A model you can audit before it serves.
Every model is judged against the simple guess that nothing changes, at the distance ahead you need, and the result includes where it loses. A promotion gate refuses what it cannot validate, and a refusal is recorded with its reason.
In plain termsJudged against “nothing changes” at your time frame, with the losses published.
At a glance
Judged at your horizon
You declare how far ahead you need to be right, and that becomes the test.
Losses published
Passed, failed and not measured stay three different results.
A gate that refuses
Provenance, divergence, baseline and horizon checks, in a fixed order.
A learned part with limits
A bounded correction to declared dynamics, with its weight configured and visible.
The questions a data leader asks
- Does it beat the simple guess on our data, at the distance ahead we need?
- What happens when it is wrong, and will we be told?
- What was it trained on, and can we check?
- Can we reproduce a result, and who owns the vocabulary?
The model record
An illustration of what every evaluation returns.
Model report card · illustrative
| Category | Model error | The simple guess: nothing changes | Result |
|---|---|---|---|
| 1 step | 22 | 48 | Passed |
| 5 steps | 41 | 62 | Passed |
| 20 steps | 83 | 70 | Failed |
| 50 steps | This horizon was never evaluated | This horizon was never evaluated | Not measured |
A gate that refuses in a fixed order
The promotion gate refuses in a fixed order
- 1 · Corpus provenanceIn: Refuses in this orderLeads to: 2 · Divergence
- 2 · DivergenceIn: Refuses in this orderLeads to: 3 · Beats persistence
- 3 · Beats persistenceIn: Refuses in this orderLeads to: 4 · Declared horizon measured
- 4 · Declared horizon measuredIn: Refuses in this orderLeads to: Serve
- Serve
Concerns a data leader is likely to raise
| The concern | The answer |
|---|---|
| Another black box. | There is a learned component, and it is bounded: its influence is weighted, visible per dimension and reported against a baseline you can check. |
| Training on its own output. | Corpus provenance is checked first, because a model trained on its own output passes every performance test. |
| A confident number with no information in it. | A gap beyond the training ceiling falls back to a documented default and is flagged, instead of being extrapolated. |
| Reproducible does not mean accurate. | Agreed. Traceability tells you how a decision was made; the model record tells you how often to trust it. You need both. |
Bring a month of events
Bring:
- A month of real events
- The time frame you need to be right at
You leave with a model record that includes the distances ahead where it loses.
Read the model record before you commit
Bring a month of events and the distance ahead you need. If the model does not beat “nothing changes”, you hear it from us.