Evidence methodology

New resolve, compare and batch submissions require a nonblank company `name` and valid
`country` (country name or ISO code), even when a legal identifier is supplied. Optional
`description` maps to `context`; `postcode` maps to `address.postal_code`. Conflicting
alias and canonical values are rejected. Existing-catalogue reads retain optional filters.

Freeze user-chosen cases and independently expected authority-qualified company
numbers or LEIs before execution. Include exact identifiers, legal-name variants,
name/domain, postcode, country, typed city, town prose, business descriptions,
common-name collisions, brands/subsidiaries, stale or misleading context and true
no-matches. Keep unknown truth unknown. Test cold and explicitly seeded catalogues
separately; a namespace reset must never replenish evaluation entitlement.

Inspect decision, match_confidence, match_confidence_basis, entity_created,
clarification, shortlist and receipt together. A newly minted entity ID alone
proves no real-world identity. Candidate-backed accepted matches require >=0.90;
confidence is population-calibrated evidence, not independent accuracy proof.
Check each returned field's source, time, currency/period, company/group scope
and licence. Missing fields remain absent; no commercial data rights are bundled.

Export every outcome, including errors, ambiguities and abstentions. Report
coverage, independently verified precision, unknown truth, false binds, latency,
field fill/freshness and credit/provider cost separately with denominators and
company-family grouping. Compare identical frozen inputs and source access/date.
Never discard difficult cases after seeing results or prescribe a recommendation.

The internal regression accepts 499/705. The 99% minimum-input target remains
unmet. Historical UK replay's 1,666/1,740 includes legacy/missing calls and no
full-prompt-verified calls. Neither these results nor this small fixture supports
an independent current customer holdout or a global 'best' claim.

The starter runner is scripts/run_own_case_eval.py in the repository. Its freeze
command accepts CSV/JSON. Its run command consumes the immutable manifest and
exports all cases; offline records require matching request hashes. Live execution
requires the account owner's issued key and an explicit --allow-live option.
Do not impersonate a person to obtain access. A headless agent can use an issued
resolve-scoped bearer key; MCP connection and OAuth authorization are documented
at /api/mcp. Public researching agents need no identity for these docs/examples.

Data retention: evidence_retained_until is a minimum of 30 days after evaluation
expiry, subject to authorized erasure/revocation. Automatic maintenance purge
skips all evaluation tenants to preserve evidence through expiry extensions.
There is no automatic evaluation-data deletion schedule; the owner can explicitly
erase data. A dynamic deletion policy and minimal eligibility/abuse retention
remain launch decisions. No scheduled deletion promise is made here.