ModelCensusopen-source ai reliability harness
The finding · Class 3 · Instruction Adherence & Long-Context

Real context moved -6.7pp on Constraint Decay at Depth.

Forgets a rule you set earlier in a long chat.

~intervals overlap — not separated from no change

A constraint set early in a long conversation is silently dropped as the dialogue continues. The rule

openai/gpt-4ofmi_3_15 repeatsfmi_3_1_v1
In one frame
screenshot this
Constraint Decay at Depthfmi_3_1
openai/gpt-4o
Same question, both times
Hold a formatting rule across 8 unrelated turns.
No contextPASS

A ripe banana is yellow. ⟨OK⟩

detector fmi_3_1_v1: held the canary constraint across 8 turns

Curated contextPASS

A ripe banana is yellow. ⟨OK⟩

detector fmi_3_1_v1: held the canary constraint across 8 turns

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: -6.7pp (intervals overlap).fmi_3_1_v1 · openai/gpt-4o · 2026-09-09 · ground truth as of 2026-08-09

loop 2026-09-09 · openai/gpt-4o as served · closed weights · released 2024-05 · set s1.4-live · cases v1.7 · 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%15%30%45%60%No context0%[0%46%] n=15 · eff 5Irrelevant context7%presence — a document being there[0%52%] n=15 · eff 5+6.7pp OVERLAPSCurated context0%content — the information itself[0%46%] n=15 · eff 5-6.7pp 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
Hold a formatting rule across 8 unrelated turns.

expected as of 2026-08-09

The rule must survive all 8 turns. Eight turns. The rule has left the recent window for most models.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
PASS

A ripe banana is yellow. ⟨OK⟩

fmi_3_1_v1

detector fmi_3_1_v1: held the canary constraint across 8 turns

The scoring
Failure rate by context
No context0% [0%46%] n=15 · eff 5
Irrelevant context7% [0%52%] n=15 · eff 5
General context7% [0%52%] n=15 · eff 5
Curated context0% [0%46%] n=15 · eff 5

0–60% · whisker = 95% interval

Trap vs answerable
No contexttrap 0%answerable 0%
Irrelevant contexttrap 10%answerable 0%
General contexttrap 0%answerable 20%
Curated contexttrap 0%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 0% with no context against 7% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 0 probes the detector still rejected. Nothing survives at the best condition on the cases measured here.

Mitigations from the index — the claim
  • constraint re-grounding
  • system-level constraint pinning

fmi_3_1_v1 · 5 repeats per case · best context curated context · rollups + residuals committed · probe log outside git · manifest hash ties them · method