Find the minimum infrastructure state — and what could still flip the decision.
The workload model predicts. The simulator explores. The Dynamic Boundary finds the modeled PASS/FAIL edge. The Value Layer quantifies economics, scale cliffs and evidence gaps. PEX replaces uncertainty with physical measurements before qualification.
Compare complete machine/runtime states, not accelerator names.
Edit a vendor or buyer candidate live. Unknown economics remain excluded from economic winners; complete but unmeasured vendor states remain PEX REQUIRED.
Buyer-specific modeled boundary
The current service contract generates the minimum simulated PASS and adjacent FAIL for the selected state.
What does the modeled choice save — if the economics are known?
Annual compute, energy and token economics use the whole machine state. Unknown values remain UNKNOWN.
Where does this state stop being enough?
Stress one service dimension until the selected machine needs more devices or leaves the modeled PASS region.
How much of this decision has actually been proven?
Measured, vendor-declared, modeled and unknown coverage stay visible separately.
Which missing measurement is most likely to change procurement?
PEX spends node-hours where evidence can actually change the recommended state, not where it merely adds more benchmark data.
Stress the service before buying hardware.
Context and concurrency sweeps show where the current candidate remains in the modeled PASS region.
Spend GPU / node hours where they can change the decision.
Experiments are ranked by decision-change probability, uncertainty reduction, economic impact, architecture learning and value per node-hour.
Physical evidence supersedes weaker evidence without deleting history.
Import a measured metric for the exact state. The evidence ledger preserves the model prediction and makes the stronger evidence effective for the decision.
IBM Power + Spyre
IBM or another hardware partner can enter the actual device count, memory, bandwidth, runtime, power and commercial inputs live in the Candidate Editor. The state immediately joins the same modeled search; physical PEX remains required before qualification. No IBM price/performance number is fabricated.
Stored scheduler replay remains context, not the buyer's answer.
Torrent Tokens is surfaced as REPLAY / PROTOTYPE. Historical S16/S17 evidence cannot determine the current buyer-specific Qualified State.
Qualification requires measured evidence.
A modeled state can guide procurement research and choose the next experiment. Measured PASS plus adjacent measured FAIL establishes the physical boundary and can promote the state into the Qualified State Library / Control Envelope.
Package the buyer or partner decision.
The report includes the recommended modeled state, next-best alternative, Procurement Delta, Scale Cliff, Evidence Confidence, Decision-Flip Unknown, highest-value PEX run, Required Physical Evidence and Deploy / Hold / Redesign status.
Use the model to choose what to test. Pay us to establish the physical boundary.
The paid case freezes the accepted service, runs the highest-value experiments, replaces simulated uncertainty with physical evidence and returns the deploy / hold / redesign decision.