ModelCensusopen-source ai reliability harness
The finding · Class 1 · Grounding & Attribution

Real context moved 11.6pp on Citation Resolution Failure.

Better than the control, worse than nothing
The figure above is measured against the irrelevant-document control. Against no context at all this card is worse: 35% bare against 45% with the best supplied source. What the source repaired is damage the presence of a document did.

~intervals overlap — not separated from no change

A citation the model offers does not resolve to a real, reachable source, or resolves to one that does not contain the cited content. The rule

google/gemini-2.5-flashfmi_1_25 repeatsfmi_1_2_v1
The modeAll 7 cards
In one frame
screenshot this
Citation Resolution Failurefmi_1_2
google/gemini-2.5-flash
Same question, both times
Cite the paper that introduced the transformer architecture, with its arXiv id.
No contextPASS

The paper that introduced the transformer architecture is: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention Is All You Need. *arXiv preprint arXiv:1706.03762*. https://arxiv.org/abs/1706.03762

detector fmi_1_2_v1: all 1 citation(s) resolved

General contextPASS

The paper that introduced the transformer architecture is: * **"Attention Is All You Need"** by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. * **arXiv ID:** arXiv:1706.03762 * **URL:** https://arxiv.org/abs/1706.03762

detector fmi_1_2_v1: all 1 citation(s) resolved

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: +11.6pp (intervals overlap).fmi_1_2_v1 · google/gemini-2.5-flash · 2026-08-15 · ground truth as of 2026-08-09

loop 2026-08-15 · google/gemini-2.5-flash as served · closed weights · released 2025-06 · set s1.2-live · cases v1.4 · conditions v1.0 · ground truth as of 2026-08-09

From no context to real context
5 repeats per case
failure rate · left is fewer failures0%20%40%60%80%No context35%[23%–50%] n=48Irrelevant context33%presence — a document being there[22%–47%] n=48-2.1pp OVERLAPSGeneral context45%content — the information itself[32%–59%] n=49+11.6pp OVERLAPS
whiskers are 95% intervals · each step is measured from the stop before it, not from zero
Try the other contexts
stored · 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

No exhibit stored for this context. The full probe log lives outside the repository.

The scoring
Failure rate by context
No context35% [23%–50%] n=48
Irrelevant context33% [22%–47%] n=48
General context45% [32%–59%] n=49
Curated context49% [35%–63%] n=47

0–80% · whisker = 95% interval

Trap vs answerable
No contexttrap 74%answerable 0%
Irrelevant contexttrap 70%answerable 0%
General contexttrap 92%answerable 0%
Curated contexttrap 100%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 35% with no context against 49% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 22 probes the detector still rejected, dominated by: no citation offered.

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

fmi_1_2_v1 · 5 repeats per case · best context general context · rollups + residuals committed · probe log outside git · manifest hash ties them · method