how it works
How a measure is made
A measure is the smallest unit the engine produces: one marker, positioned on a fixed scale, against evidence you can read. This is the whole mechanism, end to end — no metaphor required.
The scale it measures against
Every measure sits on one ruler: CMMI, 1.0 to 5.0. The numbers do not change; the words describing each band can be tuned per compass.
- 1.0 · Initial — things happen, but ad hoc.
- 2.0 · Managed — there is a process, project by project.
- 3.0 · Defined — the process is written down and shared.
- 4.0 · Measured — the process is quantitatively managed.
- 5.0 · Optimising — the process improves itself.
Three ways of taking the reading
The scale is fixed. How the engine reads a position against it is the part that’s configurable — the epistemology is a parameter.
Deterministic — you give a number
A direct question (“Do you deprovision leavers within 24 hours?”) answered on a 1–5 scale. Some markers weigh more than others; the engine weights and aggregates. Fast, broad, shallow — a written policy and a followed policy look the same here.
Stochastic — you tell what happens
An open account of how the work really happens. The engine reads it and, for each marker in the compass, asks whether the evidence is present. A described, measured SLA checks a band-4 marker; silence on orphaned accounts is a missing-evidence signal. Slow, narrow, deep — the account shows what a slider can’t.
Composite — both, weighted
Most compasses run both and weight them (say, 30% deterministic, 70% stochastic). The questions catch the broad strokes; the account catches the nuance. One setting, no code change.
Belief, updated by evidence
For each marker the engine keeps a belief over which band the organization is in — a Dirichlet posterior across the five bands. It starts knowing nothing: every band equally likely. Then each piece of cited evidence pushes belief toward the band that evidence supports.
Prior — “no idea yet”
Posterior — after the evidence
Three pieces of evidence for band 3, two for band 4, none for band 5. The expected band — Σ (band × probability) — settles near 3.0. Belief × evidence = updated belief; nothing moves without a citation.
A deterministic floor under the intelligence
The reading can use a language model to extract evidence, but it cannot conclude whatever it likes. Beneath every stochastic reading sits a fixed floor: the band definitions, the allowed vocabulary, the rules that bound a plausible interpretation. A claim of band 4 without evidence of measurement is not admitted — it is held at the band the evidence supports.
The model proposes. The floor disposes.That is how the engine uses a language model without inheriting its imagination.
Uncertainty is reported, not rounded away
A sharp belief — one band near 80% — is low variance: the engine is sure. A flat belief across bands is high variance: it is hedging, and says so.
- 3.2, low variance — “this is around band 3.”
- 3.2, high variance — “could be 2, could be 4 — not known yet.”
Same number, different sentence. The engine shows both. False precision is the house style of every tool kerte is not.
When it stops asking
More evidence keeps tightening belief — until it doesn’t. The engine stops on the first of three conditions, so it asks enough to be confident and then respects the reader’s time:
- Confident — variance drops below 0.25.
- Stable — the expected band stops moving by more than 0.05 over three rounds.
- Bounded — a hard cap of ten rounds, then stop regardless.
The human is the judge
After the evidence is read, the engine shows every cited marker: “we heard you do X — yes or no?” The reader confirms, flips, or skips. A flip tells the engine it read wrong; the reader’s confirmation overrides the calculation.
This is the step most automated assessments skip — they run a prompt and hand back a score. kerte shows its work, citation by citation, and lets you correct it. The engine is the calculator; the reader is the judge.
Why this beats a score
Most maturity tools average yes/no answers into a number that tells you nothing about how sure it is. A measure does three things instead:
- It cites its evidence. Every belief shift points to a literal quote you can see and overturn.
- It tracks uncertainty. A confident 3.2 and a hedged 3.2 are different signals, and stay different.
- It updates, it doesn’t restart. Next quarter’s account updates last quarter’s belief — long-term tracking stays honest.
Calibration is a recorded state, not a claim. A compass starts uncalibrated — uniform marker weights, validation still ahead — and is marked framework-calibrated once its weights are tuned and it has cleared the Tier-One gate. The status travels with every measure, so a reader always knows which they’re looking at.