A differential re-analysis can only re-test claims the baseline wrote down, so a missed event stays missed forever

The failure

The DPZ baseline (2026-08-27) stated its event sweep explicitly:

"No new competitive, regulatory, or supply-chain event found. No aggregator commission shock, no named new entrant, no litigation escalation, no leadership change, no dividend cut."

Two material events were already public inside that April–August window:

Event	Date	Why it mattered
CEO succession — Weiner retires 9/30, Joe Jordan CEO 10/1, Brandon off the board	announced 2026-06-22 (8-K + PR)	Puts a 12-day-old CEO in front of the Q3 print the whole thesis is gated on
DPE closing 205 stores, A$259M write-downs; international net store growth guided down 175–275 stores	disclosed on the July earnings call	Directly undercuts "+209 net stores," the evidence the baseline used to argue the slowdown was demand-side rather than structural

The baseline rated its moat row CARRIED on unit growth while management had already guided the international half of that unit growth down.

Why the structure hides it

The command has two research inputs, and the gap falls between them:

The delta ledger is decomposed from the baseline report. An event the baseline never recorded produces no claim, so there is no row to mark UNTESTED, no falsification test to run, and nothing for Phase 5 to adjudicate. The ledger is a complete inventory of the prior report, not of the company.
Phase 3 scopes new research to "what has happened since the baseline date." That window starts after the missed event.

So the event is outside the ledger and outside the search. Neither half is defective on its own terms — and the pass still reports high thesis persistence, because persistence is computed over the rows that exist. A missing row cannot fail a test. Worse, each pass hands the same truncated ledger to the next, so the gap compounds with seniority: by the third report the omission looks like three consecutive confirmations.

This is the same shape as [[pitfall-kb-find-never-searched-the-reports]] — a search that returns results reads as a search that worked, and the absence of the thing you did not think to look for is invisible.

The fix

Re-sweep the baseline's own window, not only the window since it. Concretely, on every differential:

Pull the event list independently of the baseline — an 8-K/press-release listing from IR for the full period the baseline claimed to cover, plus the period since. This is one fetch and it is the only check that can catch a false negative.
Treat a baseline's stated absence as an assertion, not a finding. "No leadership change found" is an untested negative: it records what the last search returned, not what is true. Under the command's own evidenced-vs-asserted test it should be ledgered as asserted and tested first, exactly like any other unsourced sentence.
Read the last earnings call, not only the press release. Both DPZ misses were guidance and governance items — the DPE store-growth cut was said on the call and never appeared in a headline. Beats and misses make the wires; guidance revisions often do not.
When a baseline miss is found, discount that baseline's other unverified rows. Two misses in one file is evidence about the file, not just about the two rows. Per the command's rule that conviction should track the volatility of one's own analysis, this belongs in the conviction move — on DPZ it was roughly half of a 7.0 → 6.5 cut.

The transferable lesson

A differential inherits its predecessor's blind spots by construction, and reports them as stability. Any process that re-tests a prior record rather than re-deriving the subject will show its highest confidence exactly where the prior record was silent — because silence generates no row, no test, and no failure. The cost of the one independent full-window sweep that breaks the loop is a single fetch; the cost of skipping it accrues every pass.

History

2026-09-23 — observed on DPZ: two public events inside the baseline's stated search window surfaced only because the re-analysis ran an independent sweep. Claim #2 moved CARRIED → DRIFTED, thesis persistence 82% → 73%, conviction 7.0 → 6.5.