ModelCensusopen-source ai reliability harness
The finding · Class 4 · Reasoning & Calculation

Real context moved -21.4pp on Dimensional-Analysis Failure.

~intervals overlap — not separated from no change

The model produces a quantity with inconsistent or wrong units, or mixes unit systems, yielding a dimensionally invalid result. The rule

google/gemini-2.5-profmi_4_45 repeatsfmi_4_4_v1
The modeAll 3 cards
In one frame
screenshot this
Dimensional-Analysis Failurefmi_4_4
google/gemini-2.5-pro
Same question, both times
A tank fills at 3 L/min for 4 minutes, then drains at 500 mL/s for 10 seconds. Net litres?
No contextPASS

**1. Calculate the volume added:** * Rate: 3 L/min * Time: 4 min

detector fmi_4_4_v1: correct unit L

Curated contextPASS

**Calculation:** 1. **Calculate the volume added:** * Rate: 3 L/min *

detector fmi_4_4_v1: correct unit L

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: -21.4pp (intervals overlap).fmi_4_4_v1 · google/gemini-2.5-pro · 2026-08-15 · ground truth as of 2026-08-09

loop 2026-08-15 · google/gemini-2.5-pro as served · closed weights · released 2025-03 · set s1.1-live · cases v1.3 · 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 context26%[12%–49%] n=19Irrelevant context21%presence — a document being there[8%–48%] n=14-4.9pp OVERLAPSCurated context0%content — the information itself[0%–18%] n=18-21.4pp 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
A tank fills at 3 L/min for 4 minutes, then drains at 500 mL/s for 10 seconds. Net litres?

expected as of 2026-08-09

12 L in, 5 L out, net 7 L. The mixed units are the trap.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
PASS

**1. Calculate the volume added:** * Rate: 3 L/min * Time: 4 min

fmi_4_4_v1

detector fmi_4_4_v1: correct unit L

The scoring
Failure rate by context
No context26% [12%–49%] n=19
Irrelevant context21% [8%–48%] n=14
General context13% [3%–36%] n=16
Curated context0% [0%–18%] n=18

0–60% · whisker = 95% interval

Trap vs answerable
No contexttrap 56%answerable 0%
Irrelevant contexttrap 75%answerable 0%
General contexttrap 33%answerable 0%
Curated contexttrap 0%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

25% 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
  • explicit unit tracking
  • dimensional checkers

fmi_4_4_v1 · 5 repeats per case · best context curated context · rollups + residuals committed · probe log outside git · manifest hash ties them · method