v1.0.0 · A tool for investigating high-stakes, complex, contested, and wicked problems — the kind where the important things are easy to miss, the evidence is mixed or interested, and the deciding questions are value judgements. The engine surfaces the hidden contributors, conditions, and causal structure beneath a situation — sources of success and adaptation as well as failure — verifies every load-bearing claim against current sources, and routes what cannot be verified to your judgement. It offers understanding and a defensible next step, not control or prediction: wicked problems are navigated, not solved. The large language model provides the insight; the engine makes it verified, defensible, and audience-ready.
The v1.0.0 above and the same number in the closing line are the version number of this guide, not of the plugin; do not change them when the plugin's version changes.
This is the Markdown edition, for reading on GitHub. The original styled version (with the rich diagrams) is in
originals/and on the project site.
Use the Insight Engine when a decision matters and the surface is not the whole story — a strategy call, a post-incident review, a contract or regulatory question, an investment, a contested situation where you need to know what is really going on and how strongly you actually know it.
It is not a search box or a chatbot. It is a discipline: verify what can be verified, grade how strongly each claim is established, separate fact from value, and make you own the value calls. What you get back is a graded, defensible analysis and a decision brief — and, on request, the same analysis re-voiced for a particular reader, or kept live as an investigation.
You don't need it for genuinely trivial things. It will notice triviality and right-size itself to a short answer — but you get the most from it on problems that actually have hidden depth.
The engine is the insight-engine plugin. To update, re-install the latest .plugin from Customize → Skills. You drive it with five chat commands:
/analyse— The main one. Hand it a problem in your own words, or attach a document and say "analyse this."/verify— Fact-check a set of claims or a premise before you rely on it./render— Re-voice a finished analysis for a specific reader./track— Keep an investigation live across sessions as new facts arrive./adjudicate— Put a finished high-stakes analysis to an opt-in independent second pass (blind, then adversarial) on the most decorrelated model available. It never changes the call.
Simplest use: type /analyse followed by your problem, answer the gate when it asks, and read the brief.
Say you run:
/analyseOur 40-person firm is about to sign a 5-year exclusive deal with one software vendor. The discount is 30%. Worth it?
What happens, in order:
- Triage — not trivial (a 5-year lock-in, material money), so the full pass runs.
- Scoping (optional) — up to five short questions: the trouble and why now; whose problem it is and who will read the answer; what a good outcome and a serious cost look like; what is fixed and in scope; any documents only you hold (the draft contract). Skip any or all — skipped items become stated assumptions, and if you skip scoping entirely the engine states its own one-line problem statement (trouble, owner, core question) as an assumption before proceeding.
- Findings arrive graded — for example: the vendor's "guaranteed 99.9% uptime" rests on its own marketing
[U]; a 30% discount is roughly the market norm for 5-year exclusivity[V2], so it is payment for the lock-in, not a favour; exit costs sit in the contract you hold — party-held: the draft's termination schedule would settle it. - The gate — before the brief is finalised, the engine asks you, one question at a time, things like: "This framing treats the decision as sign / don't sign. Is renegotiating scope or term a live option?" (example answer: "Yes — a 3-year non-exclusive at 20% is acceptable"), with "I don't know — explain this" and defer always available. Your one-line positions go on the record at the top of the brief.
- The decision brief — the call (provisional if values dominate), the confidence basis, the dominant unknown (here: the termination schedule), what to verify first (read that schedule), what would flip the call — then sources and the plain-language legend.
Nothing in the run predicts the future or makes the value calls for you; that part is deliberately yours.
The engine is a stack of layers (L0–L6) held together by an invariant grade-lock spine and owned, at one critical point, by you. A fail-safe proportionality triage sits in front of it and decides whether the full machinery should run at all.
