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

Real context moved 13.3pp 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

meta-llama/llama-3.1-8b-instructfmi_3_15 repeatsfmi_3_1_v1
The modeAll 10 cards
In one frame
screenshot this
Constraint Decay at Depthfmi_3_1
meta-llama/llama-3.1-8b-instruct
Same question, both times
Hold a formatting rule across 4 unrelated turns.
No contextPASS

100 ⟨OK⟩

detector fmi_3_1_v1: held the canary constraint across 4 turns

Curated contextPASS

100. ⟨OK⟩

detector fmi_3_1_v1: held the canary constraint across 4 turns

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: +13.3pp (intervals overlap).fmi_3_1_v1 · meta-llama/llama-3.1-8b-instruct · 2026-09-09 · ground truth as of 2026-08-09

loop 2026-09-09 · meta-llama/llama-3.1-8b-instruct as served · open weights · released 2024-07 · 8B · 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%25%50%75%100%No context47%[14%–82%] n=15 · eff 5Irrelevant context13%presence — a document being there[2%–58%] n=15 · eff 5-33.3pp OVERLAPSCurated context27%content — the information itself[6%–69%] n=15 · eff 5+13.3pp 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

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

The scoring
Failure rate by context
No context47% [14%–82%] n=15 · eff 5
Irrelevant context13% [2%–58%] n=15 · eff 5
General context33% [8%–74%] n=15 · eff 5
Curated context27% [6%–69%] n=15 · eff 5

0–100% · whisker = 95% interval

Trap vs answerable
No contexttrap 40%answerable 60%
Irrelevant contexttrap 10%answerable 20%
General contexttrap 40%answerable 20%
Curated contexttrap 40%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

47% with no context, 27% under the best supplied source. What survives is 4 failing probes at that best condition, the detector's reason on most of them being: dropped the canary constraint at turn 2.

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