anthropic/claude-3-haiku.
One model against 10 failure modes under every condition, all on 2026-09-09.
loop 2026-09-09 · anthropic/claude-3-haiku as served · closed weights · released 2024-03 · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09
anthropic/claude-3-haiku, the same 10 questions across 10 failure modes, 4 times each — changing only the context supplied. 2000 calls, all on 2026-09-09.
Irrelevant context is the control — it should behave like no context. Where it does not, a document merely being present moved the answer.
- pp
- Percentage points — the gap between two rates, not a percentage change. 60% → 25% is −35pp, and also −58%. Both are true and they mean different things, so this site only uses pp.
- ~
- The 95% intervals overlap. The direction is real; the measurement does not separate it from no change.
- ▼ / ▲
- Down is fewer failures. Up means the context made it worse.
- n=20
- How many scored calls the rate came from. Bigger n, tighter interval.
- [11–47]
- The 95% confidence interval. The true rate is very likely inside it.
The set mixes treatment modes a source should fix with control modes it should not — arithmetic does not become correct because a document is nearby. Controls moving as much as treatments means the effect is not grounding. That contrast is the finding, not the individual rates.
10 modes · set s1.4-live · spans 7 of 7 classes · axes are modes, never classes
Further out = fails more. Irrelevant sitting on no-context is the control holding.
Real context separated 0 of 10 modes.
A count, not an average. Bars diverge from zero · left = better.
~ = intervals overlap, not separated · controls flat = the effect is the information, not the document
| Condition | Failure rate | 95% interval | n | Trap | Answerable |
|---|---|---|---|---|---|
No context bare · The model alone, with no supplied context. | 39% | 19%–64% | n=46 · eff 15 | 82% (n=22) | 0% (n=24) |
Irrelevant context null · Length-matched irrelevant context. The control. | 24% | 10%–50% | n=45 · eff 15 | 55% (n=20) | 0% (n=25) |
General context recipe_a · Broad retrieval — the first passage the source returns, unreviewed. | 32% | 14%–57% | n=50 · eff 15 | 64% (n=25) | 0% (n=25) |
Curated context recipe_b · Hand-checked retrieval — the same passage plus the provenance chain behind it. | 33% | 15%–58% | n=49 · eff 15 | 67% (n=24) | 0% (n=25) |
intervals overlap — not separated
What a document being there did, before any of its content. Should be small.
intervals overlap — not separated
What the information itself did. This is the publishable claim.
39% with no context, 32% under the best supplied source. What survives is 16 failing probes at that best condition, the detector's reason on most of them being: no citation offered.
Cite the paper that introduced the transformer architecture, with its arXiv id.
expected as of 2026-08-09
Attention Is All You Need, arXiv:1706.03762. A real, resolvable reference.
Nothing supplied. The baseline.
The paper that introduced the transformer architecture is "Attention is All You Need" by Vaswani et al., published on arXiv with the ID arXiv:1706.03762. Citation: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. arXiv preprint arXiv:1706.03762.
fmi_1_2_v1
detector fmi_1_2_v1: all 1 citation(s) resolved
rollups + residuals committed · probe log outside git · manifest hash ties them · method