Summary

scripts/regional-snapshot/snapshot-meta.mjs produces a snapshot_confidence score that is presented to users as a trust signal for regional intelligence snapshots. The score is structurally broken in two ways that compound: stale inputs default to "fresh" when they lack a timestamp, and valid_until is unconditionally set to +6 hours regardless of input freshness. A snapshot built entirely from stale data will display high confidence and remain marked valid for 6 hours.

Defect 1 — Undated inputs treated as fresh (freshness.mjs:73-77)

// freshness.mjs lines 73-77 (classifyInputs):
// 3. If neither yields a timestamp, fall back to "fresh" (cannot prove
//    staleness). This fallback remains so we don't regress existing
//    keys that have never needed a meta key.

When an upstream seeder crashes and leaves a stale Redis payload that has no timestamp field, classifyInputs() returns "fresh" rather than "stale". The snapshot is built from the old data and snapshot_confidence is computed as if all inputs are current.

This is acknowledged in the codebase itself (todos/181-pending-p2-undated-inputs-treated-as-fresh-confidence-bug.md) but unfixed.

Defect 2 — valid_until hardcoded to +6 hours regardless of input staleness (snapshot-meta.mjs:53)

// snapshot-meta.mjs:53
valid_until: Date.now() + 6 * 60 * 60 * 1000, // 6h

valid_until is computed at snapshot write time, always set 6 hours ahead, and never reduced based on whether any inputs are stale. A snapshot that classifies 80% of inputs as stale still claims full 6-hour validity.

Combined impact

The two defects compound:

1. Seeder crashes → stale payload with no timestamp
2. classifyInputs returns "fresh" for that payload (Defect 1)
3. snapshot_confidence formula: 0.6 * completeness + 0.4 * freshness computes a high score
4. valid_until is set to +6h (Defect 2)
5. Users see a high-confidence, valid-for-6-hours snapshot built from stale data

There is no observable signal to the user that any of this happened.

Confidence formula

// snapshot-meta.mjs:43
const snapshot_confidence = round(0.6 * cCompleteness + 0.4 * cFreshness);

The weights 0.6 and 0.4 are also undocumented, but the more critical issue is the freshness default that makes cFreshness artificially high.

Suggested remediation

1. In classifyInputs(): flip the default for present-but-undated inputs from "fresh" to "stale". The existing metaKey pattern already handles keys that legitimately lack timestamps — they should be using the metaKey path.
2. In buildPreMeta(): derive valid_until from the oldest fresh input's expected age rather than a fixed offset.
3. Expose stale_inputs and missing_inputs in the UI alongside the confidence score so users can assess the basis for the score.