Sequence and stopping rules
Cheap to try, cheap to be wrong
The value of this programme is not that it is likely to succeed. It is that the cost of finding out is small, the order of the work is fixed in advance, and every exit is written down before it is needed.
The order
What runs when
01
Species, settled first
E1 runs before any device work. It is the cheapest experiment and the only one with the power to end the programme, so it goes first — not last, where a sunk-cost argument would be waiting.
02
Device figures, with error bars
E2 and E3 run together: levels, write energy, speed, drift. Results are published internally in the same format whether they flatter the thesis or not, and the ledger sheet is updated the same week.
03
One honest comparison
A single benchmark against published protonic and memristive devices on the one job we claim — holding analog weights. No comparison against CPUs, GPUs, or anything the device is not trying to be.
04
A partner who builds hardware
The route to usefulness runs through someone with a system to put this beside. Approaching them before stage 03 would mean asking them to take the physics on trust, which is the wrong way round.
The exits
Stopping rules, stated before the data
Negative species result
The compute programme ends. No computing claim about OPALBLACK is made again, on any sheet, in any register.
No repeatable switching
The analog-weight thesis is dropped and the sheets are rewritten to say so, rather than being softened into ambiguity.
Retention far below the workload
The target narrows in public to workloads that tolerate refresh, or the programme stops. It does not quietly acquire a vaguer target.
A better route appears
If someone else's device does this job better on a comparable manufacturing route, that is a result too. The point was the one per cent, not the oxide.
What success would actually look like
A layer, made by anodising, holding enough distinguishable analog levels at room temperature, stable for long enough to serve inference weights, sitting beside a silicon processor and taking one narrow job off it. Nothing larger is claimed, and nothing larger is needed for the arithmetic on the statement sheet to hold.