Assumptions
Everything the model believes, stated before the math. If you do not buy one of these, do not buy the output.
- The decision table is published, not learned: each (signal, severity, confidence) combination maps to decision language from the Section 37 action list.
- The table composes and translates numbers; it never creates one.
- Atlas recommends; the capital team decides.
- The table is versioned like a model; a change to the table is a version event.
Derivation
The mapping is a lookup over model outputs, materiality severities, confidence grades and the customer's position:
The action list (spec Section 37): proceed, proceed with conditions, continue monitoring, increase reserve, delay funding, require additional evidence, revise underwriting, reduce exposure, restructure, renegotiate terms, seek additional guarantees, require milestone completion, pause, exit. Every recommendation carries the reasons and the evidence behind the inputs that produced it.
Identification strategy
Why the estimate supports the claimed meaning - and where the model refuses to claim more than the data carries.
- A published table is identified the way a lending policy is: the same facts always produce the same language, and any change to the mapping is a reviewed, versioned event. A learned policy could not be audited this way.
- The never-creates-a-number rule keeps the engine honest: every figure in a recommendation traces to a model output or a record field, so the recommendation inherits the models' provenance rather than adding a new unaccountable layer.
Inputs and outputs
Reads
- All model outputs
- The full model suite's current outputs on the book
- Materiality
- R2 severities on recent changes
- Customer's position
- The covered book's exposure per project
Writes
- Decision language with reasons and evidence attached (spec Section 37)
- Decision snapshot recorded for later re-underwriting
Calibration and gate evidence
Validation design
- The table is reviewed against resolved cases: on projects that later failed, the table must have recommended caution before the failure was public.
Evidence on record
- Resolved-case review results are not yet published in the sealed source.
Failure modes
The ways this model declines, degrades or goes UNKNOWN - and what it prints when it does.
- Inputs carrying UNKNOWN: the table maps to evidence-demanding language (require additional evidence), never to false comfort.
- A (signal, severity, confidence) cell with no mapped action returns the conservative default and flags the gap for table revision.
Sensitivity
- Confidence grades gate aggression: the same signal at LOW confidence maps to evidence demands before it maps to exposure changes.
Iron rules
The decision engine translates model outputs into actions. It never creates a new number. Atlas recommends; the capital team decides.
Source and audit
Written from the sealed model source. Every run is recorded - model version, frozen inputs, input hash, timestamp. There is no unrecorded model execution.