Companion to Architecture.md. Written for the general reader. Nothing here requires a background in analysis, and nothing in the engine requires Anthropic's products: the engine is a written method, not software, and this document doubles as a recipe for running it on any capable language model.
Modern language models are already good at analysis. Give a strong model a hard problem and it will produce a fluent, plausible, often insightful answer. The trouble is what a fluent answer does not show: the claim it did not check, the source that benefits from being believed, the question that cannot be answered by evidence at all but gets an answer anyway, and the human judgement that quietly gets made for you instead of by you.
The Insight Engine adds no intelligence to the model. It wraps the model in discipline: check what can be checked, say plainly how strongly each claim is established, refuse to let the model predict what cannot be known, and force the human to take the judgements that belong to a human. Every part of the engine exists to serve one of those four jobs. This design philosophy has a documented basis: individuals — human or model — do not reliably correct their own biases when warned about them, but procedures imposed from outside do change outcomes (a finding associated with the psychologist Daniel Kahneman, and confirmed in the engine's own testing, where instruction-level "be careful" additions repeatedly failed while structural procedures survived).
Each part below gives: what it does, where the idea comes from, and the essence you would carry to any other model. A compact table mapping every layer to its runtime step, approach, and source is in Architecture.md, section 3.1.
The proportionality triage (Step 1; before anything else runs). The engine first asks whether the problem is genuinely trivial — routine, reversible, binding nobody else, no money or legal exposure, no contested values. Only then may it give a short answer instead of the full method; under any doubt it runs everything. The idea is satisficing — the insight, from decision research, that for many problems "good enough, cheaply" beats "exhaustive, expensively", and that heavy analysis of trivial questions is itself a failure. The design is fail-safe in one direction only: wrongly down-scoping a serious problem costs far more than mildly over-analysing a trivial one, so depth is the default. Its boundary conditions match the naturalistic decision-making tradition (Gary Klein): formal analysis pays off where experience is limited and stakes are high — which is exactly where this engine is pointed. Replication essence: begin every run with the written triviality test, defaulting to depth.
TOSCA scoping (Step 2). Before analysing, the engine offers five plain questions: what is the trouble (described as a symptom, not a diagnosis, plus "why now?"); who owns the problem and who will read the answer; what counts as success and as a serious cost; what is fixed and what is in or out of scope; and what documents the operator holds that the internet cannot verify. This is the TOSCA problem-definition frame — Trouble, Owner, Success criteria, Constraints, Actors, leading to one stated core question — from Cracked It! by Bernard Garrette, Corey Phelps and Olivier Sibony, the leading text on consulting-style problem framing, with the Actors element adapted into the "documents you hold" question. Two details matter: the questions are always optional (a skipped question never blocks the analysis), and if you skip them the engine must write its own one-line problem statement — trouble, owner, core question — and show it to you marked as an assumption, because an unstated problem is where analyses silently go wrong. Replication essence: five optional scoping questions in this shape; on skip, the model states its own problem definition visibly.
The provocation page (Step 3). A fixed, one-page set of prompts the model applies to every problem identically: for every apparent gain, ask where the cost went; mark every load-bearing claim as independently corroborated or not; ask who bears a cost but is not in the room; ask who benefits from the question being framed this way; check the base rates and whether each metric measures what it claims; ask what has become invisible through repetition; ask what is loud but undeserving of attention; generate — freshly, from the model's own knowledge, never from a stored checklist — what an expert treatment of this kind of problem would include, and flag what is missing; and ask what the way the problem was handed over makes the analyst likely to miss. The page is the distillation of a 234-framework analytical corpus built and then deliberately retired by this project, after controlled testing showed a strong model plus this one page matched the full apparatus. Its form follows the checklist-design principles Atul Gawande documented in surgery and aviation: short, judgement-activating rather than judgement-replacing, with revisions only through versioned, evidence-based change. Replication essence: copy the page verbatim (it is in the repository); apply it word for word; never let it grow per-problem.
