Accuracy: the dust rule for peel steps, and one sign as a weaker lead A peel step now has to be at least 1,000 satoshis: about a fifth of the steps in detected chains peeled exactly the dust limit (546 satoshis beside an OP_RETURN output, 330 satoshis), the look of token and inscription transfers. On a sample the change was not made on it dropped 187 of 502 steps in chains (181 dust peels, 6 collateral) and added none. A medium change guess on one sign was right 89.5% of the time against 99.5% on two signs; capping those at low waits for real-wallet data. Analysis and findings words the peel rule as "not dust-sized". Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Accuracy: the CoinJoin change and what a holdout showed; the PEEL reading The COINJOIN? rule now counts the different addresses an equal amount pays, not the outputs. A second sample it was not made on dropped 2 of 12 flags (one address paid repeatedly), kept 10 including all three Whirlpool-shaped transactions, and added none. The PEEL bullet says what the longest real chains looked like and that the default is unchanged. Analysis and findings words the rule as "to different addresses". Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Accuracy: add what real traffic and a documented case showed How often each flag fires on a sample of the real chain (both CoinJoin flags were dust-sized batches), what the change guess rests on there, and what each auto-expand rule did on the Colonial Pipeline ransom route, with what was and was not read of the documents. Development gets the commands for the new benchmark tiers. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Add the Accuracy page: how often the guesses are right, and where they break Results of the benchmark on synthetic data with known answers (change guess by confidence, clusters, the pattern badges), what the data cannot say, and what to do with the numbers. Linked from Home, the sidebar, Analysis and findings, and Development, which also gets the commands to run the benchmark. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>