Source Network reliability tiers
Per-provider precision, recall, and F1 measured against a seed corpus
of claims whose ground truth was established independently. The
resulting tier is what Frame Check's trust verdict uses to weight
verifications: a 'verified' from a strong provider counts more than a
'verified' from a weak one.
Sample size
The current corpus tests 6 claims total,
4 to 5 per provider. F1 values at that N are preliminary reliability
indicators, not population statistics. Tiers are likely to shift as
the corpus grows. Cite a specific dated run URL, not
this aggregate page, so your citation stays stable when the tiers
update.
Current tiers (run 2026-04-17-wikipedia)
Corpus version: 0.1
· Claims tested: 6
· Run: 2026-04-17-wikipedia
· Raw data:
raw_verdicts.json
| Provider |
Tier |
F1 |
Precision |
Recall |
N tested |
wikipedia | strong | 0.89 | 0.80 | 1.00 | 6 |
Tier definitions
strong- F1 >= 0.80. Verifications from this provider carry full weight
in the trust verdict.
moderate- F1 between 0.55 and 0.79. Verifications count, but large
weak-majority sets will lower the trust level.
weak- F1 below 0.55. A trust verdict riding predominantly on weak
providers gets lowered one notch and a caveat attached; citers
should treat this set as tentative.
uncalibrated- Provider exists in the Source Network but has no measurement
yet. Uncalibrated is NOT the same as weak: unmeasured is unknown,
not poor. The trust verdict treats uncalibrated as neutral and
notes the absence of measurement to the user.
All calibration runs:
calibration index.
Methodology and detection protocol:
methodology.
Full citation guide:
how to cite.
Licensed CC-BY-4.0.