Measured product lead custody receipt.
github.com/Zer0pa/Materials-WorkbenchRESEARCH-READY
LIVE
H100 ARTIFACT OPEN
Seven-Layer Physics Materials Workbench
Materials lane · Materials-Workbench · PyPI zer0pa-materials-workbench v0.1.0 stale · github.com/Zer0pa/Materials-Workbench
Materials Workbench is a CPU-side bench for in-silico battery and thermoelectric research.
Candidate chains arrive with schemas, validators, recompute checks, and dispatch wired together, so weak chemistries are rejected at the desk before they consume cluster hours. 3,547 strict-full tests pass and 588 Runpod-sim parity tests hold; the H100 campaign is the next wave, not the current result. Real GPU evidence is not claimed yet.
01 · THE GAPSPLIT WORKFLOWS
“Most candidate chemistries reach the cluster. vetting comes too late.”
02 · MARKETSWORKFLOW FIT
Workflow-fit hypotheses for a CPU pre-screen bench; not adoption, not TAM, not a discovery claim.
03 · VALUE
CPU-sidebench
Best fit: materials informatics, battery, thermoelectric, and industrial AI teams that need pre-HPC candidate checks.
04 · INSIGHT
Reject weak materials chains before any GPU runs them.
05.0 · CURRENT TECHSPECIALIZED SOLVERS, SPLIT HANDOFF
Today's stack threads DFT, phonons, MLIP, CALPHAD, phase-field, provenance, and lab automation through separate specialised tools. The gap is not a missing solver. It is a shared candidate record that lives before H100 spend.
05.1 · OUR TECHRAW-EVIDENCE RECOMPUTE
Materials Workbench at Zer0pa/Materials-Workbench is an installable alpha that holds schemas, validators, seven recompute checks, packet records, Runpod dispatch, and an H100 cutover path in one CPU-side bench. Battery and thermoelectric candidates run the same pre-screen on a laptop today that they will run against real GPU artifacts in the H100 wave.
05.2 · BENCHMARKSPOST-WAVE-F · CPU + RUNPOD-SIM
Tests3,547strict-full PASS
Parity588Runpod parity tests
Recompute7raw-evidence checks
Stress cases16plus 7 checks
Current status: 16/16 stress cases caught · 7 recompute checks hardened · no real GPU-backed runpod_rest artifact yet.
06 · MEASUREMENTRECOMPUTE + HASH HISTORY
7 recompute checks, 16 stress cases, hash history over every chain.
06.1 · COMPARATIVE PERFORMANCE · CPU WORKBENCH VS GPU CAMPAIGN
CPU strict-full and Runpod-sim parity hold today. H100 is the next campaign — no real GPU result has been produced yet. DFT/MLIP comparator parity sits outside the CPU bench.
07 · KEY METRICSMATERIALS-WORKBENCH POST-WAVE-F
07.1 · CPU STRICT CHECK
3,547/3,547
Strict-full pass · 2 pycalphad skips, 0 misses
07.2 · RUNPOD PARITY
588PASS
Runpod parity tests · mock results blocked from posing as real
07.3 · RECOMPUTE CHECKS
7hardened
Raw-evidence recompute · inputs, sources, novelty, ionic, NEB stages
07.4 · STRESS CASES
16/16
Replay-error breaches plus recompute checks · 0 misses
07.5 · GPU EVIDENCE
null
Metric pending · no real GPU result produced yet
08 · DETERMINISMRECOMPUTE BEFORE GPU
Every chain replays on CPU before any GPU touches it.
08.1 · WHAT DETERMINISTIC MEANSRAW-EVIDENCE RECOMPUTE
The seven recompute checks take a candidate chain's outputs and re-derive every layer's hash from inputs. Any divergence rejects the chain; only chains that survive the recompute move forward.
The discipline runs today on CPU and on the Runpod-sim parity surface (3,547 strict-full + 588 parity). The H100 wave will run the same checks against real GPU artifacts. The unit of bit-exactness is per-chain, per-layer-hash.
08.2 · THE FIDELITY GAP
Honest Blocker ·
No real GPU-backed runpod_rest artifact has been produced. Discovery of any material is not claimed. Public PyPI v0.1.0 is live but stale pending v0.1.1. License is SAL-7.1 / GitHub NOASSERTION; UMA/HF org, Materials Project credential, real endpoints, and EMMO cleanup remain open.
09
FIVE PATHS FROM ONE pre-GPU bench.
09.1 · THIS REPO'S AMBITION
The hinge is not one chemistry result. It is a bench where candidate generation, validators, constraints, and run history travel together from a researcher's laptop to a GPU cluster. If that path survives H100 completion, materials exploration becomes something other teams can repeat, compare, and contest.
09.2 · WHAT WORKS NOW
Working now: candidate-bench architecture, validation route, CPU and Runpod-sim parity; H100 completion target.
09.3 · WHAT'S STILL OPEN
Still open: H100 execution, real GPU artifact, candidate data, release evidence, and source context.
09.4 · LAB BUDGETS · NEAR-TERM (12–24 MO)
Weak chemistries die at the desk
A battery research lead can spend a Monday morning killing forty candidate chains on a laptop instead of carrying them into a Friday cluster reservation. The H100 queue starts holding fewer candidates that nobody really believed in.
09.5 · INFORMATICS · NEAR-TERM (12–24 MO)
Every search path carries its own paper trail
A materials-informatics team that ran a thermoelectric sweep last quarter can pull any candidate back up months later and see exactly which validator fired and which input drove it. The next investigator does not start from the team's slack scrollback.
09.6 · GPU ECONOMICS · MID-TERM (24–48 MO)
H100 hours stop funding speculation
When the same recompute discipline survives onto real GPU runs, a lab director can justify H100 spend against a candidate's pre-screen record. The conversation with procurement moves from “we need more cluster time” to “these forty chains earned it.”
09.7 · COLLABORATION · MID-TERM (24–48 MO)
One candidate, not eight notebooks
A battery candidate handed between two groups — a DFT specialist and a phase-field modeller — arrives as one object carrying schema, constraints, validator output, and run history. The receiving team can rerun and contest it without rebuilding the reasoning from emails.
09.8 · DISCOVERY DISCIPLINE · PARADIGM (48 MO+)
Materials R&D inherits release discipline
If candidate generation, validation, and execution stay coupled all the way through GPU scale, the center of gravity in materials discovery moves from a single hero result toward governed evidence flows that any independent group can pick up and continue.