freshness-corroboration
Trust-stage audit for current and corroborated brand claims. Use when AI answers show outdated or contradictory facts or content lacks trust signals. Checks machine-readable dates, staleness and fake
SKILL.mdskills/freshness-corroboration/SKILL.md
GitHubFreshness & Corroboration (trust layer)
Handout appendix D states the mechanism: machines treat a fact as more trustworthy when many independent places say the same thing, and a claim that lives in only one spot is fragile.
This layer audits both halves of that — whether a claim is current, and whether it is corroborated.
When to use
After crawl-reachability. Run standalone when a brand's facts are stale or
contradicted in AI answers.
Inputs
- Required — a run workspace from the two upstream stages:
preflight.jsonfor the scope domain andreachability.jsonfor the page inventory and saved HTML paths. - Optional —
--config(anAuditConfigJSON) and--strictness(lenient|balanced|strict) to override staleness windows and the page-type recency expectations. Defaults come fromAuditConfig. - Not consumed — no network access, no prior stage's findings. Everything here is derived from markup already fetched, so the layer is re-runnable offline and produces identical output.
python3 scripts/check_freshness.py [--workspace PATH] [--json]
Procedure
Deterministic, in this order. Staleness windows come from the page type, never from a universal constant, and no wall-clock value enters a threshold — the clock is read once, for the audit timestamp only.
- Load the inventory and pages. Resolve the scope domain from
preflight, then read every fetched page; record pages with missing HTML as a coverage limitation. - Datestamp presence (
FR001). Check for any date on page types whoseexpect_dateis true; page types carryingexpect_date: falseare exempt by design, not by accident. - Machine-readability (
FR002). Distinguish a date a human can read from a date a machine can parse —<time datetime>, JSON-LDdatePublished/dateModified, meta article timestamps. - Staleness (
FR003). Age the page against what its type implies about recency. - Fake freshness (
FR004). Compare thedateModifiedgap against actual signals of revision, to catch timestamps bumped on every deploy. - Load-bearing claims (
FR005). Flag statistics, prices and superlatives asserted with nothing behind them. - Outbound citation density (
FR006). Measure references on the page types where citing is conventional. - Corroboration profile (
FR007,FR008). Determine whether every claim is single-sourced to the site, and whether corroboration rests on press and social coverage only. - Attach the sample caveat. Every population claim carries its sample size for the advisory layer to verify.
Checks and severities
Every check the procedure above can emit, with the claim level it is allowed to assert.
| Check | Detects | Severity | Claim level |
|---|---|---|---|
FR001 |
No datestamp on page types that imply recency | high/medium | 1 |
FR002 |
Dates present but not machine-readable | medium | 1 |
FR003 |
Content stale relative to what its page type implies | medium | 1 |
FR004 |
dateModified far newer than datePublished with no visible revision |
medium | 1 |
FR005 |
Statistics and superlatives asserted with no source | medium | 3 |
FR006 |
Explanatory pages with no outbound references at all | medium | 1 |
FR007 |
Every factual claim single-sourced to the site itself | medium | 3 |
FR008 |
Corroboration resting on press and social coverage only | medium | 3 |
The recommendation people get backwards
FR008 is the non-obvious one. The instinct when a brand is invisible is to
pursue press coverage, and press coverage does help — at the selection layer.
But arXiv:2604.25707 §8.4 measures mean influence by publisher domain type:
| Domain type | Mean influence |
|---|---|
| encyclopedia | 0.2144 |
| academic_publishing | 0.1118 |
| commercial | 0.1028 |
| nonprofit | 0.0971 |
| academic | 0.0815 |
| government | 0.0769 |
| news_media | 0.0726 |
News media is the most frequently selected non-official source type (16.07% to 31.17% of citations, §7) and the least absorbed in this table. Encyclopedia pages absorb roughly three times as deeply. The paper's reading is that answer engines often need explanatory pages after they have identified timely ones.
So a corroboration profile made entirely of press hits optimises the weaker of
the two outcomes. FR008 flags exactly that shape and recommends seeding
explanatory reference coverage alongside the press.
Fake freshness
FR004 catches a specific bad practice: bumping dateModified on every deploy
so content looks maintained. It compares the modification gap against actual
signals of revision. A site where every page claims to have been updated
yesterday and none of them changed is making a claim a crawler can check and
find false, which damages exactly the trust the timestamp was meant to build.
Page-type awareness
Staleness thresholds come from what the page type implies. An undated contact
page is fine; an undated news article is not. Pages whose type carries
expect_date: false are exempt from FR001 — a blanket "no datestamp" check
fires on most of a normal site and teaches the reader to ignore the report.
Output
05-freshness-corroboration.json with per-page date inventory, staleness
distribution, outbound-reference hosts by category, unsourced-claim samples,
and the findings.
Guardrails
Read-only; consumes saved HTML only. Corroboration is assessed from signals
observable on the site itself (outbound references, sameAs, cited sources) —
the audit does not crawl third-party sites to verify them, and says so in the
report's limitations rather than implying an off-site check it did not perform.
Grounding
Handout appendix D (agreement across sources); §8.4 domain-type influence;
§8.3 semantic-role influence, where reference-only citations sit lowest at
0.0529 while definition leads at 0.1531. Details in
references/corroboration.md.