KardashevJoin Kardashev Ascent ↗Ascent experiments / 003
An idea is cheap. An oven-hour isn’t.
Compare three scheduling rules in a fictional materials lab. More commitments can keep equipment busy; fewer can leave room to use fresh results.
Equal limits / different decisions
Committed candidates consume one material unit at selection, even if unfinished. “Reported” means measurement and analysis have both finished by the deadline. Quality is a fictional score on a 0–100 scale.
From commitment to feedback
▰ Synthesis▰ Measurement▰ Analysis delay
Published run / 10 November 2026
Default inputs: seed 17, 72 hours, at most 18 material units, one oven, one measurement instrument and a six-hour analysis delay. These are chosen settings.
| Rule | Reports | Best score | Mean lead, h | Oven busy |
|---|---|---|---|---|
| FIFO / window 12 | 16 | 77.77 | 35 | 95.8% |
| Adaptive / window 12 | 17 | 77.77 | 36 | 97.2% |
| Adaptive / window 2 | 11 | 77.77 | 12.55 | 63.9% |
In this seed, all three rules found the same best reported score. The smaller window shortened selection-to-report time, while leaving the oven idle more often and returning fewer reports.
| Rule | Reports | Best score |
|---|---|---|
| FIFO / window 12 | 15.925 | 77.654 |
| Adaptive / window 12 | 15.875 | 78.044 |
| Adaptive / window 2 | 10.5 | 78.404 |
Across these 40 fictional pools, the two-job window’s mean best score was only 0.36 points above the adaptive twelve-job window, with about one third fewer reports. This small descriptive difference does not establish a generally better policy. Change the delay, equipment or budget and compare again.
Download the published paired results and inputs (JSON) · Read the complete simulation code
Method / fictional lab queue v1
Each seed generates 36 candidates in six families. A family has a hidden quality centre drawn uniformly from 20–80; each variant adds uniform variation of ±12, clipped to 0–100. Synthesis takes 2–6 hours and measurement 1–3 hours, drawn independently. There is no measurement noise.
All three policies receive the same pool and shuffled arrival order. They know candidate IDs and families. A candidate’s quality becomes available to selection only when its analysis finishes.
FIFO: select in arrival order, at most 12 outstanding commitments.
Adaptive / 12 and / 2: rank families by observed mean quality (50 before any report) + 35 / √(1 + reported family count + outstanding family count). Pick the first remaining candidate in the highest-ranked family. Both use the same heuristic; only their commitment window differs.
At each integer hour: release feedback, finish measurements, release zero-delay feedback, finish synthesis, assign waiting measurements FIFO, fill the commitment window, then start queued synthesis FIFO. At the deadline, completions and feedback count; new work does not start. Ovens and meters run independently without pre-emption. Analysis adds a delay with unlimited parallel capacity.
There are no failures, changeover costs, wafer batches or correlated durations. One candidate uses one material unit. The adaptive bonus is a chosen heuristic, not a trained scientific model. Family resemblance is built into this toy world. The real question is how results change when that assumption, the bottleneck or the delay changes.
Mulberry32 makes the pool reproducible from the seed. Full JSON includes candidate values, event timestamps, slot assignments and per-hour queue traces. Forty-seed means are paired arithmetic summaries, not confidence intervals or lab forecasts.
Real inspiration / separate evidence
NIST’s August 2026 account of AMASE describes X-ray measurements at different points on a tin–bismuth combinatorial wafer. Many material compositions can share one physical sample. This simulator’s one-candidate-per-job accounting is a fictional choice.