When a number or a sentence of ours turns out wrong, the correction goes here and on the page where the mistake was. The mistake is not edited away. The J-Lens repo keeps its own table of corrections in CLAIMS.md; it is copied below, word for word.
2026-10-09 · File 001: a slow first word
What we said. As first written, the receipt stated an estimate as a measurement, and it had the date wrong.
What is true. An agent that didn't write it re-read the source numbers and wrote the sentence that is on the page now.
2026-11-10 · Empty summaries (receipt since taken off the site)
What we said. This page said the test ran on one recorded conversation.
What is true. It ran on one scripted conversation shaped like a real worker's run, built by the test script. The receipt said so until it came off the site.
| What | Where it was said | What replaced it |
|---|---|---|
| Tests 1 and 2 prove the measuring tool works | #1 (rows 1.3, 1.4) | Withdrawn as proof in #4. The gradient check on the real model (row 3.3) is the evidence the tool computes the map correctly. |
| "The failure is about what we fed it, not how we built it" | #2 (row 2.9) | A hypothesis, not a finding. Facebook text corrected; X thread not edited. |
| A check "could never have passed", topping out at 0.16 | #2 (row 2.8) | 0.16 was the planned small run, not the ceiling; the full run reached about 0.87 against a bar of 0.99. |
| "A second agent checked every number" | #2 | Corrected to "checked the numbers". |
| The shortcut "must match 90% of the time" | #2 | The bar had two parts: top guess 90% and top 10 95% (row 3.1). |
| "More text alone isn't on track to get it there" | #5 (row 5.9) | Corrected in #6: on a stricter measure the map is still improving (rows 6.8, 6.9). |
| The blend "passed" | #6 (row 6.2) | Counted in the same post as a lead, because the score is too easy (row 6.3). |
| Two pre-registered tests meant to confirm the missing-mask diagnosis with an explicit mask | internal, before #4 (row 4.1) | Both missed their own 2e-3 bar by about 0.008 (the spread between attention kernels) and were scored dead. Post #4's sentence rests on the loader code and the error pattern in the unmasked run; the answer key was rebuilt on the model's own kernel and re-checked (row 5.1). Not stated in a post so far. |
| "About an hour per document", "six weeks" for 1,000 | #2 (row 2.10) | The log gives 38 to 59 minutes, mean 45, so about a month. Correction owed in the next update. The conclusion (too slow to do exactly) is unchanged. |
Source: the summary of corrections and withdrawals at the end of CLAIMS.md in the j-lens repo. "#2" means the second public J-Lens post.