Evidence-based decision intelligence for consequential choices.

The problem

Asked for a recommendation, a model tends to produce one confident answer — plausible, well-argued, and unmarked as to which parts are fact, which are inference, and which are simply the most common answer in its training data. For a genuinely consequential choice, that is the wrong shape of output: the decision-maker needs to see the alternatives that were considered and rejected, the evidence behind each claim, and where real uncertainty remains, not a single polished paragraph.

Why not just ask for it in the prompt

A good prompt can ask for "pros and cons" once. It does not reliably enforce that every number traces to a source, that a contradiction between two claims gets surfaced rather than silently dropped, or that the strongest counter-argument gets a fair hearing before the decision locks. Those are process guarantees, not phrasing — which is why they are encoded as constitutional rules the system cannot skip, rather than left to whichever way a particular prompt happened to be worded that day.

How it works

• —Routes every request through a triage: most decisions run at a fast, default depth; only genuine conflicts, safety consequences or high-stakes reversals escalate to a deeper, more adversarial review.
• —Follows one workflow: frame the decision, set out alternatives, gather evidence, build an evidence graph and claim ledger, run independent first passes, test scenarios and sensitivity, actively try to falsify the leading option, then lock the facts before any editorial pass.
• —Stops gathering evidence when the next piece of research is worth less than its cost and the time it takes — not when the analysis merely feels complete.

Evidence and provenance

• —Classifies every material claim and number as user-provided, externally sourced, derived, board-proposed, a hypothesis, or unknown — and blocks publication on unsupported claims.
• —Treats multiple articles that trace back to one original source as one source, not independent confirmation.
• —Uses qualitative confidence with stated reasons rather than an invented numeric percentage.

What it produces

• —A concise founder or executive brief by default, answering: what is recommended, why, what to do now, what not to do yet, what is actually known, what remains unknown, the strongest case against it, and what would change the decision.
• —A machine-checkable integrity state covering claim provenance, arithmetic, source independence and legal-claim classification. A failed check prints, literally, "BOARD DRAFT — PUBLICATION BLOCKED."

What it must not do

• —Never fabricates sources, calculations, independence or agent-run evidence.
• —Never lets editorial polish change a fact, number, evidence grade or safeguard once the facts are locked.
• —Treats a correct "unknown", "experiment first", "defer" or "provisional" as a valid, often preferred, outcome — not a failure to reach an answer.

In practice

A pricing decision with conflicting signals: internal data says one thing, a competitor benchmark says another. Rather than picking a number, DecisionOS classifies each figure by source, checks whether the benchmark is really independent evidence or three articles citing the same original report, and, if the conflict cannot be resolved with available evidence, returns a "provisional" recommendation with the specific test that would resolve it, rather than a false-confident single number.

Why a reusable skill, not a one-off prompt

A decision framework applied once in a conversation is forgotten by the next one. Packaging it as a skill means the same 21 constitutional rules, the same escalation triggers, and the same integrity gates apply every time a consequential decision comes up (across different conversations, different models, different sessions) rather than depending on whether that day's prompt happened to ask for rigour.

Mohammed's contribution

Designed the constitutional rule set and the escalation model (rapid / deep / high-stakes) that decides how much process a given decision earns, and defined the machine-checkable integrity gates a draft has to pass before it can be called final.