Measurement logs from one system, one market. Not financial advice.
Twice this week (evening 2026-08-19 to morning 2026-08-20, and evening 2026-08-22 to morning 2026-08-23) the funnel on the track record page showed a baseline of a few hundred name-days instead of the true ~6,100, with percentages computed over that tiny slice. The stored record was never lost — a network failure during the morning run overwrote outcome statuses, and the corrupted view published. Recomputed and restored 2026-08-23; three independent guards now make the same failure impossible to publish.
A specification-curve sweep: 1,440 configurations of the accumulation pattern, scored identically, with three placebo worlds (same tickers, dates shifted) defining what pure noise manufactures on this window. The median configuration nets −0.16% before costs and −6.46% after. 5.7% of real configurations beat the noise envelope — chance alone predicts 5%. One placebo world produced a +53.65% configuration from scrambled dates. The edge, wherever it lives, is not robust to specification — and that verdict is the whole point of measuring this way.
The revised rule's receipt runs on a universe frozen on 2026-07-02 (183 names). The registration never said what happens when a name joins the watchlist afterwards. Pinned here: the headline tally counts only the frozen universe; later graduates accrue in a dated, visible sub-ledger and merge at the next registration. LAC — graduated 2026-07-07, one qualifying hit on 2026-08-04 (+13.4% at 5 sessions) — is the disclosed instance, and under this reading its win sits outside the headline number.
Measured on the corrected universe, Tier 1's first fires run 36.8% win with a 0.0% median (n=95) against a 31.1% / −3.2% baseline — a real but modest edge, with half the fires going nowhere or down. That is not flagship material. The registered quiet-accumulation pattern becomes the product's centre; Tier 1 remains on the track record, measured at every export, and if its numbers improve the page will say so automatically.
Between 2026-07-22 and 2026-08-14, roughly 21 ticker pages showed a quarterly cash balance overstated 1000× — junior explorers displaying $30–80 billion in cash. Root cause: a units-marker regex that let a bare large number masquerade as an 'amounts in millions' declaration. Corrected 2026-08-14 from the original filings, with a magnitude guard so the same class of error can no longer reach the lake or the site.
The pre-catalyst radar and the measured tracker quietly became two different rules — different quiet window, different desk test, different universe. The displayed one produced 2.7 qualifying names a month and was operationally dead; the measured one worked. Here is the drift, the fix, and the recalibrated rule that re-registers on 2026-08-04.
A measurement log from one system, one market, one seven-week window: most of what I believed about my own ASX signals was wrong, and the one pattern that survived I nearly deleted as a bug.