The provocation page, probe by probe. Each line of the page counters one documented way analyses go wrong. Displacement / second-order (always on) counters the oldest accounting failure in complex systems: a measured win whose cost quietly moved to an unmeasured place, an absent party, or the future — the probe forces "where did the cost go?" every time something looks like a gain. Provenance / verification (always on) counters taking interested parties at their word: every load-bearing claim is marked corroborated or not, the single most conclusion-changing unknown is named, and a self-damaging admission outweighs a self-serving claim, because parties rarely lie against their own interest. Absent voice counters the quiet narrowing of the stakeholder set to whoever is in the room. Cui bono from the framing counters inherited problem definitions — the question as handed over often serves someone. Base rate and measurement counters two failures at once: ignoring how often situations like this usually turn out, and trusting a metric that has become a target and is now a gameable proxy optimised at the goal's expense. Normalised-as-natural counters status-quo blindness — a condition repeated long enough reads as a fact of nature rather than a choice someone keeps making. Spuriously present counters salience: the precise number, reassuring narrative, or performed rigour that pulls more attention than it has earned. Conspicuously absent — the omission sweep — counters the frame's omissions: the model must generate, freshly each time from its own knowledge, what an expert treatment of this kind of problem would contain, then flag what is missing; freshly, because a stored checklist would itself limit what the sweep can find. Reflexive counters the handover itself: what does the way this problem arrived make the analyst likely to miss? The three closing rules — keep only findings that survive their best opposing reading, name every materially affected party, and prefer the material set over length — counter, in order, confirmation bias, stakeholder omission, and the padding that lets an over-claim cost more than a miss.
Verification with graded evidence (Step 4). Every claim the conclusion rests on is sorted three ways: checkable on the public record (search, then grade it); held privately by a party (never pretend to verify it — name the document that would settle it); or unanswerable by evidence at all (route it to the human). Verified claims carry a strength tier: confirmed against the primary source itself (primary meaning close to the event, not automatically trustworthy — an interested party's own record is primary, but is graded on whether anything independent corroborates it); confirmed against reputable secondary reporting; or only weakly corroborated or actively disputed. Then the engine red-teams itself: for each critical claim it runs a second search for the counter-evidence, and downgrades the claim if credible contradiction exists. Finally, the attribution ceiling: when any official account says a named person failed or erred, that is recorded as the account's judgement — a real fact about what the body found, never a settled fact about the person, because post-incident blame is a social judgement (a principle from the complex-systems safety literature of Richard Cook and David Woods, reinforced by Baruch Fischhoff's finding that hindsight bias survives being warned about). The whole grading vocabulary is deliberately aligned with ICD-203, the published analytic tradecraft standard the United States intelligence community holds its analysts to. Finally, every graded claim in the finished document carries a reference the reader can check without help: web-verified evidence names its source with a URL; evidence from inside the provided documents names the document and the section or page. An analysis the reader cannot audit is only confident-sounding, not defensible. Replication essence: the three-way sort, the tiered grades, the mandatory counter-evidence search, the rule that official blame is never imported as fact, and a checkable reference on every graded claim.
The systems pass (Step 5, conditional). Some problems are systems: feedback loops, accumulating pressures, delays between cause and effect. Only when a problem shows at least two such signs does the engine build a small causal map — named variables, signed links, reinforcing and balancing loops, tipping points, ranked leverage points — in the tradition of systems thinking. Two disciplines make the map safe, and both came from adversarial testing rather than theory: every loop is graded twice — does it exist (often verifiable) and does it currently dominate (never graded above [U] by design, and routed to the human) — each leverage point is graded twice as well — the mechanism, and whether acting there will actually work, the latter also never above [U] by design — and tipping points are stated as conditions to watch, never dates or probabilities, because testing showed that even strong models will manufacture confident forecasts under pressure unless the prediction ceiling is written into the method. The map investigates; it never predicts. Replication essence: the two-of-six trigger, the small-map caps, the exists-versus-dominates double grade, and the hard ban on forecast language.