flowchart TD
TR["fail-safe PROPORTIONALITY TRIAGE<br/>unambiguously trivial? a short answer · under any doubt, full depth"]
TR --> L1["L1 · Provocation page<br/>displacement & 2nd-order · provenance [V]/[U] · seven probes · omission sweep"]
L1 --> L2["L2 · Verification<br/>three-bucket router · source tiers [V1]/[V2]/[V3] · disconfirmation · coverage-retention"]
L2 --> L25["L2.5 · Systems pass — conditional<br/>feedback-driven problems only · graded causal map:<br/>loops · tipping conditions · leverage · investigate, don't predict"]
L25 --> L3["L3 · Deep-core routing<br/>framing · legitimacy · opportunity-cost · value → the un-verifiable OPEN list"]
L3 --> GATE["★ Forced judgement gate — YOU<br/>you record a one-line position on each deep-core question, or defer"]
GATE --> L4["L4 · Assembly — the decision brief<br/>call · confidence basis · dominant unknown · what-to-verify-first · what-would-flip-it"]
L4 --> ADJ["Adjudication (Step 10) · opt-in, off by default<br/>independent second pass: blind → adversarial · never flips the call · YOU opt in"]
ADJ --> L5["L5 · Rendering<br/>audience re-voicing: adversarial · advocate · board · counsel · regulator · family · any"]
L5 --> L6["L6 · Workspace<br/>living dossier across sessions: OPEN · UPDATE · STATUS"]
L0["L0 · Strong model — the substrate that provides the insight"] -.->|substrate| L1
SPINE["GRADE-LOCK SPINE — the invariant through every layer<br/>[V1] primary · [V2] secondary · [V3] weak/contested · [U] unverified<br/>the single dominant unknown · the non-droppable caveat core<br/>grades are locked: set once, never changed downstream — only re-expressed"]
SPINE -.->|runs through every layer| L2
classDef gate fill:#fdf1df,stroke:#e8a33d,stroke-width:2px;
classDef cond fill:#fff7ec,stroke:#eac489,stroke-width:2px;
classDef green fill:#e9f5ee,stroke:#a9d4bd;
class GATE gate;
class L25 cond;
class ADJ cond;
class L2 green;
The five skills (the interface): /analyse orchestrates the whole pipeline (triage → L1–L6 with the gate); /verify is L2 standalone; /render is L5 standalone; /track is L6 standalone; /adjudicate is the opt-in Step 10 adjudication layer standalone.
In short: L0 the model is the substrate; L1–L6 are the processing layers; the spine is the invariant that never changes; the gate is where you act; the triage decides how much machinery to run.
Here is what actually happens when you hand it a problem. Most of the pipeline runs silently — the one place it stops for you is the gate.
flowchart TD
U["You · /analyse your problem"] --> T{"⚑ STEP 1 · Proportionality triage<br/>trivial, self-contained, reversible,<br/>no third parties or legal exposure?"}
T -->|"YES — rare"| SH["A short, proportionate answer<br/>+ an offer to run the full pass"]
T -->|"NO / any doubt → full depth (the default)"| S0["STEP 2 · Optional scoping<br/>a few skippable questions"]
S0 --> S1["STEP 3 · Provocation pass<br/>surface what's material but sits where attention doesn't go; mark each claim [V] or [U]"]
S1 --> S2["STEP 4 · Verification<br/>route each claim · tier it V1/V2/V3 · search the negation of the load-bearing few"]
S2 --> S25["STEP 5 · Systems-investigation pass — conditional<br/>only if genuinely systemic (feedback · stocks · delays · tipping):<br/>a graded causal map of loops, tipping conditions, leverage · investigate, don't predict"]
S25 --> S3["STEP 6 · Deep-core routing<br/>set aside the questions no evidence can settle: framing · legitimacy · opportunity-cost · value"]
S3 --> S4["STEP 7 · Grade-lock + resilience check<br/>fix every grade; stress-test whether the call survives its assumptions moving together"]
S4 --> G["★ STEP 8 · THE FORCED GATE — YOU<br/>take a one-line position on each deep-core question, or explicitly defer.<br/>The brief does not finalise until you do."]
G --> S6["STEP 9 · Assemble → the decision brief<br/>coverage-retention re-scan · your positions · graded findings · call · confidence basis · dominant unknown · verify-first · flip-it · sources"]
S6 --> S10["STEP 10 · Offer independent adjudication<br/>opt-in, off by default · high-stakes only · blind → adversarial · never flips the call"]
S10 --> S11["STEP 11 · Offer /render"]
S11 --> S12["STEP 12 · Offer /track"]
classDef gate fill:#fdf1df,stroke:#e8a33d,stroke-width:2px;
classDef cond fill:#fff7ec,stroke:#eac489,stroke-width:2px;
classDef green fill:#e9f5ee,stroke:#a9d4bd;
class G gate;
class T cond;
class S25 cond;
class S10 cond;
class S2 green;
class S4 green;
From Step 7 onward the grade-lock spine holds: grades never change — they are only re-expressed.
