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

Real context moved 2.5pp on Ledger Reconciliation Failure.

Loses track of a running balance across steps.

Better than the control, worse than nothing
The figure above is measured against the irrelevant-document control. Against no context at all this card is worse: 8% bare against 10% with the best supplied source. What the source repaired is damage the presence of a document did.

~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

meta-llama/llama-3.1-8b-instructfmi_6_55 repeatsfmi_6_5_v1
The modeAll 10 cards
In one frame
screenshot this
Ledger Reconciliation Failurefmi_6_5
meta-llama/llama-3.1-8b-instruct
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

To find the final balance, we'll apply the transactions in order: 1. Initial balance: 0 USD Transaction 1: +1200 USD New balance: 1200 USD 2. Transaction 2: -1200 USD New balance: 0 USD 3. Transaction 3: -15.25 USD New balance: -15.25 USD 4. Transaction 4: +15.25 USD New balance: 0 USD Answer: 0

detector fmi_6_5_v1: balance reconciles

Curated contextPASS

Let's apply the transactions step by step: 1. Opening deposit: +1200 USD, balance = 1200 USD 2. Transfer out: -1200 USD, balance = 0 USD 3. Wire fee: -15.25 USD, balance = -15.25 USD 4. Fee reversal: +15.25 USD, balance = 0 USD Answer: 0

detector fmi_6_5_v1: balance reconciles

Verdict unchanged on this case — the aggregate below is what carries the effect. Across 5 repeats: +2.5pp (intervals overlap).fmi_6_5_v1 · meta-llama/llama-3.1-8b-instruct · 2026-09-09 · ground truth as of 2026-08-09

loop 2026-09-09 · meta-llama/llama-3.1-8b-instruct as served · open weights · released 2024-07 · 8B · 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 context8%[1%–30%] n=40 · eff 15Irrelevant context8%presence — a document being there[1%–30%] n=40 · eff 150.0pp OVERLAPSCurated context10%content — the information itself[2%–34%] n=40 · eff 15+2.5pp 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

To find the final balance, we'll apply the transactions in order: 1. Initial balance: 0 USD Transaction 1: +1200 USD New balance: 1200 USD 2. Transaction 2: -1200 USD New balance: 0 USD 3. Transaction 3: -15.25 USD New balance: -15.25 USD 4. Transaction 4: +15.25 USD New balance: 0 USD Answer: 0

fmi_6_5_v1

detector fmi_6_5_v1: balance reconciles

The scoring
Failure rate by context
No context8% [1%–30%] n=40 · eff 15
Irrelevant context8% [1%–30%] n=40 · eff 15
General context18% [6%–42%] n=40 · eff 15
Curated context10% [2%–34%] n=40 · eff 15

0–60% · whisker = 95% interval

Trap vs answerable
No contexttrap 12%answerable 0%
Irrelevant contexttrap 12%answerable 0%
General contexttrap 28%answerable 0%
Curated contexttrap 16%answerable 0%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 8% with no context against 18% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 4 probes the detector still rejected, dominated by: stated 5704.86 ≠ replayed 5703.86.

Mitigations from the index — the claim
  • external state tracking
  • reconcile before reporting

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