How the Validation Engine Works
Every number in a PhysicsCheck report — and across all our products — is verified against experimental databases. Our 12 physics constraint rules are deterministic: hardcoded physical constants and peer-reviewed reference data from 3 open databases, not AI predictions.
12
Physics Rules
5.4M
Materials
3
Open Databases
0
AI-Generated Data
The 8-Stage Validation Pipeline
From claim input to branded PDF — every stage is transparent, auditable, and deterministic.
01
Parse
Claim decomposed into composition, properties, conditions
02
RAG
Cross-referenced against 5.4M materials from 3 databases
03
Rules
12 deterministic physics constraints check every claim
04
Draft
AI drafts sections with explicit source attribution
05
Verify
Every number verified against database evidence
06
Enrich
Crystal structures, Ashby plots, comparison tables
07
Review
Named PhD specialist writes expert commentary
08
Render
8–12 page branded PDF with full audit trail
Triple Verification — Zero Hallucinated Data
The validation engine does not generate data. Every product validates claims against data that already exists in peer-reviewed databases.
Deterministic Physics
12 constraint rules use hardcoded physical constants and peer-reviewed reference data. Not AI predictions. Deterministic.
Cross-Verification
Every number in the AI-drafted report is extracted and checked against database evidence. Unverified values are flagged and regenerated, never published.
Human Expert Review
A named PhD specialist reviews every report. They interpret, contextualise, and write the expert commentary that makes each report actionable.
Three Databases — Named, Counted, Linked
Every database is open-access, peer-reviewed, and independently verifiable. CC BY 4.0 licensed.
AX
Alexandria
Ruhr University Bochum · Microsoft Research
5.1M
near-hull structures
CC BY 4.0
View source →
MP
Materials Project
Lawrence Berkeley National Lab (DOE)
211K
inorganic materials
CC BY 4.0
View source →
Material Coverage by Class
| Class | Classification | Materials | Ashby Bounds |
|---|---|---|---|
| Loading coverage data… | |||
Alexandria v2024.3 — synced 2026-02-18
OQMD v1.6 — synced 2026-02-18
Materials Project v2024.01 — synced 2026-02-18
Daily sync via pg_cron
Coverage counts and Ashby bounds auto-regenerate on each database sync. Classification uses the SQL rules defined in the constraint engine. ● fresh (<7 days) · ● aging (7–30 days) · ● stale (>30 days)
12 Constraint Rules — Categories
Categories and representative examples shown. Exact thresholds are kept private to prevent claim-tuning.
✓
Thermodynamic Checks (3 rules)
PASSValidates energy, stability, and thermal limits against experimental references and DFT computations.
Formation energy · Phase stability · Thermal limits
✓
Structural Checks (3 rules)
KNOWNVerifies crystal structure, composition validity, and material existence against known databases.
Crystal structure · Composition · Novelty detection
!
Transport Properties (2 rules)
WARNCross-checks claimed physical properties against Ashby limits and fundamental physics laws.
Ashby bounds · Wiedemann-Franz verification
✓
Data Validation (4 rules)
PASSEnsures internal consistency across databases, literature benchmarks, and synthesis feasibility.
Cross-DB agreement · Literature consensus · Synthesis routes · Dimensional analysis
Technical Documentation
For scientists evaluating methodology rigour and potential SME candidates.
All 12 Constraint Rules
R1 · Ashby Upper Bound — OOM percentile check against Ashby envelope
R2 · Formation Energy — Cross-database ΔHf comparison
R3 · Composition Validity — Charge balance and element validation
R4 · Crystal Structure — Polytype and lattice parameter verification
R5 · Phase Stability — Convex hull distance analysis
R6 · Thermal Limit — Operating temp vs decomposition/melting
R7 · Dimensional Analysis — Unit consistency via pint library
R8 · Wiedemann-Franz — Metal-specific transport law verification
R9 · Synthesis Feasibility — Known routes and precursor availability
R10 · Literature Consensus — Multi-source citation agreement
R11 · Cross-Database Agreement — Multi-DB value convergence
R12 · Novelty Flag — Known / Variant / Novel classification
Error Margin Disclosure
PhysicsCheck reports include uncertainty where available. Formation energy comparisons use ±0.10 eV/atom tolerance. Ashby bounds use percentile ranges (p95, p99). Literature consensus reports both mean and variance across sources. Where DFT accuracy limits apply (typically ±0.1–0.3 eV for GGA functionals), this is stated in the methodology appendix.
Database Schema Philosophy
Local-first PostgreSQL with pgvector. All three databases are synced locally for sub-300ms query latency. No external API calls during validation. Full GDPR compliance by architecture — no user data leaves the EU-hosted infrastructure.
LD
Luc Durand, PhD
PhD in materials science specialising in coatings, thin films, and surface engineering. Research experience at CEA Saclay. Built PhysicsCheck because VCs deserve physics, not guesswork.