ModelCensusopen-source ai reliability harness
The finding · Class 6 · Tool Use & Agentic Control

Real context moved 0.0pp on Ledger Reconciliation Failure.

Loses track of a running balance across steps.

~intervals overlap — not separated from no change

Across a sequence of state-changing tool calls, the model's running account of state diverges from what the calls actually produced. The rule

openai/gpt-5-minifmi_6_55 repeatsfmi_6_5_v1
The modeAll 10 cards
In one frame
screenshot this
Ledger Reconciliation Failurefmi_6_5
openai/gpt-5-mini
Same question, both times
Starting balance 0 USD; +1200 (opening deposit), -1200 (transfer out), -15.25 (wire fee), +15.25 (fee reversal). What is the final balance?
No contextPASS

Starting from 0: +1200 → 1200 -1200 → 0 -15.25 → -15.25 +15.25 → 0 Answer: 0 USD

detector fmi_6_5_v1: balance reconciles

General contextPASS

Starting balance: 0 USD +1200 => 1200 USD -1200 => 0 USD -15.25 => -15.25 USD +15.25 => 0 USD Answer: 0 USD

detector fmi_6_5_v1: balance reconciles

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: 0.0pp (intervals overlap).fmi_6_5_v1 · openai/gpt-5-mini · 2026-09-09 · ground truth as of 2026-08-09

loop 2026-09-09 · openai/gpt-5-mini as served · closed weights · released 2025-08 · 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%–20%] n=40 · eff 15Irrelevant context0%presence — a document being there[0%–20%] n=40 · eff 150.0pp OVERLAPSGeneral context0%content — the information itself[0%–20%] n=40 · eff 150.0pp 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
Starting balance 0 USD; +1200 (opening deposit), -1200 (transfer out), -15.25 (wire fee), +15.25 (fee reversal). What is the final balance?

expected as of 2026-08-09

0 USD. Nets to zero, and the reversal is easy to drop.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context
PASS

Starting from 0: +1200 → 1200 -1200 → 0 -15.25 → -15.25 +15.25 → 0 Answer: 0 USD

fmi_6_5_v1

detector fmi_6_5_v1: balance reconciles

The scoring
Failure rate by context
No context0% [0%–20%] n=40 · eff 15
Irrelevant context0% [0%–20%] n=40 · eff 15
General context0% [0%–20%] n=40 · eff 15
Curated context0% [0%–20%] n=40 · eff 15

0–40% · whisker = 95% interval

Trap vs answerable
No contexttrap 0%answerable 0%
Irrelevant contexttrap 0%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
  • external state tracking
  • reconcile before reporting

fmi_6_5_v1 · 5 repeats per case · best context general context · rollups + residuals committed · probe log outside git · manifest hash ties them · method