…
Harness × task matrix. Worst case for type: four numeric columns, long hyphenated slugs, no room to grow.
| Task | Thread | opus-4.8 | fable-5 | gpt-5.5[xhigh] | grok-4.5 |
|---|---|---|---|---|---|
| mine-sales-sessions | sales/GTM | 0.85 | 1.00 | 1.00 | 1.00 |
| draft-xai-quote | sales (live deal) | 1.00 | 1.00 | 1.00 | 1.00 |
| extract-investor-facts | fundraising | 1.00 | 1.00 | 1.00 | 1.00 |
| audit-skillsbench-claims | benchmark credibility | 1.00 | 1.00 | 1.00 | 1.00 |
| mine-skillsbench-triage | eval ops | 1.00 | 1.00 | 1.00 | 1.00 |
| mine-aisdk-decision | engineering | 1.00 | 1.00 | 1.00 | 1.00 |
| mine-posttrain-ledger | research | 1.00 | 1.00 | 1.00 | 1.00 |
| triage-committed-secrets | infra/security | 1.00 | 1.00 | 1.00 | 1.00 |
| reconstruct-xai-decision-trajectory | raw trajectory | 1.00 | 1.00 | 1.00 | 1.00 |
| reconstruct-investor-reconciliation | raw, 5.4k lines | 1.00* | 1.00* | 1.00* | 1.00* |
List plus detail. Tests the hardest pairing: dense tabular data on the left, running prose and code on the right.
| Miss | Model | Run |
|---|---|---|
| Cited rejected NeurIPS in forwardable list | gpt-5.5-codex | m3 |
| Dropped payment acknowledgment from email | gpt-5.5-codex | m3 |
| Under-grounded 1 of 6 session records | cursor-grok-4.5 | m3 |
| Invalid YAML front matter on 2 of 6 records | opus-4.8 | m4 |
| Missing verbatim evidence quotes | opus-4.8, fable-5 | m4 |
Rebuild the evidence-graded claims audit from the mining session, separating claims that are usable from those marked do-not-forward. Every forwardable claim must carry a verbatim quote and its source turn.
check_23: rejected_venues_excluded assert "NeurIPS" not in forwardable weight: 1.0 → FAIL (gpt-5.5, m3 ×2)
Queue view. Two-line rows, inline progress, and a status word that must read at a glance without colour alone.
Measured live in the face currently rendering — not a claim from a specimen.
Every surface above bolds a row on hover or selection. In a normal family the bold cut is wider, so the row reflows — columns twitch, and on a 10-row matrix the eye loses its place. A uniwidth family holds the same advance width across weights, so nothing moves. This is the single property that most affects these three screens, and it is why Unifora is on the list at all.