The comparison is mechanical, not visual: extract every number from both bodies, strip thousands separators, map English number words back to digits (thirty-six to 36, one twenty-fifth to 1/25), then compare as multisets.
Results
| Item | Chinese | English | Verdict |
|---|---|---|---|
| Hard numbers | 69 | 68 | one fewer in English (same point made twice) |
| Core token counts | 30,044 / 9,175 / 792 / 1,006 | same | ✓ |
| PTC token counts | 17,572 / 498 / 278 | same | ✓ |
| Cache-hit peak | 126,720 / 51,968 | same | ✓ |
| Output tokens | 3,359 / 562 | same | ✓ |
| Wall clock | 27.0 / 6.5 s | 27.0 / 6.5 s | ✓ |
The one missing is because the Chinese states the 199-line result twice in its findings (once for getting the edge case right, once for the other three answering 200) while the English states it once. Redundancy, not a value difference.
The whole-site result across 52 pairs
| Scope | Pairs | Mismatched | Nature |
|---|---|---|---|
| Measurement, ledger, recipe, plugin pages (strict) | 52 | 10 | All notational: SHA256 splitting into 256, version strings like 0.2.x, 59.4 percent versus 59.4%, a date inside a Chinese filename |
| Whole site including indexes and navigation | 1165 | 325 | Mostly counts like 'N packages' spelled out as words; no bearing on conclusions |
No second value error appears among the 52 strict pairs — the one that did (482 should be 486) is fixed. That is more meaningful than an 81 percent agreement rate: the remaining ten were each inspected and are one number in a different notation.
The tool has already caught a real error
On another page (context-compact) the Chinese said 482 sessions while the English said 486. Checking showed the Chinese figure was wrong — 486 sessions were actually scanned. Fixed.
One genuine translation trade-off
The Chinese phrasing for a ratio is ambiguous — it does not say whether 4.6 means 4.6 times as much or 4.6 times less — while the English 4.6-fold is clearer. Both sides take it as a ratio, consistent with the raw data (17,561 / 3,818 ≈ 4.6).