Step 5 (added in v0.1.7, numbered Step 2.5 until v0.1.16) is conditional. When a problem is genuinely systemic — driven by feedback loops, accumulating pressures, delays, or a possible tipping point — the engine adds a systems-investigation pass: a graded causal map of the loops, the conditions under which the system could tip (never a date or a probability), and the highest-leverage places to intervene. It investigates the structure and routes the "will it / when" judgement to you; it never forecasts. On a static, one-shot problem it skips this step and says so.
What to type: /analyse and your problem in plain words, the specific questions you want answered, or a document to analyse — it will answer your questions inside the graded analysis. The more raw the problem (rather than a pre-digested summary), the more the engine can surface what you might have framed out.
What it does: runs the pipeline above. You may see it search the web — that is verification, grading each load-bearing claim and searching for evidence that would disprove the ones the conclusion rests on.
What you get: a written analysis led by your gate positions, the graded findings (each marked and tiered), and a five-part decision brief.
Part-way through, the engine stops and puts two to four deep-core questions to you, one at a time — about the framing, the legitimacy, the opportunity cost, or the values at stake. These have no verifiable answer; they are judgements only you can make. You can:
- Take a position — a single line on each is enough.
- Override the engine's framing — encouraged. If you disagree with how it has framed the problem, say so; it will reshape the brief around your judgement.
- Ask for help — every question arrives with a one-line example answer; choose "I don't know — explain this" and the engine explains it in plain language, then asks again.
- Defer — logged, not ignored; the brief notes it as open.
The gate opens by asking for your own standpoint in one line. Where the systems pass ran, it also asks at what level to intervene: this problem, or the problem it is a symptom of.
Why it's forced. The engine learned, by test, that if these questions are merely shown to you they get read and then ignored — attention shifts to the verifiable. Being made to take a position is what gets the value questions actually addressed. In real use, an operator's override at the gate materially reshaped the final brief. This is the engine's most valuable interaction; treat it as the decision, not a form to clear.
Every load-bearing claim is marked. [V] = verified, independently corroborated; [U] = unverified, meaning not independently verified (five situations, set out below). A [V] always carries a strength tier: [V1] primary (the regulation, the official statistic, the study), [V2] secondary (reputable reporting on top of primaries), [V3] weak or contested — the primary/secondary distinction marking proximity of the source, not trust in it ([V3] is a judgement about strength, not proximity): an interested party's own document does not earn [V1] on a claim it has a stake in, and a claim resting only on that party's own account is [U], not [V3]. An [U] does not tell you why the claim is unverified — the claim's own line does. It may be uncorroborated — or contradicted — after a real search; rest only on an interested or self-reporting party; be waiting on a named document that would settle it; inherit the cap from an unverified thing it depends on; or be a judgement or prediction the engine never grades higher, however strong the evidence — which loop dominates, whether acting at a leverage point will work, and whether or when a system tips are permanently [U] by design. Only the first three can be moved by looking harder; the fourth moves only when the unverified thing it depends on is itself verified. The last kind is not a gap in the research; it is the engine declining to pretend.
A related point where causal maps are involved: a tier on a map link grades the support for that mechanism in general — not that this particular arrow is the operative cause in your case. The same symbol carries a weaker promise there than it does on a claim.
Map elements and the analysis's numbered items carry fixed identifiers whose shape — a capitalised word stem plus a number: Node1, Link1, LoopR1, ClaimReg1, Find1 — cannot be confused with a grade symbol, a domain acronym, or a layer tag. A bracketed [V1] is always a grade; Node1 is always a map variable.
The [V3] and [U] items are where the analysis is exposed — that is where to direct further verification. One further rule protects you: an official finding of personal fault (a coroner's or inquiry's "X failed") is reported as that body's recorded judgement, never as established fact — blame is always yours to judge at the gate. Every full-pass deliverable also ends with a fixed plain-language legend restating these markers in one sentence each — you never need this guide open to read a brief.
- The call — the actual recommendation.
- Confidence basis — confidence in the fact-chain and the conditions under which it changes (it names the weakest grade the call leans on), plus, separately, the action contingency: what must hold true after acting for the call to succeed. Evidence-quality transparency, not a track-record calibration.
- The dominant unknown — the one fact that would most change the conclusion. Verify this first.
- What to verify first — the cheapest next step that most reduces risk.
- What would flip the call — including any assumptions that, moved together, the call could not survive.
