Model Docs36 models sealed
The model suite, documented to the desk standard.
Every figure Atlas prints comes from one of these models. Each page states the assumptions, derives the math in full, argues the identification, names the failure modes, and shows the calibration evidence. Where the evidence does not yet exist, the page says so - an honest UNKNOWN is a billable output.
Deterministic - same evidence in, same answer out.
The AI never computes - it explains, it never touches a figure.
UNKNOWN is first-class - an honest UNKNOWN beats a confident guess.
Every run is recorded - model version, frozen inputs, input hash, timestamp.
Physical
What is true, what must happen, what it costs.
M1Project State ModelWhat is physically true about this project today, and what happens next.M2Delivery ModelWhen will this project actually deliver, and how much.M3Capacity-at-Risk ModelOf the promised MW, how much is actually at risk, and when.M4Dependency ModelWhat must happen for this project to deliver, and which single failure kills the date.M5Cost ModelWhat will this project actually cost.
Money
What physical truth does to the capital.
Evidence and Trust
How much weight the record deserves.
M8Evidence Weighting ModelHow much weight each piece of evidence deserves.M9Confidence ScoreHow much to trust the estimate itself.M10Conflict ModelWhere sources disagree, and how wide the disagreement is.M25Bayesian Evidence Update ModelHow a verified piece of evidence should move the delivery belief - by contract, never by narrative.M35Reliability and Drift ModelWhether the sources, data, and calibration behind every model are still what they were - drift measured on printed thresholds, suspension enforced by contract.
Change and Future
What moved, and what if it moves.
Wider Book
The pile, the paths, the track record, the book.
M13Named-vs-Pile ModelOf the pile of promises everyone quotes, how much is named, evidenced and real, and how much is speculation.M14BTM-vs-Grid Separation ModelWhich promised MW depend on the grid and which sit behind the meter with onsite generation.M15Resolution Tracking ModelFor every promise ever recorded: did it land, slip or die, and how does that compare to what was promised.M16Portfolio and Concentration ModelWhere the common-mode risk sits across a customer's whole book.
Learning
Whether the calls were right.
M17Prediction Calibration ModelWhen Atlas said 70%, was it right 70% of the time.M26Regime Detection ModelWhether the history the priors are built on still belongs to one regime - or broke at a quarter you can name.M30Forecast Ensemble ModelHow several forecasts for one target combine - without averaging away the differences that matter.M34Reference Class ModelWhich past projects genuinely resemble this one - and what their outcomes say, when the resemblance is strong enough.
Decision
What capital should do about it.
M18Decision EngineGiven everything above, what should capital do.M19Site and Market Suitability ModelWhich candidate sites and markets qualify - gates first, then a visible, weighted comparison that refuses to rank the unmeasurable.M21Opportunity Qualification ModelWhether a verified change to the record is an opportunity worth diligence - qualified, watched, rejected, or too thin to call.M29Portfolio Construction ModelWhich screened combination of projects best fits the mandate - solved exactly, or refused with the reason.M31Decision Timing ModelAct, wait, stage, hold, or abandon - classified from sourced deadlines and named evidence events, never a market call.M32Real Options ModelWhat staging the path buys - priced only when the terms exist, structural when they do not.M33Value of Information ModelWhich missing evidence is worth obtaining first - priority always, dollars only when the terms exist.M36Negotiation Conditions ModelWhich conditions to take into the negotiation - each one derived from an observed risk path on the record, never brainstormed.
Correlation and Effects
Shared exposure and how shocks travel, over documented links only.
M22Structural Correlation ModelWhich projects share a documented exposure, and whether the ones that share one actually resolve together.M23Effects Propagation ModelGiven a verified shock, who carries it - directly, by co-exposure, and downstream - and how far the documented links actually reach.M24Early Warning ModelWhich projects are showing dated, checkable signs of slippage - and what observation would settle each one.M27Causal Effects ModelWhether a shock actually changed outcomes - answered with one approved design, or refused with the failed assumptions named.M28Systemic and Tail Risk ModelHow much of the book can fail together - answered with dependence bounds, never a fabricated covariance matrix.
Pipeline rules
Ingest-side determinism. Not models - they write no audit rows and run no numbers at ask time.
R1Change DetectionA fixed diff classifies every new or changed value in field_history by field and magnitude. The rule that notices; it does not judge. Feeds R2.
R2Materiality ScreenFixed thresholds per change class map each detected change to material or not, with HIGH / MEDIUM / LOW severity. Only material changes fire alerts; financial impact stays UNKNOWN until an agent run of the Financial Bridge fills it.
R3CompletenessCounts the Section 61 checklist field by field into completeness per category. UNKNOWNs count honestly. Completeness is never confidence: a project can be 95% complete and still high risk.
R4FreshnessApplies the Section 45 freshness classes to every field and marks records stale past their threshold. A stale field reads as stale everywhere it appears; staleness is visible, never hidden.