| name | briefing-scorecard |
| description | Weekly audit of the Daily briefings: score every claim in the trailing week against the domain vault reports and open-items ledgers using the forensic taxonomy (verified / stale-echo / fabricated / unverifiable-personal), run the mechanical assertions, and write the scorecard to /vault/Briefings/Scorecards/. This is how prompt and model changes stay testable — the next week's real runs ARE the test.
|
Briefing Scorecard — Weekly Claim Audit
You audit what was published against what was true. Sources of truth,
in order: the domain vault reports for that day (/vault/Briefings/Security| Intel|HomeLab|Home|Finance|Career|Wellness/, matched by mtime — Intel
filenames are sometimes future-dated), the open-items ledgers
(/vault/Briefings/Ledger/<domain>.json), and the verifier reports where
they exist. You never rescore a claim as true because it sounds plausible —
only because a source confirms it.
Scope
The 7 most recent /vault/Briefings/Daily/YYYY-MM-DD.md files by mtime
(skip -home variants and fallback files, but note their existence).
Per-claim taxonomy
Extract every factual claim (a statement checkable in principle — statuses,
Day-N counters, incidents, resolutions, "missing report" assertions,
citations). Classify each:
- verified — a same-day source (vault report, ledger, verifier report)
supports it.
- stale-echo — was true once, wasn't true that day (carried items,
outdated statuses, counters incremented without confirmation).
- fabricated — no source supports it (invented citations, invented
numbers, phantom incidents, wrong knight-liveness).
- unverifiable-personal — personal-domain claim with no checkable source;
not an error, tracked separately.
Attribute every non-verified claim to the stage that introduced it:
knight (wrong in the vault report too), summary-hop (right in the
vault report, wrong/missing in what reached synthesis), synthesis
(right in the inputs, wrong in the briefing), chain-spec (prompt
told the knight something false).
Mechanical assertions (pass/fail per briefing)
- No unconfirmed Day-N: every "⏳ Day N" item has same-day confirmation
(ledger
open + last_verified = that day, or an explicit summary
bullet). Day-N values equal date - first_seen where the ledger has one.
- No counter arithmetic drift: no counter decreases, jumps by more than
the elapsed days, or contradicts another instance in the same briefing.
- No pregnancy-era tokens: no due-date/countdown/pregnancy-week/
nursery-prep/hospital-bag language (Emma was born 2026-04-14).
- Liveness honesty: no domain listed missing that has a same-day vault
report (by mtime), and no domain presented as reporting that has none.
- Citations exist: every cited report/file exists with a plausible
mtime; the citation's content direction matches the claim.
- Numbers traceable: spot-check ≥5 numbers per briefing back to a
source; any untraceable number fails this assertion.
- Length discipline: no padding; a quiet day reads short; over 150
lines fails.
- Required sections present: TOP 3, CROSS-DOMAIN, DOMAIN SNAPSHOTS,
MISSING REPORTS.
If the ledgers or verifier reports don't exist yet for part of the week,
run what's runnable and mark the rest "not assessable — ledger absent";
never fail an assertion on missing infrastructure.
Output — /vault/Briefings/Scorecards/YYYY-Www.md
ISO week of the Monday you run on. Create the directory if missing. Contents:
- Headline table — per day: claims, verified, stale-echo, fabricated,
unverifiable-personal, corrupt % (corrupt = stale-echo + fabricated).
- Stage attribution — corrupt-claim counts per stage
(knight / summary-hop / synthesis / chain-spec).
- Assertion matrix — the 8 assertions × 7 days, pass/fail/n-a.
- Trend line — vs the previous scorecard file if one exists; call out
any error class that grew.
- Worst 5 claims — quoted, with the evidence that disproves each.
- One-line verdict — is the pipeline getting more honest or less?
Keep the whole file under ~120 lines. Numbers you print are subject to the
same rules you're enforcing: every one traceable to a check you ran.
Baseline
The pipeline's pre-hardening baseline (2026-07-02/03/04) was 47% / 53% /
52% corrupt — scored with this same taxonomy. That is the number to beat;
week-over-week comparisons are valid only if the taxonomy above is applied
unchanged.