Deep-core routing and the gate (Steps 6 and 8). Some questions have no evidence-based answer: how the decision should be framed, who is bound by it without a say, what is being given up, which value outranks which. The engine separates these into an explicit open list — the fact/value separation that tradecraft standards require — and then does something unusual: it stops, and puts each question to the human, one at a time, in plain words, with an example answer, until the human records a position or explicitly defers. This is forced because testing showed that merely displaying such questions does not work — readers, human and machine, drift to the checkable items and leave the value questions untouched. Forcing an answer is a cognitive-forcing device, the same family of remedy medicine uses against diagnostic tunnel vision. The gate opens by asking for the operator's own standpoint — the analysis has audited the problem; this audits the analyst's reader, an idea owed to The Blind Spot (Frank, Gleiser and Thompson): never leave out the person doing the knowing. Replication essence: an OPEN list the model may not resolve, and a mandatory human checkpoint before the final document — in a chat interface, simply ask the questions one at a time and wait.
Assembly (Step 9). The final document leads with the human's recorded positions, then the graded findings, then a decision brief in five fixed parts: the call; the confidence basis and what changes it; the single dominant unknown; the cheapest next verification; and what would flip the conclusion. If the problem is wicked — the term Horst Rittel and Melvin Webber coined in 1973 for problems with no definitive formulation, no stopping rule, and no true-or-false answers — the call is issued as a provisional stance, never a verdict: it names the value it trades away, and names what would re-open it. Before finalising, the engine re-scans once for material it deprioritised while concentrating on the load-bearing few (focus narrows attention; the re-scan is the counter-measure), stress-tests the conclusion by varying its main assumptions together, and — where the evidence shows an outcome is decided by power rather than merits — says so plainly instead of recommending accommodation the powerful party has no need to grant. Every full document ends with a fixed plain-language legend explaining the grades, so no reader needs training to use it — based on the research on how novices actually read expert documents (John Hayes; Bransford and Stein). Replication essence: the fixed five-part brief, the wicked-call guard, the joint assumption stress-test, the closing legend.
Rendering and tracking (Steps 11 and 12; standalone skills). A finished analysis can be re-voiced for any named reader — a board, a regulator, a family member, an opponent — under one inviolable rule: grades and claims never change, only register. That is the "sell" half of problem-solving (again Cracked It!), disciplined by the grade-lock. And an analysis can be kept alive over weeks: a single dossier holding the graded spine, a watch-list of what would change the call, and a dated log where every grade change is explicit — the tradecraft standard of showing your judgement change rather than silently updating. Replication essence: re-voicing may touch tone only; a living analysis is one file with an explicit change log.
Independent adjudication (Step 10 — opt-in). For the highest-stakes results, the engine offers a second pass: a different model, told the facts and the graded spine but not the call and not the first model's reasoning, re-derives the analysis blind; only then is it shown the call, and it attacks it; and its disagreements are folded back in without ever overturning the human's call. The idea is old and well-evidenced — devil's advocacy, red-teaming, and independent review — with one modern addition: the check is only as good as how differently the second model thinks, so the strongest form uses a model from a different lab entirely (a different training lineage has the most decorrelated blind spots), stepping down to a different in-house model, then to a same-model cross-check, and the engine declares which rung it managed. It is off by default, asked as a one-word-declinable question only on a finished high-stakes analysis, honest that its ability to catch a missed error is unproven, and — for the cross-lab rung — held back by default on privileged or confidential material that would leave the boundary, released only by a deliberate, logged act of the operator's. Replication essence: hand a fresh, independent model the facts and the graded spine — never the call, never the first model's reasoning; have it re-derive blind; only once its verdict is in, reveal the brief and have it attack; fold the disagreements back under grade-lock; keep the call with the human; ask first, always.
