Assumptions
Everything the model believes, stated before the math. If you do not buy one of these, do not buy the output.
- A promise is never overwritten; resolutions attach to the promise. The append-only history is the record of truth.
- Sponsor statistics are empirical base rates with published shrinkage toward the sector base rate.
- The shrinkage weight is documented, not tuned per case: shrink = n / (n + K) with K = 3.
Derivation
Per sponsor and per project the model keeps the promised-vs-actual ledger: original COD, revised COD, actual COD, original capacity, delivered capacity, and the resolution (delivered, delayed, downsized, cancelled). Sponsor statistics - delay frequency, slip distribution, cancellation and downsize rates, capacity-attainment ratio - are empirical frequencies over that ledger.
The sponsor hazard multiplier the Delivery Model consumes is the sponsor's own rate shrunk toward the sector base rate:
A sponsor with three resolved projects is half-shrunk; with thirty, the shrinkage barely binds. The sector base rate is computed across all resolved outcomes, so the shrinkage target is itself identified from the corpus.
Identification strategy
Why the estimate supports the claimed meaning - and where the model refuses to claim more than the data carries.
- Track record is identified from resolutions only - promises that actually landed, slipped or died. Open promises contribute nothing, so a sponsor cannot build a record on announcements.
- Shrinkage is what makes small samples honest: a sponsor with one resolved project cannot distinguish skill from luck, so the estimate borrows strength from the class in proportion to how little the sponsor's own data can say. The published K makes that trade explicit and arguable.
Inputs and outputs
Reads
- Outcome data
- Spec Section 40: resolutions with promised and actual dates and capacities
- Append-only history
- Field history; promises are never overwritten
Writes
- Promised vs actual per project and per sponsor
- Sponsor track record: delay frequency, slip distribution, cancellation and downsize rates, capacity-attainment ratio
- Sponsor hazard multiplier (consumed by the Delivery Model)
Calibration and gate evidence
Validation design
- The shrinkage weight is validated by leave-one-out prediction across resolved sponsors: each sponsor's resolution is predicted from the shrunk estimate fit without them.
Evidence on record
- Leave-one-out results are not yet published in the sealed source. K = 3 is the published constant.
Failure modes
The ways this model declines, degrades or goes UNKNOWN - and what it prints when it does.
- Sponsor with no resolved projects: the multiplier is the sector base rate, and the page of record says the sponsor's own sample is empty.
- Promises with missing resolution dates do not enter rate denominators; they are counted as unresolved.
Sensitivity
- K controls how fast a sponsor earns their own record: K = 3 means the fourth resolved project still carries 43% sector weight. Changing K moves every thin sponsor's multiplier, which is why it is published rather than tuned.
Iron rules
A promise is never overwritten; resolutions attach to the promise. This model is the sponsor track record.
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.