Reconciler

Reconcile a finding by hand

This runs lib/reconcile.ts only, the same deterministic checker used in production, in your browser, with no model call. Build a finding, pick a scan history, and see every check.

Example cases

Each button loads a ready-made finding and switches to the eval record so its numbers line up. Press Reconcile after loading one.

Everything lines up. The cited value is in the scan and matches exactly, the model's risk level agrees with the one computed from reference ranges, and the write-up only names backed numbers. Expected result: grounded (shown normally).
The finding cites LDL 4.2, but the scan says 3.6. A hard check fails. Expected result: rejected (kept out of the clinical view).
The scan shows LDL rising, but the claim says it fell. The direction is recomputed and disagrees. Expected result: rejected.
The value is real, but the model marks it 'priority' while the reference range puts it at 'watch'. The computed level wins for display; the disagreement is surfaced. Expected result: flagged (review carefully).
The claim ties out, but the sentence mentions an HbA1c reading that is not one of the finding's cited sources. A soft check. Expected result: flagged.
The finding cites 'blood.vitaminD', a measurement the scan does not contain at all. Expected result: rejected.

Inputs

Base recordThe scan history the finding is checked against. Every scan and value the finding cites has to exist here.

Claim kindlevel = one value at one date. trend = a value moving between two dates. observation = a note with no number to verify.

leveltrendobservation

metricThe path to the measurement inside a scan, for example blood.ldl or heart.bpSystolic.

valueThe number the finding claims for that measurement.scanDateWhich scan it reads that value from (YYYY-MM-DD).

proposedTierThe model’s guess at the risk level. Never displayed to a clinician. Shown here only so you can watch it agree or disagree with the level computed from reference ranges.rationaleThe model’s plain-language write-up. Checked so that every number and measurement it names also appears in the sources below.

provenance (sources)The exact measurements the finding is allowed to cite as evidence. Each row is a metric path, the value found there, and the scan date.

metricvaluescanDate

Result

The five checks it runs

Referential integrity - every measurement the finding cites actually exists in the scan on that date.
Value tie-out - the numbers it cites match the scan exactly.
Trend consistency - if it claims a direction, the record moves that way when the difference is recomputed.
Tier derivation - the model’s risk level matches the one computed from reference ranges (or is flagged if no range exists for that metric).
Prose coverage - every number and measurement named in the write-up is one of the cited sources.

The first three are hard: a failure rejects the finding. The last two are soft: a failure flags it for review.

Edit the finding and press Reconcile. Nothing is sent anywhere.

Example cases

Each button loads a ready-made finding and switches to the eval record so its numbers line up. Press Reconcile after loading one.

Everything lines up. The cited value is in the scan and matches exactly, the model's risk level agrees with the one computed from reference ranges, and the write-up only names backed numbers. Expected result: grounded (shown normally).

The finding cites LDL 4.2, but the scan says 3.6. A hard check fails. Expected result: rejected (kept out of the clinical view).

The scan shows LDL rising, but the claim says it fell. The direction is recomputed and disagrees. Expected result: rejected.

The value is real, but the model marks it 'priority' while the reference range puts it at 'watch'. The computed level wins for display; the disagreement is surfaced. Expected result: flagged (review carefully).

The claim ties out, but the sentence mentions an HbA1c reading that is not one of the finding's cited sources. A soft check. Expected result: flagged.

The finding cites 'blood.vitaminD', a measurement the scan does not contain at all. Expected result: rejected.