost per run is computed from real usage returned by the SDK and persisted
under cost_breakdown in every report.json, split by lab
(anthropic_usd / openai_usd / embedding_usd) and by phase. A
representative full v0.6 run lands at roughly USD 1.00–1.50, ≈50%
Anthropic / ≈50% OpenAI. Pricing version is recorded against the report
(currently 2026-05).
- Python ≥3.11. macOS still ships Python 3.9. The Claude Code paste prompt detects this and installs Python 3.11.9 via pyenv automatically (no admin/sudo, but the build takes ~10 min the first time).
- Two API keys, both mandatory.
ANTHROPIC_API_KEY— Haiku 4.5, Sonnet 4.6, Opus 4.7 (3 frames + blind tenth-man + canonical + consensus + claims-extract).OPENAI_API_KEY— gpt-5 (6 frames + meta-frame + adversarial scoping + informed tenth-man + claim verification) andtext-embedding-3-large. v0.6 is cross-lab by design; both keys are required at boot. The startup validator pings gpt-5 and embeddings up front so you fail loudly, not 60s into a/decidecall.
- Restart Claude Code fully after install. Close ALL terminals running
claude, then reopen. The MCP catalog is loaded once at startup; a running session will never pick up a freshly-registered server.
v0.6 routes work across two labs (Anthropic + OpenAI) by design:
- Synthesis stays in Anthropic — scoping (Haiku), canonical context (Opus), consensus (Haiku), claims extraction (Sonnet), tenth-man blind (Opus). Same-lab consistency where structure matters.
- Audit crosses to OpenAI — adversarial scoping, meta-frame, tenth-man informed, claim verification. Cross-lab specifically catches the case where the synth lab hallucinates: gpt-5 has no output-style affinity with Haiku/Sonnet and surfaces orphan claims.
The full mapping is in the Models & costs table above and in
henge/config/frame_assignment.py.
Add OPENAI_API_KEY to .env. That's the only required action.
Everything else is automatic — the startup validator pings gpt-5 and
embeddings up front, so a missing-access account fails loudly at boot
instead of 60s into a /decide call. Schema bumps to 0.6; the legacy
henge.pricing.total_cost lookup is kept for back-compat through v0.7.
Every HTML report ships with two subtle highlighter-style markers and a toggle pill (bottom-left) to switch them on/off:
- 🟢 green (conclusion) — first paragraph after a
Conclusión / Inclinación neta / Recomendación / Veredicto / Síntesis / Takeawayheading. What to believe. - 🔵 cyan (action) —
<strong>blocks and bullets opening with imperative verbs (Priorizar / Asignar / Empaquetar / Posponer / Resistir / Embeber / Decisión / Segmento / Asignación / Secuencia, plus English equivalents). What to do.
Conservative heuristic — passages without an explicit conclusion heading or imperative opener stay unmarked.
Henge is a structural tool with a pre-registered metric. It is not a validated decision-quality instrument. Specifically: embedding distance is not propositional disagreement, all ten advisors share an Anthropic backbone, and verdict thresholds are pre-registered defaults rather than calibrated against ground truth.
The full list lives in LIMITS.md. Read it before you trust the output for anything load-bearing.
- founder & operator decisions
- hiring / scaling / firing
- product strategy and prioritisation
- risk analysis & pre-mortems
- counterfactual reasoning
- AI agent orchestration where you need a structured second opinion
Tracked in the issues page.
v0.6 (current). Cross-lab multi-model — 6× gpt-5 + 2× Sonnet + 1× Opus
across the 9 frames; adversarial scoping; canonical context; meta-frame
audit with early-exit; dual tenth-man (blind + informed); claim
verification; honest cost accounting split by lab; K-runs distribution
mode for CFI with temperature > 0; takeaway markers in the HTML report.
v0.7 (planned).
- Henge-50 outcome benchmark (50 historical decisions with known outcomes) — the validity claim that v0.5 promised and v0.6 still owes.
cross_lab_agreementanddelta_signalmetrics on the tenth-man pair (high/medium/low).- Inline claim annotation in the consensus body (vs. a separate panel).
--force-full-runMCP flag to bypass meta-framereformulate.- Add Gemini to the frame pool; remove the Voyage embeddings path if no demand surfaces.
v0.x. Local embeddings (sentence-transformers, no API key), PDF / shareable web report, streaming results, adaptive frame selection.
Why Henge exists, the role/method split, and why steel-man over devil's advocate live in MANIFESTO.md.
The methodology paper is WHITEPAPER.md.
The reproducible protocol is METHODOLOGY.md.
For Tool API, output structure, manual installs (Claude Desktop / Cursor), embeddings provider config, architecture, and troubleshooting, see DEVELOPER.md.
MIT