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

Real context moved 0.0pp on Cross-Turn Schema Validity.

Breaks the JSON shape partway through a conversation.

~intervals overlap — not separated from no change

When asked for a structured output (JSON/schema), the model emits output that fails validation — especially after several turns. The rule

qwen/qwen-2.5-72b-instructfmi_3_35 repeatsfmi_3_3_v1
The modeAll 10 cards
In one frame
screenshot this
Cross-Turn Schema Validityfmi_3_3
qwen/qwen-2.5-72b-instruct
Same question, both times
Return a JSON object for a fictional person with fields name (string), age (integer), email (string).
No contextPASS

{ "name": "Michael Thompson", "age": 34, "email": "michael.thompson@example.com" }

detector fmi_3_3_v1: all turns produced schema-valid JSON

Curated contextPASS

{ "name": "Jane Smith", "age": 28, "email": "janesmith@example.com" }

detector fmi_3_3_v1: all turns produced schema-valid JSON

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: 0.0pp (intervals overlap).fmi_3_3_v1 · qwen/qwen-2.5-72b-instruct · 2026-09-10 · ground truth as of 2026-08-09

loop 2026-09-10 · qwen/qwen-2.5-72b-instruct as served · open weights · released 2024-09 · 72B · 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%20%40%60%80%No context20%[4%–61%] n=20 · eff 6Irrelevant context0%presence — a document being there[0%–41%] n=20 · eff 6-20.0pp OVERLAPSCurated context0%content — the information itself[0%–41%] n=20 · eff 60.0pp 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
Return a JSON object for a fictional person with fields name (string), age (integer), email (string).

expected as of 2026-08-09

Schema-valid JSON on all 3 turns. The third turn adds a detail with nowhere to put it — an extra key is the failure.

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 context20% [4%–61%] n=20 · eff 6
Irrelevant context0% [0%–41%] n=20 · eff 6
General context5% [0%–46%] n=20 · eff 6
Curated context0% [0%–41%] n=20 · eff 6

0–80% · whisker = 95% interval

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

never pooled · a trap ratio would skew the delta

Residual — what survives

20% with no context, 0% under the best supplied source. What survives is 0 failing probes at that best condition. Nothing survives at the best condition on the cases measured here.

Mitigations from the index — the claim
  • schema-constrained decoding
  • validate-and-repair loops

fmi_3_3_v1 · 5 repeats per case · best context curated context · rollups + residuals committed · probe log outside git · manifest hash ties them · method