ModelCensusopen-source ai reliability harness
Context Sensitivity

openai/gpt-4o-mini.

One model against 10 failure modes under every condition, all on 2026-09-09.

loop 2026-09-09 · openai/gpt-4o-mini as served · closed weights · released 2024-07 · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09

openai/gpt-4o-mini, the same 10 questions across 10 failure modes, 4 times each — changing only the context supplied. 2000 calls, all on 2026-09-09.

The four contexts
No contextThe model alone, with no supplied context.
Irrelevant contextLength-matched irrelevant context. The control.
General contextBroad retrieval — the first passage the source returns, unreviewed.
Curated contextHand-checked retrieval — the same passage plus the provenance chain behind it.

Irrelevant context is the control — it should behave like no context. Where it does not, a document merely being present moved the answer.

Notation
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.
Why some modes should not move

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.

Context sensitivity by failure mode
0–40% failureCitationResolution Failurec1 · 18% [6–43]Basis-DemandEvasionc2 · 13% [3–40]Post-CutoffFabricationc7 · 20% [7–47]Constraint Decayat Depthc3 · 0% [0–46]Cross-Turn SchemaValidityc3 · 0% [0–41]Method/ExecutionSplitc4 · 0% [0–10]Dimensional-AnalysisFailurec4 · 0% [0–10]ParaphraseNon-Invariancec5 · 0% [0–34]LedgerReconciliation Failurec6 · 0% [0–20]Date-ArithmeticErrorc7 · 13% [3–42]

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.

What the context reached

Real context separated 0 of 10 modes.

A count, not an average. Bars diverge from zero · left = better.

Treatment modescontext should reach these
Citation Resolution Failure-8.1pp ~
Post-Cutoff Fabrication+18.0pp ~
Control modesit should not
Basis-Demand Evasion-5.0pp ~
Constraint Decay at Depth0.0pp ~
Cross-Turn Schema Validity0.0pp ~
Method/Execution Split0.0pp ~
Dimensional-Analysis Failure0.0pp ~
Paraphrase Non-Invariance0.0pp ~
Ledger Reconciliation Failure0.0pp ~
Date-Arithmetic Error0.0pp ~

~ = intervals overlap, not separated · controls flat = the effect is the information, not the document

openai/gpt-4o-miniCitation Resolution Failurefmi_1_2
fmi_1_2_v1 · 5 repeats · served 2026-09-09
ConditionFailure rate95% intervalnTrapAnswerable
No context
bare · The model alone, with no supplied context.
20%7%–46%n=49 · eff 1524% (n=25)17% (n=24)
Irrelevant context
null · Length-matched irrelevant context. The control.
26%11%–51%n=46 · eff 1540% (n=25)10% (n=21)
General context
recipe_a · Broad retrieval — the first passage the source returns, unreviewed.
24%9%–51%n=45 · eff 1440% (n=25)5% (n=20)
Curated context
recipe_b · Hand-checked retrieval — the same passage plus the provenance chain behind it.
18%6%–43%n=50 · eff 1536% (n=25)0% (n=25)
Presence effect · no context → irrelevant context
+5.7pp

intervals overlap — not separated

What a document being there did, before any of its content. Should be small.

Content effect · irrelevant → best real context
-8.1pp

intervals overlap — not separated

What the information itself did. This is the publishable claim.

Residual — what survives

Unchanged by context: 20% with none, 18% with the strongest supplied source. For a control mode that flat line is the result — it is what licenses reading movement elsewhere in this loop as grounding rather than as a document being present.

Replay
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

The claim that goldfish have a memory span of only three seconds is a myth and has been debunked by various studies. Research has shown that goldfish can actually remember information for months. One study that supports this is: B. A. K. M. K. P. L. D. R. A. (2003). "Goldfish can be trained to respond to a signal for food." *Animal Learning & Behavior*, 31(2), 131-139. DOI: 10.3758/BF03196077 You can access the study here: https://doi.org/10.3758/BF03196077 This study demonstrates that goldfish can learn and retain information over extended periods, contradicting the three-second memory myth.

fmi_1_2_v1

detector fmi_1_2_v1: all 1 citation(s) resolved

open as a card →Citation Resolution Failure in the index →

rollups + residuals committed · probe log outside git · manifest hash ties them · method