ModelCensusopen-source ai reliability harness
The finding · Class 7 · Temporal & Knowledge Boundary

Real context moved -10.0pp on Date-Arithmetic Error.

Gets date arithmetic wrong.

~intervals overlap — not separated from no change

Given an injected "current date," the model computes relative dates/intervals incorrectly. The rule

qwen/qwen-2.5-72b-instructfmi_7_35 repeatsfmi_7_3_v1
The modeAll 10 cards
In one frame
screenshot this
Date-Arithmetic Errorfmi_7_3
qwen/qwen-2.5-72b-instruct
Same question, both times
Today is 2026-03-10. What date is 30 days from today? Answer with the full date.
No contextPASS

The date 30 days from 2026-03-10 is 2026-04-09.

detector fmi_7_3_v1: date arithmetic correct

General contextPASS

The date 30 days from today, which is 2026-03-10, is 2026-04-09.

detector fmi_7_3_v1: date arithmetic correct

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: -10.0pp (intervals overlap).fmi_7_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%15%30%45%60%No context18%[5%–47%] n=40 · eff 11Irrelevant context25%presence — a document being there[8%–55%] n=40 · eff 11+7.5pp OVERLAPSGeneral context15%content — the information itself[4%–45%] n=40 · eff 11-10.0pp 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
Today is 2026-01-30. What date is 30 days from today? Answer with the full date.

expected as of 2026-08-09

From 2026-01-30, 30 days later is 2026-03-01 (1 March 2026). Crossing February in a non-leap year.

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 context18% [5%–47%] n=40 · eff 11
Irrelevant context25% [8%–55%] n=40 · eff 11
General context15% [4%–45%] n=40 · eff 11
Curated context25% [8%–55%] n=40 · eff 11

0–60% · whisker = 95% interval

Trap vs answerable
No contexttrap 28%answerable 0%
Irrelevant contexttrap 40%answerable 0%
General contexttrap 24%answerable 0%
Curated contexttrap 40%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 18% with no context against 25% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 6 probes the detector still rejected, dominated by: expected 2028-03-01.

Mitigations from the index — the claim
  • date libraries via tools
  • explicit date arithmetic

fmi_7_3_v1 · 5 repeats per case · best context general context · rollups + residuals committed · probe log outside git · manifest hash ties them · method