The six soak evidence documents are withdrawn
None of the six is presented as current evidence any longer. Every original file is preserved, unchanged and in full, and each remains reachable from this page as an archived historical version.
Withdrawn 30 August 2026 · coordinated corrected re-issue pending
Why all six, and not just the summary
The defect is shared. In each run, real measurements were passed through verdict logic that could not reach a failing outcome — a gate that reads a field nothing writes, a timer that expires before the run begins, a check that does not exist. Leaving five documents standing as current evidence while withdrawing the sixth would publish a set that disagrees with itself about whether the same runs passed. So the whole set is withdrawn together, and the corrected set will be re-issued together.
No corrected figure is published here — deliberately
This notice says what is wrong with each document. It does not print a replacement cycle count, completeness ratio or corrected verdict for any run, and neither does any other page. Publishing corrected numbers one document at a time is how a set comes to contradict itself; the corrected figures are released only as a coordinated set, when the runs have been re-gated against controls that can report a failure.
What is not wrong — and this is part of the record
- No fabrication was found at the source. The individual figures came verbatim from the machine-generated closing records that were the document generator's input. They were read, not invented. The one exception is named below, under the portfolio summary.
- An earlier generation of these documents refused to guess. It carried
[PENDING]markers where the numbers now sit, together with a stated release rule barring distribution until the closing metrics were inserted from the machine snapshot. The team wrote the markers in and held the documents back. - The measurements are largely sound. The judgements placed on top of them are not. That distinction is the whole finding, and it is why these documents are preserved rather than deleted.
Document by document
The six are not identical and the findings differ. Each entry below states its finding and how that finding was established, including where it could not be re-verified on the machine that produced this notice.
24-Hour Hybrid Soak
DCS_Soak_24h_hybrid_Evidence.pdfThe run's cycles are complete — it executed every cycle its configured duration and cycle interval called for, and no cycle is recorded as failed. But the record issues fewer receipts than it ran cycles, while reporting zero failed cycles. One of the missing receipts is plausibly the deliberately refused partial signature, which the run counts as correct behaviour; the record does not account for the remainder. None of the ten verdict checks reads receipt coverage at all, so the gap could not have affected the verdict no matter how large it grew.
Confirmed directly from DCS_Soak_24h_hybrid_Closing_Record.json, which is published alongside this notice and remains linked from the evidence page. The ten check names are listed in that file; none of them reads the receipt count.
7-Day Continuous Soak
DCS_Soak_7day_Evidence.pdfThe run's receipt verifier was a function that returned true unconditionally, and the closing record shows the run was configured to use it (mode.verify: "local-true"). The cycles were executed; nothing about them was verified. Every pass/fail figure this document reports — failed cycles, receipt coverage, fault recovery — came from a control that could not return false, so none of them is evidence of a pass.
Finding as established by the reconciliation of 30 August 2026, which recomputed it from the closing record. That closing record is not present on the machine that produced this notice, so the finding is reported here on that authority and has not been independently re-derived. The absence of the underlying record is itself part of why the document is withdrawn rather than defended.
Archived historical version (PDF, withdrawn) →30-Day Internal Soak
DCS_Soak_30day_Evidence.pdfThe configured run duration in milliseconds exceeds 231−1, the largest delay the timer accepts. On overflow the timer fires within a few milliseconds and the run marks itself passed. A thirty-day PASS was obtainable in roughly the time it takes to start the process. The shorter runs are below the ceiling and are unaffected by this particular defect.
The overflow is arithmetically certain for the duration this run was configured with: 2,592,000,000 exceeds 231−1 = 2,147,483,647, and that comparison is verified here. The configured value itself is taken from the reconciliation of 30 August 2026; the closing record is not present on the machine that produced this notice and the value has not been independently re-read.
Archived historical version (PDF, withdrawn) →72-Hour Triple-Hybrid Soak
DCS_Soak_72h_triple_hybrid_Evidence.pdfThe run completed materially fewer cycles than its configured duration and cycle interval called for, at an observed cadence well slower than the one configured. It was nonetheless gated PASS, because no verdict function reads cycle completeness at all — the shortfall is invisible to every check that decides the outcome. A separate uptime threshold is declared in the gate configuration and read by nothing.
Finding as established by the reconciliation of 30 August 2026, which computed the expected cycle count from the record's own duration and interval fields and compared it against the completed count. That closing record is not present on the machine that produced this notice; the shortfall is reported on that authority and has not been independently re-derived. The corrected counts are withheld here under the no-corrected-figures rule above.
Archived historical version (PDF, withdrawn) →15-Day Continuous Soak
DCS_Soak_15day_Evidence.pdfThe run fell short of its scheduled cycle count, and it reports a no-error check as passing while the panel's own last-error field is not empty. A check that reports "no error" over a record that contains one is not reading the field it claims to read.
Finding as established by the reconciliation of 30 August 2026. The closing record is not present on the machine that produced this notice, so neither the shortfall nor the non-null error field has been independently re-read here.
Archived historical version (PDF, withdrawn) →Soak Evidence Portfolio Summary — the aggregate
DCS_Soak_Evidence_Portfolio_Summary.pdfThis document repeats the withdrawn 7-day and 30-day results, so it inherits both defects above. It prints a 72-hour cycle count that matches neither the completed count in that run's machine record nor the count the run was scheduled to perform — a third position on a figure that already had two. And it carries three aggregate figures that appear in no closing record and in no other document: 1,875,024 RECORDED CYCLES, 78/78 FAULTS SURVIVED and 5/5 FINAL PASS. These are sums over the withdrawn counts. They inherit every defect above and add the appearance of a portfolio-wide result that was never separately measured. This is the one place where the fairness finding above does not hold: these three figures were not read from a record, because no record contains them.
The three aggregate figures and the 72-hour discrepancy are as established by the reconciliation of 30 August 2026. The underlying closing records are not present on the machine that produced this notice.
Archived historical version (PDF, withdrawn) →All six PDFs are unmodified. None has been deleted, moved, renamed or edited, and none is going to be: they are the record of what was published, and the correction history is part of the evidence. They are reachable only through this notice so that no reader meets the figures without the withdrawal attached to them.