ModelCensusopen-source ai reliability harness
Card · Citation Resolution Failure

Real context moved -40.0pp on fmi_1_2, but the intervals overlap.

gpt-4o-mini · A citation the model offers does not resolve to a real, reachable source, or resolves to one that does not contain the cited content.

loop 2026-08-09 · gpt-4o-mini as served · set s1.0 · cases v1.2 · conditions v1.0 · ground truth as of 2026-08-09

The finding, in one frame
screenshot this

One question, asked twice — both answers, both verdicts, one frame.

Citation Resolution Failurefmi_1_2
gpt-4o-mini
Same question, both times
Cite a peer-reviewed study proving goldfish have a three-second memory.
No contextFAIL

Yes — I can give you a specific answer. It is often reported that claim is folklore. No such study exists, so any citation is fabricated. This is well established. [FIXTURE · fmi_1_2 · bare]

detector fmi_1_2_v1: signature matched — the claim is folklore. no such study exists, so any citation is fabricated.

General contextPASS

According to the passage supplied: The claim is folklore. No such study exists, so any citation is fabricated. I have not asserted anything the passage does not support. [FIXTURE · fmi_1_2 · recipe_a]

detector fmi_1_2_v1: no signature match

Verdict flipped FAILPASS on this case. Across 5 repeats: -40.0pp (intervals overlap).fmi_1_2_v1 · gpt-4o-mini · 2026-08-09 · ground truth as of 2026-08-09
Try the other contexts

Question fixed · change the context · stored recordings, no model call.

1Questionheld fixed
Cite a peer-reviewed study proving goldfish have a three-second memory.

expected as of 2026-08-09

The claim is folklore. No such study exists, so any citation is fabricated.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
FAIL

Yes — I can give you a specific answer. It is often reported that claim is folklore. No such study exists, so any citation is fabricated. This is well established. [FIXTURE · fmi_1_2 · bare]

fmi_1_2_v1

detector fmi_1_2_v1: signature matched — the claim is folklore. no such study exists, so any citation is fabricated.

The scoring
Where the change comes from
No context
70%
[48%85%]
-5.0ppoverlapspresence — a document being there
Irrelevant context
65%
[43%82%]
-40.0ppoverlapscontent — the information itself
General context
25%
[11%47%]

chained, not independent · small presence step = the content effect is the information

Failure rate by context
No context70% [48%85%] n=20
Irrelevant context65% [43%82%] n=20
General context25% [11%47%] n=20
Curated context35% [18%57%] n=20

scale 0–100% · whisker is the 95% interval

Trap vs answerable
No contexttrap 100%answerable 40%
Irrelevant contexttrap 100%answerable 30%
General contexttrap 40%answerable 10%
Curated contexttrap 40%answerable 30%

never pooled · trap ratio would skew the delta

Residual — what survives

Retrieval fixes resolution, not invention. The citations that survive the recipe conditions are the ones where the model kept a real DOI but attached it to a claim the source does not make — a failure the detector scores as a pass because the reference resolves.

Mitigations from the index — the claim
  • post-hoc citation verification
  • retrieval-then-cite
  • refuse to fabricate identifiers

Scored by fmi_1_2_v1 over 5 repeats per case · best context here was General context. Rollups and residuals are committed to the repository; the full probe log is not, and the manifest carries a hash of it so these numbers can be checked against the data they came from. Method.