A one-line reliability caveat naming what the analysis rests on; and, for a wicked problem with no verifiable solution, the "call" arrives as a provisional stance — the value it trades and what would re-open it — rather than a false-confident verdict. That is deliberate.
Findings in a brief may carry an optional tag borrowed from systems engineering (Madhavan, 2024), naming what kind of component each finding is and what a response can honestly aim at:
- hard — bounded and optimisable; the matched verb is solve.
- soft — behavioural or political; there is no optimum, only a defensible accommodation; the verb is resolve (by satisficing).
- messy — rooted in conflicting values or beliefs; the verb is dissolve (reframe the situation until the problem cannot arise in that form).
A wicked problem is usually all three colliding, which is why a single verdict cannot fit every finding. Tags are reporting vocabulary only — they change no grade and add no pipeline step.
/verify— Hand it a set of claims or a premise. It routes each (web-checkable / your-document / value-judgement), grades and tiers them, searches for counter-evidence, and tells you which are exposed. Use it to check something before you rely on it./render— Takes a finished analysis and re-voices it for a reader. It will ask who — or you name them: a board, counsel, a regulator, a journalist, the family, an adversary. The grades never change; only the register does./track— For inquiries that evolve. OPEN starts a dossier; UPDATE folds in new facts and re-verifies; STATUS reads back where it stands. (Newest skill — sound, but the least tested in real use.)/adjudicate— Put a finished high-stakes analysis to an independent second pass: a different-lab model — you choose the lab (OpenAI GPT-5.6 Sol, Google Gemini, or xAI Grok, or any OpenAI-compatible endpoint under "Other"; a Claude-lineage model is refused, since a same-lineage check is not cross-lab) — re-derives it blind, then attacks it, and its disagreements fold back in — it never changes a claim, a grade, or the call. It is opt-in and off by default; the engine offers it in one line you can decline, it costs credits or set-up, and "don't offer again" is remembered by the plugin. For privileged or confidential material the cross-lab pass is blocked by default — it would leave the boundary — and lifts only when you deliberately override it, with the exact package shown for confirm / edit / cancel and the override logged. The pass runs as two calls, not one: the first carries the graded spine and not the call, so the blind pass is blind by what was sent rather than by instruction, and the brief is revealed only in the second. Every ask, gate, adjudicator rung, and panel size is a switchable setting, standing or per-run. Honest note: it adds defensibility and a decorrelated second look; whether it catches an error the first pass missed is unproven. (OpenAI is live-verified; the Gemini and xAI adapters are built to each lab's current API and live-verified only when you run the smoke with that lab's key.)- Chaining them — A natural flow:
/analysea problem → (optionally/adjudicatea high-stakes result) →/renderit for each stakeholder →/trackit as the situation develops./verifycan be used at any point to re-check a claim.
- Give it the raw problem, not a tidy summary — it verifies, and it surfaces what a tidy summary has already framed out.
- Upload the documents you hold (contracts, figures). It will not pretend to verify what it cannot see, but it will use what you give it and name what would settle the rest.
- Answer the gate honestly, and override it when you disagree. That is the point of the gate.
- Don't rewrite for a different reader — use
/render; the grades stay intact. - For something that will change, use
/trackrather than re-running from scratch. - Trust the tiers. A
[V1]is solid; a[V3]or[U]the call leans on is the weakest point — verify there first.
- It won't over-analyse a genuinely trivial decision — it right-sizes, so don't be surprised by a short answer (you can always ask for the full pass).
- It won't pseudo-verify a private document — it names what would settle it.
- It won't resolve your value questions — it forces you to.
- It won't change a grade to suit an audience.
- It won't convene real stakeholders — it names every affected party and can render their perspectives, but a simulated perspective is labelled as such and is not participation; the people bound by a wicked call still have to be brought in by you.
- It won't design interventions or trials — it tells you what to verify first and what would flip the call; staging, piloting, and reversible-trial design are routed to you and your specialists.
- It won't simulate quantitatively — the systems pass investigates and grades structure; when genuine modelling is needed it says so and routes out.
- It won't intervene in power — where evidence shows outcomes are power-determined, the brief says "power problem, not an analysis problem" instead of pretending analysis can settle it.
