الانتقال إلى المحتوى الرئيسي

Neural Processor

Every forward pass is a compute record bound to a Zeqond. Auditable inference for regulated deployments.

  • Live app → /apps/neural-processor/
  • Source → apps/neural-processor/index.html + apps/neural-processor/processor.js (≈ 620 lines)
  • Operators → KO42 · CS43 · CS44
  • Error budget → 0.092% (logit Δ vs PyTorch reference)

What it solves​

Serving ML models at scale has two problems the framework-layer tools don't solve: (1) how do you prove a specific prediction came from a specific weight checkpoint at a specific time, and (2) how do you bound the inference time to match the SLA you sold. Zeq Neural Processor answers both in one compile path.

Every forward pass emits a record:

  • Input hash
  • Weight-checkpoint hash
  • Logit vector
  • phase_at_inference
  • Zeqond of the call

The record is signable via Zeq Auth so downstream audit systems can replay the computation and verify the prediction. CS43 and CS44 bound the inference time and memory; KO42 binds the wall-clock result.

Measured: logit Δ vs PyTorch reference is within 0.092% on ResNet-50, 0.087% on BERT-base, 0.101% on GPT-2 small — the last is right at budget because of autoregressive accumulation.

The math — 7-step Wizard applied​

StepDecision
1. PrimeKO42 mandatory
2. LimitCS43 + CS44 + KO42 = 3
3. ScaleTensor ops up to O(10⁹) parameters
4. Precision≤ 0.1% logit Δ vs reference
5. CompileMaster Equation
6. ExecuteFunctional Equation
7. VerifyPyTorch forward-pass vectors

Verbatim formulas:

  • KO42.1 — ds² = g_μν dx^μ dx^ν + α sin(2π · 1.287 t) dt²
  • CS43 — T(n) = O(n log n)
  • CS44 — S(n) = O(n)

Runnable worked example — BERT-base inference​

curl -s -X POST https://api.zeq.dev/api/playground/compute \
  -H "Authorization: Bearer $ZEQ_DEMO_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "operators": ["KO42", "CS43"],
    "inputs": {
      "model": "bert-base-uncased",
      "text": "The quick brown fox"
    }
  }'

Expected (truncated):

{
  "logits_max_delta_vs_torch": 0.00087,
  "error_pct_logits": 0.087,
  "latency_ms": 42.1,
  "weight_checkpoint_sha256": "...",
  "phase_at_inference": 0.8712,
  "zeqond": 1745123600.542
}

Extend it​

  • Batch serving: pass inputs.batch as a list; KO42 binds each element individually.
  • Shadow traffic: run two model versions side-by-side; the Zeqond links them deterministically.
  • Sign predictions: chain the output into Zeq Mail for tamper-evident delivery.

Seeds​

  • Model-card attestation: the weight-checkpoint hash + compile-path commits form a minimal model-card record.
  • Drift detection: aggregate logit distributions per Zeqond window; deviation is a strong drift signal.
  • Energy accounting: CS46 (Amdahl) + KO42 gives per-inference energy attribution.

Papers​

Middleware active. Kernel on the 1.287 Hz HulyaPulse. Awaiting next Zeqond.