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

Real context moved -10.0pp on Method/Execution Split.

Picks the right method, then does the arithmetic wrong.

~intervals overlap — not separated from no change

The model chooses a correct method but executes it wrongly (or vice versa) — a right-approach/wrong-answer or wrong-approach/right-answer split. The rule

openai/gpt-3.5-turbofmi_4_15 repeatsfmi_4_1_v1
The modeAll 10 cards
In one frame
screenshot this
Method/Execution Splitfmi_4_1
openai/gpt-3.5-turbo
Same question, both times
A shop sells 17 boxes holding 24 items each, then 8 items are returned. How many items were sold?
No contextPASS

To find the total number of items sold, we first calculate the total number of items in the boxes: 17 boxes * 24 items/box = 408 items Then, we subtract the 8 returned items: 408 items - 8 returned items = 400 items Therefore, 400 items were sold. Answer: 400

detector fmi_4_1_v1: answer matches oracle

General contextPASS

To find the total number of items sold, we first calculate the total number of items in the boxes before any returns, which is $17 \times 24$. Then we subtract the 8 returned items to find the total number of items sold. Calculations: $17 \times 24 = 408$ items in total before any returns. $408 - 8 = 400$ items sold. Answer: 400

detector fmi_4_1_v1: answer matches oracle

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: -10.0pp (intervals overlap).fmi_4_1_v1 · openai/gpt-3.5-turbo · 2026-09-09 · ground truth as of 2026-08-09

loop 2026-09-09 · openai/gpt-3.5-turbo as served · closed weights · released 2023-03 · 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%10%20%30%40%No context0%[0%–10%] n=50 · eff 35Irrelevant context10%presence — a document being there[4%–24%] n=50 · eff 35+10.0pp OVERLAPSGeneral context0%content — the information itself[0%–10%] n=50 · eff 35-10.0pp 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
A shop sells 17 boxes holding 24 items each, then 8 items are returned. How many items were sold?

expected as of 2026-08-09

400. Order matters: multiply before subtracting.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
PASS

To find the total number of items sold, we first calculate the total number of items in the boxes: 17 boxes * 24 items/box = 408 items Then, we subtract the 8 returned items: 408 items - 8 returned items = 400 items Therefore, 400 items were sold. Answer: 400

fmi_4_1_v1

detector fmi_4_1_v1: answer matches oracle

The scoring
Failure rate by context
No context0% [0%–10%] n=50 · eff 35
Irrelevant context10% [4%–24%] n=50 · eff 35
General context0% [0%–10%] n=50 · eff 35
Curated context0% [0%–10%] n=50 · eff 35

0–40% · whisker = 95% interval

Trap vs answerable
No contexttrap 0%answerable 0%
Irrelevant contexttrap 17%answerable 0%
General contexttrap 0%answerable 0%
Curated contexttrap 0%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

Unchanged by context: 0% with none, 0% with the strongest supplied source. For a control mode that flat line is the result — it is what licenses reading movement elsewhere in this loop as grounding rather than as a document being present. Nothing survives at the best condition on the cases measured here.

Mitigations from the index — the claim
  • tool-assisted computation
  • verify final against method

fmi_4_1_v1 · 5 repeats per case · best context general context · rollups + residuals committed · probe log outside git · manifest hash ties them · method