The terms needed to read a brief are below. The full corpus vocabulary — every term, symbol, identifier series and abbreviation, with the multiple-meaning audit — is in the Glossary.
| Term | Meaning |
|---|---|
[V] / [U] |
A claim is [V] (verified — independently corroborated) or [U] (unverified — not independently verified). That one mark covers five situations: uncorroborated or contradicted after a real search; resting only on an interested or self-reporting party; awaiting a named document that would settle it; inheriting the cap from an unverified thing it depends on; or permanently capped by design — which loop dominates now, whether acting at a leverage point will work, and whether or when the system tips. The first three move on further verification; the fourth moves only when the thing it depends on is verified; the fifth never moves. The claim's own line says which of the five it is, and records any contradiction found. Questions of value, framing or blame, and forecasts made outside the systems map, carry no mark at all: they are routed to you as open questions, never graded. |
[U] by design |
The permanent kind. Which loop dominates now, whether acting at a leverage point will work, and whether or when the system tips carry a hard ceiling of [U] and route to you. No further research lifts them: the engine is declining to manufacture confidence, not reporting a gap in its work. |
| Tier on a map link | On a causal-map link, [V1]/[V2]/[V3] grade the support for that mechanism in general — not that this arrow is the operative, correctly-directed cause in your case. A weaker promise than the same symbol on a claim. |
[V1] / [V2] / [V3] |
Source-strength tiers on a [V] — how strong the corroborating source itself is. The tier is part of the grade: it travels with the claim and is never dropped. |
[V1] — primary |
The strongest tier: the claim is confirmed against the authoritative source itself — the text of the regulation or statute, the court ruling, the official statistic or regulator publication, the peer-reviewed study, the primary dataset or standard. Nothing stands between the claim and its evidence. Note: primary means the source is proximate, not that its content is trustworthy — an interested party's own primary material does not by itself earn [V1] on a claim that party has a stake in; grade that claim on independent corroboration. |
[V2] — secondary |
Confirmed against reputable reporting or interpretation resting on primary sources — an established news organisation, a professional body, a textbook, a well-regarded analyst. Reliable, but one step removed; where it can, the engine names the primary source that would lift the claim to [V1]. |
[V3] — weak or contested |
The weakest [V]: the claim has only weak but independent corroboration (a single blog, a content aggregator, advocacy material) — or credible sources genuinely disagree, so it is true-but-disputed. A claim resting only on an interested party's own materials, with nothing independent behind it, is [U], not [V3]. The specific weakness is named and travels with the claim. A [V3] the call leans on is where the analysis is most exposed: treat it as the first thing to verify further. |
| Dominant unknown | The single fact that, if known, would most change the conclusion. |
| Deep core | The questions no evidence can settle — framing, legitimacy, opportunity cost, value. Yours to judge. |
| The gate | The forced step where you take a position on each deep-core question before the brief finalises. |
| Grade-lock | Once set, a claim's grade never changes downstream — only its wording does. What makes the output defensible. |
| Disconfirmation | For load-bearing claims, a second search aimed at disproving them — the counter to confirmation bias. |
| Coverage-retention | A re-scan for material issues that focusing on the load-bearing few may have dropped. |
| Systems-investigation pass | A conditional step (Step 5; added v0.1.7) for genuinely systemic problems: a graded causal map of feedback loops, tipping conditions, and leverage points — investigating structure, never predicting whether or when the system tips. |
| Proportionality triage | The fail-safe pre-flight that right-sizes the analysis — short answer for the unambiguously trivial, full depth otherwise. |
| Provocation page | The fixed one-page method applied verbatim every run — the engine's analytical L1. |
| Wicked problem | One with no verifiable solution, where the deep core dominates. The call comes as a provisional stance, not a verdict. |
| Adjudication (Step 10) | An opt-in, off-by-default independent second pass on a finished high-stakes analysis: a separate different-lab model (OpenAI / Gemini / Grok, or an "Other" endpoint; a Claude-lineage model is refused) re-derives it blind, then adversarially; disagreements fold back under grade-lock; it never changes the call. Its decision-flip benefit is unproven; what it adds is defensibility and a decorrelated check. |
| Independence ladder / rung | The order of adjudicators, strongest first — A a different-lab model · B a different in-house model · C a panel of same-model instances · D a self-adversarial reset — chosen by what is set up; the engine declares which rung it reached. |
Insight Engine v1.0.0 — User Operating Guide. The large language model provides the insight; the engine makes it verified, defensible, and audience-ready. For the full design specification, layer-by-layer detail, and the controlled-testing record behind each feature, see the companion Architecture document.