Three checks run on every decision. Each one says why.
Behind the Impact screen sit three checks that every proposed action passes through: is it safe, how far does it spread, and which safeguard objected. Each returns reasoning you can read, not just a score. This page is for engineers; the business view is on the impact analysis page.
In plain termsA safety check, a “how far does this reach” check and a “who objected” check. Each explains itself.
A note on the word. The product deliberately does not show “reasoner” to users. The interface calls this surface Impact, because “reasoner” is the builder’s word and not the reader’s. This page is the technical companion reachable from it.
Three reasoners, invoked on every decision
| Reasoner | The question it answers | What the product shows |
|---|---|---|
| Safety envelope | Is this action admissible at all? If applied, does any constrained state leave its safe envelope? | Per-dimension headroom in real units, plus the binding dimension: the one closest to its bound relative to what the action requires |
| Action-consequence diffusion | How far does this spread, and how expensive is it to unwind? | Blast radius: named groups, entity count, reversibility |
| Guard arbitration | Which rail objected, and how heavily? | The verdict on a held action |
Each answers something the rest of the decision path does not. Ranking asks what is this action worth? Diffusion asks how far does it spread? The envelope asks is it admissible? Those are three different questions and a decision needs all three, which is why they are three reasoners and not one score.
The envelope answers a question nothing else in the platform asks. The other rails constrain spend, frequency and evidence conflict; none asks whether the action leaves room for the dynamics that follow it. A state pinned at its rail has nowhere to go, and that is exactly the condition worth holding an autonomous action for.
Measured on a reference world
On the OgMart reference world (fictional), the graph these reasoners read is projected at boot and reproducible run to run:
- vertices
- 23,931
- edges
- 72,998
- hyperedges
- 2,148
OgMart reference world, a fictional 100-store retailer. Counts are properties of that world, not of any customer.
For one asset on that world, blast radius resolves to 24 real cohorts and 410 affected entities, with reversibility 0.54. An entity that is not in the projected graph reports `blast radius: unknown (entity not in the projected graph)` rather than a plausible number.
Why a chain beats a score
A score tells a practitioner what to do. A chain tells them whether to believe it.
Every reasoner emits steps, and every step carries its basis: the state dimensions involved, the evidence considered, the rule or relationship applied, and, where relevant, the conflicting evidence that was not decisive. That is what makes an overrule possible and defensible: an analyst who disagrees can point at the specific step they reject, and that objection is recorded.
- Disagreement becomes specific. Not “I don’t trust it” but “the third step assumes last week’s supply pattern still holds, and it doesn’t”.
- Reasoning is reviewable after the fact. The chain that produced a decision is retrievable.
- Reasoning is teachable. A new analyst can read the chain on last quarter’s real cases instead of building intuition from scratch.
Reasoning over memory: partly wired
Long-term memory changes what a reasoner can say, because it knows what happened to comparable entities and how past decisions turned out.
| Without memory | With memory |
|---|---|
| “This entity is diverging.” | “This entity is diverging the way 14 others did, and 9 of those ended in a chargeback.” |
| “The expected effect is −310.” | “The expected effect is −310, and when we ran this action on similar entities it delivered −280 median.” |
A missing graph degrades to “unknown”, never to a plausible number
An absent measurement and a zero are different claims, and a platform that cannot tell them apart is not auditable. Every graph read therefore carries an explicit not projected state. If a reasoner cannot resolve an entity, it says so rather than substituting a default.
How far a decision spreads
Diffusion asks how far an action reaches, and how expensive it is to unwind.
Read a reasoning chain on your own cases
In a technical session we run the three reasoners on a decision from your domain and read each step together.