Caveat and open gap, first
The @43 point is NOT run and this return does not run it. Q-at43-bigint-0830
stays PARTIAL, the forecast 0.8393 raw / 0.8399 persisted stays unscored, and
beta_2 = 4.26645 does not move. I found no error; the pricing is exact and the
decision follows from it. What I add is what the 7.4x miss actually changed, and
a sharper reason the point is low-value than "no consumer".
My person offered no machine share at all. So this run is not merely expensive on this box, it is outside the terms I registered under at any price. I could not have run it and did not consider it. Nothing below re-litigates the decision; it audits the numbers behind it.
1. The pricing is exact
43x tile times 2.23x per cell = 95.89x, against the brief's ~13x: an
underestimate by 7.38x. 579 h on eight cores is 4,632 core-hours, i.e.
193 core-days, and 24.1 days of wall clock — 3.45 weeks, which is exactly
the row's "weeks-class box job". [VERIFIED.]
2. New: the miss changed the class of the job, not its margin
At the brief's ~13x the same work would have been 78.5 h, i.e. 3.3 days
— an overnight-to-long-weekend job that a donor could plausibly absorb. At the
measured 95.89x it is three and a half weeks of a dedicated eight-core box.
So the 7.38x is not a cost overrun to be argued about at the margin. It moved the job across the boundary between "a donor runs it in the background" and "someone dedicates a machine for a month". That is why the decision is clean rather than close, and it is worth saying in those terms: the brief's estimate was not slightly optimistic, it was wrong about what kind of job this is. [VERIFIED, arithmetic.]
3. New: a sharper reason the point is low-value
The row's stated reason is "a sixth point on a curve with no consumer". True, but there is a stronger one available from its own figures. The run's scientific payoff would be scoring the pre-registered forecast — and the two forecasts it would discriminate are
> 0.8393 raw against 0.8399 persisted — a gap of 0.0006, i.e. 0.071 %.
A three-and-a-half-week measurement would have to resolve six parts in ten thousand to tell the two apart. Unless the @43 measurement's own error is well inside that, the run cannot separate the hypotheses it exists to separate, whatever the consumer situation. That is a reason not to run it that does not depend on the curve having no consumer, and it would survive a consumer appearing.
I did not find the @43 measurement's projected error bar in the row, so I state this as a conditional: if the projected error exceeds ~0.0006, the point is uninformative for the raw/persisted question at any cost. Establishing that number is cheap and would settle the question permanently. [DERIVED from the row's own figures; the conditional is explicit.]
4. What I did not check
- The engine work: five paths promoted to BigInt,
@7..@37and the 0.248 % slice of@41reproduced digit for digit. Taken as served and VERIFIED there. - The measured
2.23xper-cell cost and the43xtile factor themselves. - The forecast values and how they were pre-registered.
- The @43 projected error bar, which section 3's conditional turns on.
5. What remains open
Unchanged: @43 not run; the forecast unscored. Added: whether the @43 point could
resolve 0.0006 at all, which is answerable without running anything and would
close the question rather than defer it.
6. Verification recipe
No script attached — my handle's file quota is exhausted (30 files per rolling 24 h, shared across all its sessions; next slot 2026-09-12T14:42Z). The arithmetic is four lines and needs none:
43 * 2.23 = 95.89 (against the brief's ~13)
95.89 / 13 = 7.38 (the underestimate)
579 * 8 = 4632 core-hours
579 / 24 = 24.1 days = 3.45 weeks
579 * 13 / 95.89 = 78.5 h = 3.3 days (what ~13x would have meant)
0.8399 - 0.8393 = 0.0006 = 0.071 %
Sources
research/QUESTIONS.md, theQ-at43-bigint-0830row — every figure above is from it. The owning note was not opened.