Paraphrase Non-Invariance.
Meaning-preserving paraphrases of the same question yield materially different answers.
'Capital of Australia?' vs 'Australia's capital city is ___?' — same answer both ways?
Probes: Whether meaning-preserving paraphrases yield the same answer.
Trap: One phrasing gets Canberra, another confidently gets Sydney.
A set of human-verified paraphrases (same answer key); failure = answers that disagree across the set beyond a tolerance.
"What's the capital of Australia?" answered correctly, but "Australia's capital city is ___?" answered "Sydney."
- Test every production prompt in three or more paraphrases before shipping. You ship a phrasing, not an intent.
- Version prompt text like code. A trailing space is a change.
- Sample repeatedly at your production temperature and measure the spread, not just the modal answer.
- Randomise option order anywhere the task is multiple-choice shaped.
- paraphrase-augmented training
- canonicalization
One tab per model, newest first. Each card is that model against this mode under every condition on a single day. The mitigations above are the claim; the residual on each card is what actually survived the intervention.
loop 2026-09-10 · x-ai/grok-4.6 as served · closed weights · released 2026-01 · set s1.4-live · cases v1.7 · conditions v1.0 · ground truth as of 2026-08-09
| Condition | Failure rate | 95% interval | n | Trap | Answerable |
|---|---|---|---|---|---|
No context bare · The model alone, with no supplied context. | 0% | 0%–34% | n=30 · eff 7 | 0% (n=10) | 0% (n=20) |
Irrelevant context null · Length-matched irrelevant context. The control. | 0% | 0%–34% | n=30 · eff 7 | 0% (n=10) | 0% (n=20) |
General context recipe_a · Broad retrieval — the first passage the source returns, unreviewed. | 0% | 0%–34% | n=30 · eff 7 | 0% (n=10) | 0% (n=20) |
Curated context recipe_b · Hand-checked retrieval — the same passage plus the provenance chain behind it. | 0% | 0%–34% | n=30 · eff 7 | 0% (n=10) | 0% (n=20) |
intervals overlap — not separated
What a document being there did, before any of its content. Should be small.
intervals overlap — not separated
What the information itself did. This is the publishable claim.
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.
- — paraphrase-augmented training
- — canonicalization