What was built
The design in the last entry is now code. The specimen corpus is 51 documents from eleven P&I clubs, and there is a reader for each club and each kind of document that club issues. Every one of the 51 now reads.
A reader is a small piece of code that understands one club's layout and nothing else. That is a deliberate choice over one general system that tries to understand every club at once: the number of clubs is small and changes slowly, while the layouts differ from one another in ways no single set of rules expressed well. The first attempt used one general rule language and reproduced fewer than a third of the figures. The readers reproduce all of them.
Every figure is checked against an answer worked out separately
For each of the 51 documents there is an expected reading, kept beside the code and checked automatically: 1,072 values in total, covering vessels, periods, tonnages, premiums, commissions and the rest.
The expected readings were written by people who had not written the readers, and were then verified again by a third party. That separation earned its keep. The verification found sixteen real defects in work that had already passed every automated check, including four documents filed under the wrong kind of document entirely. A checker that only tests what its author believed would find none of those.
Proving it reads rather than remembers
A reader tested only against the documents it was written for may simply have memorised them, and would then fail on the first real document that arrives. So we tested it on documents it had never seen.
We rebuilt the corpus: 48 of the documents regenerated with different vessel names, different assureds, different references, different dates and different figures, while every layout, label and position stayed exactly as it was. The readers had seen none of it.
-
48 documents rebuilt with changed details
-
776 values read from them
-
463 of those values different from the originals
Zero wrong
The 463 is the number that matters. It says the readers were genuinely finding labels and reading what sat beside them, rather than recalling what the original said. The Japanese-language documents were left out of the exercise, because rebuilding them would have lost their typesetting and tested nothing.
What a figure carries with it
Every value a reader produces carries where it came from: the page, the exact place on that page, which reader produced it, and whether the figure was read from the document or calculated from other figures. A value that cannot show its place on the page is refused before anyone sees it, so the evidence cannot quietly go missing.
That distinction between read and calculated is deliberate. A figure the document states and a figure we worked out from other figures are different things, and a book that presents them identically is inviting someone to rely on the wrong one.
The club comparison tool is back
The P&I club comparison is published again, at ajgmanifest.com/comparison, and it opens straight away with no sign-in. It reads published market data and nothing else: it never touches the renewal book, which is why it needs no sign-in to protect. The figures in it remain illustrative until real market data replaces them, and the tool says so on the page.
Also in this period
Admission and authentication is settled. Signing in and proving who you are to the services that compute keys are two separate acts, and neither stands in for the other, so an identity provider that is fully compromised yields an application shell and no readable record.
Authorisation was designed and then adversarially reviewed, and the review broke it. Three reviewers reading the design and the specification found that two of its central mechanisms did not exist as written. The design has been revised and the reviewers' findings are recorded in full alongside it. We would rather report that than report a clean pass we had not earned.