07 · Valuation and strategy
A range you can take apart, never a number to trust.
Claims, elements, contested facts and comparable outcomes are authored into a model, simulated, and published as an interval with its assumptions attached.
A range is a record of assumptions, not a promise.
The question in the room
“Why would I believe a number a model produced?”
How it works
The mechanism, not the promise.
You cannot get a point estimate
A currency figure not bound to a second, different figure raises. It parses magnitude words, so the same number written two ways is caught, and it requires the two ends to actually differ after a real matter produced an identical low, middle and high that every range-shaped check waved through.
Five authors, and the median model kept whole
Extended reasoning cannot be pinned to a temperature, so five independent authors run and the median member's model and simulation are published together rather than blended, which keeps the published model internally coherent.
A census that admits a thin run
A run where four of five authors failed once published a median from one author that was indistinguishable from a healthy run. Every snapshot now carries how many authors returned, and an older snapshot with no such record reads as unknown rather than healthy.
An interval, never a point
ShippedThe honest interval can only widen the simulated band, never narrow it, and the method that produced it travels with the number.
backend/src/lib/analytics/conformal.ts
One kill shot per matter
ShippedFour surfaces once answered what single thing decides this case and disagreed on the same matter. Now there is one answer, and it is stored.
backend/src/lib/analytics/killShot.ts
Adversarial war-gaming
ShippedThe opponent wins each pivotal issue in turn, under shared random draws so the comparison is not simulation noise.
backend/src/lib/analytics/warGame.ts
The Red Brief
ShippedThe case against your own position, written from the same record.
backend/src/lib/redbrief/generate.ts
Provable precedence
ShippedA prediction is committed to an append-only ledger before the outcome is known, which is what makes a later score mean anything.
backend/src/lib/analytics/provablePrecedence.ts
Refusals
What it will not do, and why.
- Publish a calibration result.
- The public ledger has no scorable outcomes, and publication refuses below ten resolved matters. It states the insufficiency instead of filling the space with a number.
- Narrow the band because the calibration set is thin.
- A thin set returns an infinite width before any clamp, so it degrades into a wider honest interval rather than a falsely tight one.
- Hide a cap that bound the outcome.
- Where a coverage ceiling applied, the share of simulated worlds it bound and the mean amount it removed are both reported. A cap applied without saying so is its own dishonesty.
- Re-author a pinned valuation on its own.
- A pinned matter returns the frozen model, and if that model cannot be read it raises rather than quietly authoring a fresh number.
One record
Nothing here stands on its own.
What feeds it
- Established facts and contradictions
- Comparable outcomes in the corpus
- Coverage and damages records
What it feeds
- Settlement posture
- The client conversation
- The committed prediction ledger
Limits
What this does not establish.
- This is an estimate, not a measurement. Nothing here has been scored against a resolved outcome.
- The band is calibrated and widened by disagreement. That is not a claim of coverage at a stated level.
- The default outcome model averages over liability, so only the mean is sound and the published limitation sentence travels with any percentile.
Check the work before you believe it.
Run a sample manifest in your own browser, or read where publication refuses for want of evidence.