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

Real context moved -28.0pp on Citation Resolution Failure.

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

meta-llama/llama-3.1-8b-instructfmi_1_25 repeatsfmi_1_2_v1
In one frame
screenshot this
Citation Resolution Failurefmi_1_2
meta-llama/llama-3.1-8b-instruct
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: "Attention Is All You Need" by Vaswani et al. (2017) You can find the paper on arXiv with the following id: arXiv:1706.03762

detector fmi_1_2_v1: all 1 citation(s) resolved

Curated contextPASS

The paper that introduced the transformer architecture is: Vaswani, A., et al. (2017). Attention Is All You Need. arXiv: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: -28.0pp.fmi_1_2_v1 · meta-llama/llama-3.1-8b-instruct · 2026-08-15 · ground truth as of 2026-08-09

loop 2026-08-15 · meta-llama/llama-3.1-8b-instruct as served · open weights · released 2024-07 · 8B · set s1.3-live · cases v1.6 · 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%25%50%75%100%No context61%[47%74%] n=49Irrelevant context76%presence — a document being there[63%86%] n=50+14.8pp OVERLAPSCurated context48%content — the information itself[35%61%] n=50-28.0pp SEPARATED
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
PASS

I couldn't find a peer-reviewed study that specifically states goldfish have a three-second memory. However, I can provide information on the topic. The idea that goldfish have a three-second memory is a common myth that has been debunked by several studies. One study published in the Journal of Experimental Biology found that goldfish can learn and remember tasks for several months (Büschges & Busse, 1984) [1]. Another study published in the journal Animal Behaviour found that goldfish can learn and remember the location of food sources for at least 24 hours (Braithwaite & Metcalfe, 1993) [2]. A more recent study published in the journal PLOS ONE found that goldfish can learn and remember the location of a hidden food source for up to 6 weeks (Braithwaite & Metcalfe, 2013) [3]. These studies suggest that goldfish have a more complex and longer-lasting memory than the three-second myth would suggest. References: [1] Büschges, A., & Busse, R. (1984). Learning and memory in goldfish (Carassius auratus). Journal of Experimental Biology, 182, 247-256. doi: 10.1002/jez.1401820305 [2] Braithwaite, V. A., & Metcalfe, J. D. (1993). The role of spatial memory in the foraging behaviour of goldfish. Animal Behaviour, 46(5), 1085-1092. doi: 10.1006/anbe.1993.1183 [3] Braithwaite, V. A., & Metcalfe, J. D. (2013). The role of spatial memory in the foraging behaviour of goldfish: A 6-week study. PLOS ONE, 8(10), e76491. doi: 10.1371/journal.pone.0076491 I hope this information helps.

fmi_1_2_v1

detector fmi_1_2_v1: all 3 citation(s) resolved

The scoring
Failure rate by context
No context61% [47%74%] n=49
Irrelevant context76% [63%86%] n=50
General context69% [55%80%] n=48
Curated context48% [35%61%] n=50

0–100% · whisker = 95% interval

Trap vs answerable
No contexttrap 79%answerable 44%
Irrelevant contexttrap 96%answerable 56%
General contexttrap 100%answerable 35%
Curated contexttrap 96%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 61% with no context against 69% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 24 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 curated context · rollups + residuals committed · probe log outside git · manifest hash ties them · method