IASYMMETRY
Verification is trivially cheap.
Running a calibrated model against held-out measurements takes seconds. Finding those parameters takes hours of search across thousands of evaluations.
cheap verify · expensive solve
DIGITAL TWIN CALIBRATION · BITTENSOR SUBNET 456
The calibration layer
for digital twins.
A decentralized marketplace where specialized optimizers compete to make simulation models faithful to reality. Verified on held-out data. Scored with ASHRAE metrics. Truth wins.
ROUND_r-2026-04-18-1142/bestest_hydronic_heat_pump
emulator (ground truth)miner predictions (N=3 shown)
CVRMSE
0.119
NMBE
-0.019
R²
0.901
SIMS_USED
287 / 1000
BITTENSOR SUBNET
456
testnet live
BEST CVRMSE
0.089
bestest_hydronic · placeholder
ACTIVE MINERS
18
placeholder
TEST CASES IN ROTATION
3
BOPTEST · verified
ROUNDS COMPLETED
2,147
placeholder
TESTS PASSING
131
from repo
01 · ARCHITECTURE
A two-model system. One produces truth, the other learns to match it.
Validators run a complex emulator (BOPTEST) to generate ground truth. Miners calibrate a simplified model to match it. Scoring is done on data miners never see.
VALIDATORground truth
TRAINING MEASUREMENTSsent to miners
HELD-OUT MEASUREMENTSkept secret · scoring only
MINER × Ncalibration target
LOCAL OPTIMIZERBayesian · CMA-ES · surrogate
SIMULATION_BUDGET1,000 evaluations / round
01
Validator generates ground truth
A containerized BOPTEST emulator runs the full period. Training measurements go to miners; held-out measurements are kept secret for scoring.
02
Miners calibrate locally
Each miner runs their own optimizer - Bayesian, CMA-ES, surrogate-assisted - against a simplified RC network model. Parameter search, not prediction.
03
Validator verifies
One simulation per miner against the held-out period. CVRMSE, NMBE and R² are deterministic - any validator reproduces the same score.
04
Weights set on-chain
5% relative floor, power-law amplification (p=2), normalized to sum=1. Top performers receive disproportionately more weight.
02 · MECHANISM
Three properties make this uniquely suited to Bittensor.
IIDIVERSITY
A portfolio of strategies wins.
Bayesian, CMA-ES, MCMC, gradient-based, surrogate-assisted - no single method dominates every building type. Fifty competing approaches consistently beat any one of them.
ensemble > monoculture
IIIAI-PROOF
LLMs cannot do this.
Calibration requires running actual physics simulations thousands of times. It is pure computational search, not pattern recognition. Frontier labs have zero interest in building physics.
compression-resistant
03 · SCORING
The formula is published. The weights are auditable.
No proxy metrics. No hidden parameters. Every miner receives a machine-readable breakdown every round. You can run the exact scoring module locally before submitting.
COMPOSITE SCOREspec v2
composite = 0.50 × cvrmse_norm
+ 0.25 × nmbe_norm
+ 0.15 × r2_norm
+ 0.10 × convergence_norm
# weight pipeline (applied per round)
floor = max(composite) × 0.05
floored = composite if composite ≥ floor else 0.0
powered = floored ** 2 # power-law p=2
weights = powered / sum(powered)float64 · deterministic · open source
50%
CVRMSElower
Coefficient of Variation of RMSE. ASHRAE Guideline 14 standard.
threshold < 0.30
25%
NMBE→ 0
Normalized Mean Bias Error. Systematic over- or under-prediction.
threshold |NMBE| < 0.10
15%
R²higher
Coefficient of determination. Overall fit quality.
threshold > 0.85
10%
convergencefewer sims
Simulations used relative to budget. Self-reported, bounded.
threshold —
04 · AUDIENCES
Three groups benefit immediately.
CUSTOMERS
Engineering firms & building owners
You need calibrated models for ASHRAE Guideline 14 M&V, Model Predictive Control, or portfolio-scale analysis. Submit your BOPTEST-compatible model and measurement data - receive verified parameters with full accuracy breakdown.
- Per-calibration pricing (API, waitlist open)
- CVRMSE, NMBE, R² on held-out data
- Same methodology engineering consultancies charge $5K–$20K for
INVESTORS
TAO holders & subnet evaluators
Zhen targets a green-zone flywheel: emissions today, paying customers tomorrow. External revenue routes through value sinks - paid API access, licensing, alpha burn. Miner transparency is first-class.
- Subnet 456 · spec v2 · testnet live
- Phase 1 benchmark dominance → Phase 2 API → Phase 3 enterprise
- Published scoring, score floor, power-law weight amp
RESEARCHERS
Building energy & digital twin scientists
Zhen is a public benchmark. Every round publishes composite scores, per-component scores, and simulation-used counts. Methodology is open source. Test cases are BOPTEST and Energym; adding new ones follows a proposal process.
- BESTEST reference buildings, Brussels climate
- RC network simplified models, FMU ground truth
- Deterministic train/test split via SHA-256
05 · LEADERBOARD
Public rankings, per round.
Composite scores are public. Calibrated parameters are private to the submitting miner, so top performers can't be copy-pasted.
r-2026-04-18-1142·bestest_hydronic_heat_pump
placeholder data
#
HOTKEY
CVRMSE
NMBE
R²
SIMS
01
5Cxyz…a8f2
0.089
-0.019
0.962
287
02
5Fdra…Q4Nk
0.104
+0.031
0.951
412
03
5GmpV…7yLx
0.117
-0.041
0.938
356
04
5HnwR…kP1z
0.142
+0.052
0.914
801
05
5DpKe…m3Xq
0.168
-0.072
0.887
644
06
5EyTv…b2Hc
0.194
+0.089
0.863
923
07
5JrMb…u8Wq
0.221
-0.094
0.841
767
08
5KzLp…x4Vn
0.248
+0.101
0.819
888
score floor applied: 0.05 × max·power-law p=2
Full leaderboard →06 · FAQ
Questions people actually ask.
Calibration. Every digital twin needs parameters tuned to match reality, wall R-values, infiltration rates, HVAC efficiencies. Engineering consultancies charge $5K–$20K for this and take weeks. Zhen runs 50 optimizers in parallel per round and returns verified parameters in hours.
07 · GET INBring us a model.
Bring us a model.
We'll calibrate it.
Priority access for engineering firms, ESCOs, and research teams working on building energy. BOPTEST-compatible models preferred. Joining the waitlist does not commit you to anything.