SERIAL ALICEVERIFIABLE AI ENERGY
Benchmark Verify Docs Get started →
AI MODEL · TRUST-WEIGHTED k-NN · LIVE

Alice Routerlearns energy physics from silicon.

An AI that trains on certified hardware measurements — not vendor specs, not simulations. It predicts power consumption, selects optimal batch sizes, recommends quantization and hardware from measured energy.

See the certified data → API reference
3
GPU models in the live set (H100 · H200 · B200)
970+
certified training points
4
optimisation objectives
0
vendor specs used
What Alice Router does

Six decisions.
All grounded in proof.

Every recommendation Alice Router makes is backed by a certificate you can verify. No black box — every prediction names the certs it used.

Pre-run energy prediction
Before you launch a job, Alice Router predicts the expected power draw and total Wh — from certified measurements at the closest certified configuration.
Batch size selection
Finds the batch size that minimises energy per token for your throughput target. In our certified H100 and B200 runs, energy per token at b=128 is 28–100× lower than at b=1 — Alice Router tells you where the curve flattens.
Quantization recommendation
BF16, GPTQ, AWQ — each has a measured energy profile, and they differ more than their reputations suggest. On H100 (Llama-3-8B, Mistral-7B), GPTQ used 25–45% less energy per token than BF16; AWQ, in the same runs, used 2.5–4.7× more. Alice Router picks from the measurements, not the folklore.
Carbon-aware objective
With optimize_for=carbon, Alice Router turns predicted energy into gCO₂ per billion tokens using the grid intensity you pass (grid_carbon_now). Automatic time-shifting to greener grid windows is on the roadmap — not live.
Hardware routing
Different models have different efficiency curves on different hardware. Alice Router recommends the GPU with the lowest certified energy for the workload, where certified points exist for that model on that GPU.
Length-aware prediction BETA
Prompt and answer length change the bill: on Qwen2.5-14B (H100), a 4,096-token prompt with a 128-token answer cost 2.6× more energy per token than a short exchange. A length-aware mode uses this; it stays behind a flag while it is validated.
Layer 01 — Certified training data

Every cert is a training point.

Alice Router doesn't train on datasheets. It trains on Serial Alice certificates — Ed25519-signed, Polygon-anchored measurements from real GPUs running real models. Each certificate contributes a CertifiedPoint: model, hardware, batch size, quantization, power (W), energy per token (µWh/tok), trust score.

Ed25519 signedPolygon anchoredtrust-weightedno simulations
certified-point · sa-d64b2e20 (signed fields)TRAINING
job_name vllm-phi3-bf16-b128
gpu NVIDIA H200
avg_power_w 516.9
gpu_energy_wh 8.81
trust_score 0.80 hardware_attested
tee_attested true · TDX · MRTD ✓ · TCB UpToDate
Layer 02 — Trust-weighted k-NN

Closer certs, higher trust, more weight.

When you request a prediction, Alice Router finds the k nearest certified configurations by model similarity, hardware match, and batch distance. Each point is weighted by model closeness × certificate trust score. Under policy evidence-weights-v1, TEE-attested certificates weigh 1.0, hardware-measured 0.35, and operator-declared points carry no weight. The prediction is a trust-weighted average of certified points.

k-NN on cert spacemodel-similarity distancetrust × evidence
router · weight computation (illustrative)
request: phi-3-mini · H200 · b=128 · optimize_for=energy
cert sa-d64b2e20 w=0.80 exact match · trust 0.80
cert sa-1751b088 w=0.52 b=64 · trust 0.80
cert sa-1a051cf4 w=0.34 b=16 · trust 0.80
cert results_h100 w=0.21 H100 · trust 0.55
predicted_power_w 516.9 W
confidence 0.86 (exact cert hit)
Layer 03 — Objective optimisation

Four modes. One honest answer.

Set your objective: energy, carbon, latency, or cost. Alice Router evaluates every feasible (batch, quant, hardware) configuration against the objective and picks the winner — with progressive constraint relaxation if the strict set is empty. The decision names its basis certificates, so you can check the maths.

energycarbonlatencycostprogressive relaxation
router · decision · optimize_for=energy
request: llama-3-8b · H100 · BF16 / GPTQ / AWQ
winner b=128 · GPTQ · H100
energy 10.7 µWh/token (measured)
vs BF16 −25% (14.2 µWh/token, b=128)
vs AWQ AWQ +145% (34.8 µWh/token, b=128)
basis sa-5f64133f, sa-65e9860b
tier hardware_measured
Layer 04 — The flywheel

More runs. Better model.
Better decisions.

Every certified run adds a point to the training set. The live Alice Router is a trust-weighted k-NN over 970+ certified points, including a 240-point certified H100 grid (5 models, 3B–32B, batch and length sweeps). A learned model (FT-Transformer ensemble) exists as an offline prototype; it will only serve predictions once it beats k-NN on models it has never seen.

k-NN · livelength-aware · beta flagFT-Transformer · offline prototype
ALICE ROUTER WORKLOAD RUNS SA CERTS ROUTER TRAINS BETTER PICKS
Version roadmap

Getting smarter
with every certified run.

v0.1 · live
Trust-weighted k-NN
970+ certified points: the H100 + B200 benchmark dataset, the Amália benchmark wave, the H200 Alice Benchmark and the 240-point H100 grid. Exact model lookup + progressive relaxation. Pure stdlib, no ML deps.
v0.2 · offline prototype
FT-Transformer prototype
Trained offline on H100 + B200 certified points. Not serving: held-out (unseen-model) error has not yet been validated against the k-NN baseline.
v0.3 · next
FT-Transformer on the full certified set
Retrain on the full certified set (incl. H200 and the 240-point H100 grid) and publish leave-one-model-out error next to k-NN. Promote only if it wins.
v0.4 · future
Online learning
Router updates weights continuously as new certs arrive. Tenant-specific fine-tune for your workload profile. Real-time carbon schedule optimisation.
API endpoints

One call.
A certified recommendation.

POST
/v2/router/recommend
Core router. Pass model_id, constraints, optimize_for. Returns recommended (batch, quant, hw) + predicted power + basis certificate IDs + confidence.
POST
/v2/router/plugin/vllm
Returns a copy-paste-ready vllm serve command with certified power evidence. Phi-3-mini on H200: direct TDX-certified overlay (trust=0.8).
POST
/v2/settlement/evaluate
Post-run settlement oracle. Evaluates a batch of certs against SLA conditions (energy, carbon, power, trust). Returns PASS/FAIL/PARTIAL + Ed25519-signed verdict for Polygon.
GET
/v2/router/dataset
Export the full certified training dataset as JSON. All certified points with cert IDs, so you can verify every data point the model trained on.

Every prediction.
Verifiable.

Alice Router names the certificates it used. Call /v2/certificates/{id}/verify on any of them. Each certificate belongs to a batch whose root is anchored on Polygon, so any later change to it is detectable.

See the benchmark data → Verify a certificate Get API access