What we predicted, and what happened

Visitors is external traffic only — our own verification hits are flagged and excluded. 4 of these 4 are VOID: registered, never run, and contributing nothing to any calibration in either direction. Deadline voided by founder direction 2026-08-06. The test was never run: 18 external visitors across all four probes, no tagged link ever shared, and no outreach performed against the registered threshold. Recording FAIL would enter "demand was tested and absent" into the calibration record when nothing was tested. Void contributes nothing in either direction. 3 of these 4 have had fewer than 25 visitors, so a threshold missed on the deadline would record that distribution was never attempted, not that demand was absent. An unrun test is not a failed one and will not be filed as one. 3 carry no freeze record: the metric and threshold are set, but nothing captured what was committed to or when. (GA4 Fix, AD Sync Fix, Fail2ban for journald.)

Each of these was recorded at the moment the verdict was issued. None has been run. They are published now so that what was predicted cannot be revised after the fact.

20 of 36 decided verdicts carry no usable pre-registered test. The cause is ours and it is documented: the arbitration agent nested the test object inside its dimensions block on 8 of 10 runs, and the pipeline read it from where it was expected to be rather than normalising it at the boundary, writing NULL. The payload is now normalised at a single intake point and a regression case holds it there. The affected verdicts need re-running, not patching — editing them would destroy the calibration record, which is the only thing here worth having.

mention_count — whether a problem was raised once or fifty times — feeds three confidence functions at up to 20 points. It was a stored number that defaulted to 1 at creation and was updated only by a clustering step that never ran, so 19 of 36 verdicts were scored as though one person had raised the problem when no signal was linked at all. The database now derives it from what is actually attached. These verdicts are named rather than rescored: editing a frozen verdict destroys the calibration record it exists to be. 96 of 170 demand signals have never been processed into a candidate — not failed, never reached.

No prediction has resolved yet, so the rubric has never been graded. The first four resolve 21 and 30 August 2026. Until then every score this system publishes is uncalibrated against its own outcomes, and saying so is the difference between this and a confidence number with nothing behind it. Every verdict is written into a prediction ledger at the moment it is issued — the score, the confidence, the dimension scores as they stood, and the rubric version that produced them — and the outcome is written back automatically when a test resolves. Nothing in this category does this: fourteen tools were surveyed and none publishes a calibrated error bar, a pre-registered test, or its own failure history, let alone uses them to re-weight what it scores.