Log Coverage Is 1.0000 at Every Schema Width, Including the Width Where the Record Is Empty

Arc Ops Level 6 is audit trails: reconstruct why it did that from the logs alone. 24 runs, 156 steps, a 15-field schema swept at every width.

Level page: https://dev48.infy.uk/arcops/level6-audit-trails.html · Repo: https://github.com/dev48v/arc-ops

Every number the log reports about itself is perfect

Record coverage is 1.0000 at every width — including width 0, where the record is empty. The hash chain verifies. Records written, bytes written and fields declared all improve monotonically as the schema grows.

And the log reconstructs only 84 of 156 decisions.

	reconstruction
overall	53.85%
on the 38 irreversible steps	36.84%
on the robust steps	81.40%
on the fragile steps	43.36%

It is 1.46× worse exactly where a review opens it.

Why, and it is not carelessness

The steps a review opens are the irreversible ones, and those are disproportionately the fragile ones — 78.95% of irreversible steps are fragile against 72.44% overall. A fragile decision is precisely the one whose explanation needs the most fields, and the fields that explain it are the large, sensitive ones that get truncated or redacted first.

So the pressure that shrinks a log — size, retention, privacy — removes exactly the records that would have answered the question, while leaving coverage at 1.0000.

What the fields are worth

• Nothing reconstructs below width 12. The last three fields cost 5,264 bytes a run and buy 0 reconstructions.
• Six fields are worth exactly zero reconstructions: actor, outcome, run_id, step, target, ts — the six every logging standard insists on.
• The three sensitive fields carry 90 between them, which is why the privacy review and the incident review want opposite things.

Ten questions a review actually asks

Three are answered on no record at the reference settings. One of those cannot be answered by any schema at any width: what did it NOT see — because a log records what happened.

Redaction

mode	reconstructions	bytes
none	101	72,852
hash	84	45,884
drop	84	40,620

Hashing and dropping cost exactly the same 17 reconstructions, and hashing still writes 5,264 more bytes. You pay for the hash and get none of the answer back.

One pen, proved by parsing all ten modules with ast: 0 imports that could reach a clock, a logger, a socket or a file; exactly 1 call site appending to a Journal.

1,061 tests, standard library only, MIT. 53 verifier asserts, 17 page checks, 0 failures.