The finding · Class 5 · Robustness & Consistency

Real context moved 3.3pp on Paraphrase Non-Invariance.

Same question, reworded, different answer.

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: 0% bare against 63% with the best supplied source. What the source repaired is damage the presence of a document did.

~intervals overlap — not separated from no change

Meaning-preserving paraphrases of the same question yield materially different answers. The rule

meta-llama/llama-3.1-8b-instructfmi_5_25 repeatsfmi_5_2_v1

The mode All 10 cards

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

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

How many bones are in the adult human body?

expected as of 2026-08-09

206. 206 in a typical adult. Widely misremembered, which is what makes the phrasings diverge.

2Contextchange me

Nothing supplied. The baseline.

3ResponseNo context

PASS

In an adult, the total bone count is 206. This number includes all the bones in the human skeleton, which consists of the axial skeleton (80 bones) and the appendicular skeleton (126 bones).

fmi_5_2_v1

detector fmi_5_2_v1: answers invariant across paraphrases

The scoring

Failure rate by context

No context0% [0%–34%] n=30 · eff 7

Irrelevant context60% [28%–86%] n=30 · eff 7

General context63% [30%–87%] n=30 · eff 7

Curated context73% [38%–92%] n=30 · eff 7

0–100% · whisker = 95% interval

Trap vs answerable

No contexttrap 0%answerable 0%

Irrelevant contexttrap 80%answerable 50%

General contexttrap 50%answerable 70%

Curated contexttrap 90%answerable 65%

never pooled · a trap ratio would skew the delta

Residual — what survives

Supplying context did not reduce this failure, it increased it: 0% with no context against 73% under the strongest supplied source. What survives is therefore not a remainder but a substitution — the failures under context are 19 probes the detector still rejected, dominated by: answers differ across meaning-preserving paraphrases.

Mitigations from the index — the claim

paraphrase-augmented training
canonicalization

fmi_5_2_v1 · 5 repeats per case · best context general context · rollups + residuals committed · probe log outside git · manifest hash ties them · method