Indicators
Named numbers. Open formulas.
The indicators Cascadian publishes about its own honesty — defined now, populated at launch.
Self-measurement
Six numbers we answer to.
Each indicator is named, defined, and given its formula before a single value exists — so when the numbers arrive, the math is already public.
Definitions final — values ship with the live pipeline
| Indicator | Definition | Formula | Populates at launch |
|---|---|---|---|
| Calibration Drift | How far frozen expectations sit from settled outcomes, on average. | mean(expected − actual) · trailing 20 closed decisionsreview_calibration() | — |
| Leakage Rate | The share of reconstructed decisions where any input postdates the question. Target: zero. | % decisions with any item available_at > as_ofdetect_leakage() | — |
| Edge Acceptance | How many causal edges proposed to the graph survive human review. | approved edges ÷ proposed edgespropose_edge() | — |
| Reconstruction Fidelity | How often a past decision’s context replays exactly as it was assembled. | % context replays byte-exactreconstruct_decision() | — |
| Abstention Rate | How often a context pack says “not enough” instead of guessing. | % packs carrying an abstention reportget_context_pack() | — |
| Approval Latency | How long a requested approval waits for a human to grant it. | median(granted_at − requested_at)request_approval() | — |
No value appears here until the pipeline that computes it is live. The formulas are the commitment: when the numbers land, they are scored against definitions that were published first.