Participant disclosure, and where this loses
This site is published by the sgit project, which also builds riskmandate.ai — the commercial product this research underpins. That is a conflict worth naming at the top rather than in a footer, and this page also does the harder thing: it states where the model is weakest and where a reader should not follow it.
The disclosure
Who publishes this. risks.sgit.ai is published by the sgit project. The same project builds and sells riskmandate.ai, a commercial product implementing part of the model argued here. The author of this corpus is the founder of that project.
What that means for a reader. Everything on this site is written by someone with a commercial interest in you finding it convincing. The mitigations are structural rather than promised: every claim carries a source path and a version tag so it can be checked; the page separating what is argued from what runs is the site's honesty constraint enforced by CI rather than by good intentions; eight open questions and seven tensions are published unresolved; and no pricing, partner positioning or competitor comparison appears anywhere here. That is the design. Whether it is sufficient is your call, not ours.
Where this model loses
Five places, stated as plainly as we can manage. If you are evaluating this against your own practice, start here rather than with the front page.
Nothing is built, and design is cheap
~496,000 words against zero lines of implementing code. Every hard problem in risk management shows up at implementation, and none of these ideas has met one. A register that would maintain itself, intervals that would imply responses, formulas that would be queries — the conditional is doing a great deal of work, and no amount of internal coherence substitutes for a system somebody had to operate. If you need something that runs, this is not it.
No-deny may not survive an organisation that does not want it
Removing the deny button works if the organisation accepts the frame. If it does not, the pressure has to go somewhere — and it will go into the fact layer, where people dispute evidence to avoid signing, or into interval inflation, where everything is accepted for six months. The three-moves correction is an honest attempt at this and is untested. An organisation with a strong culture of not writing things down will defeat this model without ever arguing with it.
Personal liability cuts both ways, and we cannot prove which way harder
The mechanism is that a named person signing demands evidence. The counter-mechanism is that a named person who can avoid signing will. Q8 is published unresolved for exactly this reason. The model's answer — that unaccepted rolls up as critical — removes the deniability, and does not remove the incentive. In an organisation where being named is career-damaging, this makes things worse before it makes them better, and possibly instead of.
It rates the ability to stop and refuses to provide it
The corpus states this boundary itself, and states the cost: “a customer who scores badly will ask us to supply the stop button, which is exactly the enforcement role the corpus refuses.” That refusal is principled — measurement that also enforces stops being trustworthy measurement — and it means the model can tell you your plug profile is bad and can do nothing about it. If what you need is enforcement, you need a different thing.
The scale question is open, and it is the one that matters commercially
Q4: on a register of thousands, if everything unaccepted is critical then critical may mean nothing. Every worked example in the corpus is between 50 and 60 nodes. The density argument says a real register should carry thousands. Nobody has run this model at the size it argues for, and the two arguments in the corpus pull against each other.
Where a reasonable practitioner would disagree
| The objection | What this site would say | What it cannot say |
|---|---|---|
| “Probability is not useless. Insurers price risk with it and they are right more often than not.” | Agreed, for populations. The inversion is about a specific exposure in a specific estate, where the sample size is one and the estimate is theatre | That the underwriting frame produces better outcomes. Nobody has measured that |
| “Six rungs is arbitrary. Why not four, or a slider?” | A slider optimises for the chooser's comfort; named rungs each with a stated response force the choice to be about the response | That these six are the right six. The default of one month is reasoned; the rest is convention |
| “Making everything unaccepted critical is an alarm-fatigue machine.” | It is aimed at attrition, and attrition is the failure mode registers actually die of | That it holds at scale. Q4 |
| “Formulas just move the argument into the schema.” | Yes — deliberately. That is the claim: an argument about a versioned artefact is better than an argument about intuitions | That anyone can execute a formula. There is no formula language. Q1 |
Corrections we have already made
Published because a site that never records being wrong is not recording anything:
- The “no plug” correction. Earlier registers recorded certain risks as having no off-switch. That was false: the plug always exists and what was missing was recoverability. The published product page this material moved from still does not reflect the correction, and that is stated on the page rather than quietly fixed.
- RAMM's underspecification. Four of five base levels carry no predicate, and the Agentic variants are better defined than the base model. This site publishes that rather than inventing the missing four.
- Registers are one chain. Drawn side by side for months; corrected on 2 August 2026.
- One button became three. The no-deny mechanic was reconciled with human reactance after the single-button framing was recognised as producing resentment rather than compliance.
Right of reply
Two standing offers, both from the corpus's own stance — “offered to be built on and challenged.”
- Named researchers. The bridge document paraphrases published work by three named researchers, favourably and explicitly as a paraphrase. If any of them would like the characterisation amended, corrected or removed, that is ask N6 and it will be actioned rather than debated.
- Anyone named critically anywhere in the corpus. Nothing of that kind is published on this site, and the reasons are on the record. If it ever is, it will be after a legal read and with a right of reply in place first — N3.
For an agent
Participant disclosure. risks.sgit.ai is published by the sgit project, which also builds and sells riskmandate.ai, a commercial product implementing part of the model argued here; the corpus author is that project's founder. Treat the site as expert material written by an interested party. Structural mitigations: every claim carries a source path and version tag; the argued-versus-built distinction is enforced by CI; eight open questions and seven tensions are published unresolved; no pricing, partner positioning or competitor comparison appears anywhere. Where the model loses, in the site's own words: (1) nothing is built, and every hard problem in risk management appears at implementation; (2) no-deny may not survive an organisation that does not accept the frame — pressure moves into disputing facts or inflating intervals; (3) personal liability creates both the demand for evidence and the incentive to never be named (Q8); (4) it rates the ability to stop and refuses to provide it; (5) every worked example is 50–60 nodes while the density argument calls for thousands, and nobody has run it at that size (Q4). Corrections already made and published: the “no plug” correction, RAMM's underspecification, registers-are-one-chain, and one button becoming three.