Five invariants run through every layer, and they are the actual product: (1) the grade-lock — once evidence grades are fixed, nothing downstream may alter them, only re-express them; (2) prediction ceilings — what cannot be known is marked unknowable and routed to the human, never estimated to please; (3) offer, never block — every question to the operator is skippable, and skipping never stops the analysis, it only makes the assumptions explicit; (4) the human owns the values — the engine articulates value questions and refuses to answer them; (5) procedures, not exhortations — every rule lives as a written step the model must walk through, because "try harder to be unbiased" measurably does nothing.
The engine's analytical substance is a handful of plain-text files in this repository — the five skill instructions (analyse, verify, render, track, adjudicate) and the provocation page — released under the MIT licence. (The adjudicate skill also ships a few small Python helpers, for the operator's switchable settings and opt-out, making the cross-lab call, and logging runs; these are conveniences around a method that is still just written text — in plain chat you ask the opt-in question yourself and run the second model as another session.) To run it on any other capable model:
- You need: a strong general model; a web-search or browsing tool it can use; the ability to give it a long instruction; and a human willing to answer a handful of questions mid-run. Nothing else.
- Set-up: paste the
analyseskill text (with the provocation page inserted where it is referenced) as the system instruction or opening message, followed by your problem. - Run: the model works the steps in order. Where the text says to ask the operator, it asks you in the chat; you answer or say "defer". Where it says to search, it uses whatever search tool the platform provides.
- What does not carry over: the packaging only — automatic skill loading, form-style question widgets, scheduled re-verification, and the adjudication skill's opt-out memory, cross-lab harness, and run log. All are conveniences; asking in plain chat, re-running
verifyby hand, and running a second model as a fresh session reproduce them. - One finding worth knowing: in this project's controlled tests, the discipline added more on a strong-but-lesser model than on the strongest one tested — the strongest models already self-supply some of the checking, so the written procedure matters more, not less, as substrate strength falls. The method is, if anything, more valuable the further you move from the strongest models.
- What to expect: the engine makes analyses slower, longer, and more honest. It does not make the model smarter, and it does not remove the need for you: its most tested single feature is the moment it stops and hands the judgement to a human.
Rittel & Webber, "Dilemmas in a General Theory of Planning" (1973) — wicked problems. ICD-203, Analytic Standards, Office of the Director of National Intelligence (2015) — the analytic tradecraft standard this engine is built around. Garrette, Phelps & Sibony, Cracked It! (2018) — TOSCA framing and audience-fit. Cook & Woods, complex-systems safety papers (1994-2005) — attribution as social judgement; the blunt end behind every sharp end. Fischhoff — hindsight bias survives warnings. Kahneman, Thinking, Fast and Slow (2011) — procedures over self-correction. Klein, Sources of Power (1998) — when formal analysis pays off. Gawande, The Checklist Manifesto (2009) — checklist form and revision discipline. Simon — satisficing. Senge and the systems-thinking tradition — causal loops (used as method, with archetype libraries deliberately declined). Frank, Gleiser & Thompson, The Blind Spot (2024) — audit the knower. Hayes; Bransford & Stein — how non-specialists actually read and decide. Madhavan, Wicked Problems (2024) — the hard/soft/messy finding tags. Tang, "Complex, Messy, Wicked, Systems: Management Principles and Tactics" (Procedia Computer Science, 2025) — a complex-to-messy-to-wicked typology and a satisficing-first management doctrine that independently matches the engine's triage, stakeholder, and risk-versus-uncertainty disciplines. And the engine's own predecessor: a 234-framework analytical corpus (the "SAF"), built, tested against its own distillation, and retired; its tested content survives as the one page.
Every design choice above was also tested: the Architecture document records, feature by feature, the controlled comparisons behind what shipped — and the tested candidates that did not.