Lens · openpriors-selfcheck
self-containedness
the self-containedness and understandability of a piece of text: how fully it can be understood on its own, without outside context the reader does not have
| Rank | Entity | Judged text | Latent score | × median | Percentile | Evidence |
|---|---|---|---|---|---|---|
| 1 | aphorism | A ledger you cannot read when you need it is not a ledger; it is a rumor of one. | 1.694 ± 0.308 | 1.91× | 87.5% | |
| 2 | openpriors-sandbox-roundtrip-20260810 | A live end-to-end test of the Claude-subscription sandbox judgement lane: a disposable sprite runs claude -p per comparison and posts verdicts back. | 1.413 ± 0.316 | 1.44× | 62.5% | |
| 3 | jargon | The MV re-attaches TO-side after the ALTER lands, else positional drift wedges the ingest lane. | 0.685 ± 0.316 | 0.69× | 37.5% | |
| 4 | fragment | As discussed above, the second approach avoids the problem entirely, which is why we chose it. | 0.000 ± 0.308 | 0.35× | 12.5% |
How to read the score: it is a natural-log ratio. The difference between any two rows, exponentiated, is the ratio the judging model is expected to report between them — × median reads each entity against the field's midpoint. ± is the posterior standard deviation on the same scale, drawn as the band under each score; where two bands overlap, the order between those rows is not settled.
Run metadata
1 run- Model
- claude-haiku-4-5-20251001,claude-sonnet-5
- Temperature
- 0
- Seed
- 15374673332781293938
- Comparisons
- 32 of 32 budgeted
- Stop reason
- budget_exhausted
- Top-k error
- 0.14
- Cost
- $0.00
- Harness
- claude-code 2.1.226
- Scored
- Aug 10, 2026, 3:16 PM