Governed speculative decoding — speedup model + receipt

Probe the speculative-decoding surface: get the acceptance-rejection cost curve, the expected per-step token yield, and a receipted snapshot — all honestly labeled MODELED until a same-family draft+target pair is live on the sovereign tower.

Headline number: GET /specdec/health → label: "SPEC-DECODE ROADMAP" (draft not yet pulled) or "SPEC-DECODE MEASURED" once the llama3.2:1b draft is present on the tower.

Speculative decoding (Leviathan et al. 2023, arXiv:2211.17192; Chen et al. 2023, arXiv:2302.01318) pairs a small draft model with a large target — the draft proposes k tokens in one pass; the target accepts or rejects them via sampling-preserving verification. Throughput scales as (1 + k\alpha) where α is the per-token acceptance rate. The SZL surface computes this curve, stores the speedup accounting, and wraps the whole result in a governance receipt.

Honest posture. Until both draft and target are on the same sovereign node (shared tokenizer required), the surface returns label: "SPEC-DECODE ROADMAP" and measured: "UNAVAILABLE". The acceptance-rate curve is MODELED (Leviathan Lemma 1 applied to the configured k). Λ = Conjecture 1 (advisory). No joules fabricated. No throughput fabricated.

---

Prerequisites

python3 -m pip install httpx

Live base: https://a-11-oy.com. Check current state:

curl -s https://a-11-oy.com/api/a11oy/v1/specdec/health \
  | python3 -m json.tool \
  | grep -E '"status"|"label"|"same_family_pair"|"measured_run_possible"'

Expected (tower down today):

"status": "LIVE",
"label": "SPEC-DECODE ROADMAP (no same-family pair)",
"measured_run_possible": false

---

Quickstart — run the surface

import httpx, json

BASE = "https://a-11-oy.com"

# POST a prompt (tower reachable → MEASURED; tower down → MODELED cost curve)
r = httpx.post(
    f"{BASE}/api/a11oy/v1/specdec/run",
    json={"prompt": "The sovereign GPU mesh uses governed speculative decoding to"},
    timeout=60,
).json()

# Always present (even when tower is down)
acctg  = r["payload"]["accounting_modeled"]
k      = acctg["k"]
label  = r["payload"]["label"]
digest = r["payload_sha3_256"]

print(f"k (draft steps):         {k}")
print(f"State label:             {label}")
print(f"Expected speedup @ α=0.7: {acctg['expected_speedup_at_alpha']['0.7']:.3f}×")
print(f"Receipt digest (SHA3-256): {digest}")
print(f"Signed: {r['signature']['signed']}")

# Offline verify the receipt (zero server round-trip)
import json as _j, hashlib
payload_bytes = _j.dumps(
    r["payload"], sort_keys=True, separators=(",", ":"), ensure_ascii=False
).encode("utf-8")
recomputed = hashlib.sha3_256(payload_bytes).hexdigest()
assert recomputed == digest, "RECEIPT TAMPERED"
print("Offline receipt verify: OK ✓")

---

Full walkthrough

Step 1 — health check

h = httpx.get(f"{BASE}/api/a11oy/v1/specdec/health", timeout=15).json()
print(h["same_family_pair"])   # {"found": false, "reason": "tower unreachable"} until pair pulled

Step 2 — cost curve (always available)

The accounting_modeled block applies Leviathan et al. 2023 Lemma 1 (cited, not re-derived):

α (acceptance rate)	Expected tokens/step	Expected speedup
0.5	1.94	1.08×
0.7	2.77	1.54×
0.9	4.10	2.28×

These are MODELED projections — not measurements. Honest label: "MODELED".

Step 3 — offline receipt verification

Every /specdec/run response includes a payload_sha3_256 field and a digest_canonicalization note. Verify in one line:

import json, hashlib
payload_bytes = json.dumps(
    r["payload"], sort_keys=True, separators=(",", ":"), ensure_ascii=False
).encode()
assert hashlib.sha3_256(payload_bytes).hexdigest() == r["payload_sha3_256"]

Step 4 — activate MEASURED mode (founder / tower operator)

When the tower (RTX 4060 Ti, https://gpu.a-11-oy.com) is reachable and holds a same-family draft+target pair, the endpoint auto-detects and runs a MEASURED block:

# On the tower
ollama pull llama3.2:1b    # draft (same llama family as existing llama3.1:8b)
ollama pull llama3.1:8b    # target (likely already present)

# Confirm
curl -s https://gpu.a-11-oy.com/api/tags | python3 -m json.tool | grep '"name"'

After this, POST /specdec/run returns label: "SPEC-DECODE MEASURED" with a real accepted_rate, mean_accepted_len, and speedup_modeled derived from live token sampling.

---

Expected response shape

{
  "schema": "szl.a11oy.specdec.receipt.v1",
  "payload": {
    "schema": "szl.a11oy.specdec.v1",
    "label": "SPEC-DECODE ROADMAP",
    "measured": "UNAVAILABLE",
    "accounting_modeled": {
      "label": "MODELED (acceptance-rejection accounting curve, not a measurement)",
      "k": 4,
      "expected_speedup_at_alpha": {"0.5": 1.0764, "0.7": 1.5406, "0.9": 2.2751}
    }
  },
  "payload_sha3_256": "<sha3-256 of canonicalised payload>",
  "signature": {"signed": false, "status": "DSSE_PLACEHOLDER"}
}

When tower is live and same-family pair present:

{
  "payload": {
    "label": "SPEC-DECODE MEASURED",
    "measured": "MEASURED",
    "accepted_rate": 0.73,
    "mean_accepted_len": 3.1,
    "speedup_modeled": 1.63,
    "n": 50,
    "draft_model": "llama3.2:1b",
    "target_model": "llama3.1:8b",
    "quality_delta": {"identical_rate": 1.0, "label": "MEASURED (lossless)"}
  }
}

---

Doctrine

Field	Value
Locked-proven	8 ({F1,F4,F7,F11,F12,F18,F19,F22})
Λ	Conjecture 1 (advisory, ≤0.99, NOT a theorem)
Energy	UNAVAILABLE until NVML meter live
Throughput	MODELED until same-family pair on tower
Citations	arXiv:2211.17192 (Leviathan 2023); arXiv:2302.01318 (Chen 2023); arXiv:2406.02532 (SpecExec, MIT reference); arXiv:2402.12374 (Sequoia, MIT reference)

---

Doctrine v11 LOCKED · Λ = Conjecture 1 · SLSA L1