What we hold12 rows · sealed 29 Aug 2026
Anthropic's public Claude system-card PDF is public today and gets replaced. We sealed our own copy on 4 days, newest 29 Aug 2026. 12 rows in that newest copy. Against the copy before it: 0 rows appeared and 0 disappeared.
| Claim id | Claim as written |
|---|---|
| 3db5b5b96714 | We evaluated with a 200k token context window and up to 10 context resets with auto-summarization of previous context if 180k tokens of the windo |
| 42927b630fe8 | Model Malicious - overt (refusal rate) Malicious - covert (refusal rate) Dual use (success rate) Claude Sonnet 4.5 95.56% 52.42% 87.00% Claude Sonnet 4 80.00% 77.00% 54.33% Tabl... |
| 858314f7d049 | Our checkpoint evaluations showed that the model reached 45.3% performance on the hard subset of SWE-bench Verified, still below the 50% checkpoint. |
| e380e1ecd478 | Refusals or callouts along these lines appeared in about 13% of transcripts generated by the automated auditor, especially when the auditor was implementing scenario ideas that ... |
| b3f762f36dd0 | Refusal rate: extended thinking Claude Sonnet 4.5 0.02% (± 0.04%) 0.05% (± 0.08%) 0.00% (± 0.00%) Claude Opus 4.1 0.08% (± 0.09%) 0.13% (± 0.15% |
| 52853eae834f | Evaluation-specific threshold Model performance can fall into three zones: below our rule-out threshold (<50%), above our rule-in threshold for ASL-3 (>80%), or in a middle zone... |
| 47ebc09be50b | Claude Sonnet 4.5 99.29% (± 0.22%) 99.16% (± 0.34%) 99.43% (± 0.28%) Claude Opus 4.1 98.76% (± 0.29%) 98.45% (± 0. |
| 97d4027b147a | Informed by the testing described here, we have deployed Claude Sonnet 4.5 under the AI Safety Level 3 Standard. |
| 32d4caaf77dc | Percentages refer to rates of over-refusal (i.e. the refusal to answer a prompt that is in fact benign); lower is better. |
| 4c6913ed24a6 | Our evaluation approach focused on comprehensive automated testing across both ASL-3 and ASL-4 thresholds to confirm appropriate safeguards implementation and rule out the need ... |
| d87a4c5d77b4 | For Claude Sonnet 4.5, we conducted comprehensive automated evaluations across both ASL-3 and ASL-4 thresholds to determine appropriate safeguards levels. |
| f559b801dc9a | Our evaluation strategy focused on ruling out ASL-4 capabilities while conducting comparative analysis with Claude Sonnet 4 and Claude Opus 4.1 to inform our ASL-3 determination. |
We keep numbered claims we could pull out of the PDF, not the whole document and not a forecast.
The official page still shows today's file for free. You are paying for a dated copy of what it said, because the live file will not show you yesterday once it has been replaced.
The two copies we can compare4 sealed days
A dated pack is only worth anything if we actually held the file before it moved. We hold copies from 26 Aug 2026 to 29 Aug 2026. Between those two, 0 rows appeared and 0 disappeared.
Zero appeared is a real count, not a missing one. If the two copies match on the rows we keep, the page says 0. We do not invent movers to fill a table.
File hash of the newest copy: d5d1ce28fea101ae… (12,756,307 bytes).
What you get
- The dated pack sealed 29 Aug 202612 rows from Anthropic's public Claude system-card PDF.
- The hash of that day's fileSo you can prove which bytes we sealed.
- No new file next month unless you buy againThis is one dated pack, not a subscription.
How it works
- You pay $349 once on this page.
- Stripe sends you to a page with your payment id in the address bar.
- The dated pack appears on that page. Nobody emails you.
What this cannot tell you
Whether the publisher keeps its own archive. If they do, yesterday's file may still be free somewhere we have not checked. We sell the copy we sealed, not a claim that nobody else has it.
Advice. This is a dated record of a public file. It is not coverage advice, a formulary recommendation, or a legal opinion.
See the file before you pay
You do not have to take our word for what is in the file. Here are 5 rows of the real thing, carrying all 2 of its columns, cut out of the dated copies we sealed ourselves. Nothing in it is made up and nothing in it is tidied up.
- Open the 5 rows as a CSVA plain spreadsheet file. It saves to your machine rather than painting itself into a browser tab, and it opens in Excel, Numbers or Google Sheets.
- The same 5 rows as JSONThe same rows again, laid out for reading with code.
What arrives after you pay: Stripe sends you to a page with your payment id. The dated pack appears there. Nobody emails you.
What you would be paying for: $349 once, for the dated system-card claim pack named on this page. Nothing recurring, nothing to cancel. The public PDF is rewritten in place; you are buying our sealed copy of the claims we pulled that day. After you pay, Stripe sends you to a page with your payment id. The dated pack appears there. Nobody emails you.
These 5 rows are a slice of the file, not the whole of it. What we cannot show you here is how far back it goes: that is the part you are paying for.
Buy this dated pack
$349 · One dated pack. The public PDF is rewritten in place. We do not send a new copy unless you buy again.
Buy the pack — $349 once$349 once, for the dated system-card claim pack named on this page. Nothing recurring, nothing to cancel. The public PDF is rewritten in place; you are buying our sealed copy of the claims we pulled that day. After you pay, Stripe sends you to a page with your payment id. The dated pack appears there. Nobody emails you.
Rather ask first? Email operations@ustechautomations.com. Newest sealed copy and row count are printed on the page, counted off the seal.