Fine-tune a customer compliance regime

Take the doctrine pretrain set, project new receipts through the a11oy Sentinel's σ-algebra of governance signals (ships inside a11oy), and emit a customer-specific Λ-extension that runs as an overlay — without touching the locked kernel.

Headline number: 1 customer regime, 0 changes to 749/14/163.

A "compliance regime" in SZL is not a model fine-tune that mutates the kernel. The kernel (c7c0ba17, 749/14/163) is LOCKED. Instead you define an overlay: a measurable function on the receipt-bus σ-algebra that tightens gates for your context (e.g., EU AI Act high-risk, NIST AI RMF; CMMC alignment is roadmap). This recipe builds and tests one.

Honest scope. The a11oy Sentinel has no standalone Space or repo. The live policy/drift inspector is exposed inside a11oy via /api/a11oy/v1/mcp/call (tool a11oy_gate). Λ axis scores come from the live killinchu gate. An overlay is your engineering policy on top of these signals — never a change to the locked kernel.

Honest framing. Λ is Conjecture 1, not a theorem. The overlay you build is an engineering policy on top of the gate; it inherits Λ's measured axis scores, not a proof of optimality. We say "extension," never "improvement to the proof."

---

Prerequisites

python3 -m pip install requests

No kernel write access required (and none is possible — that is the design).

---

Quickstart

import requests

# a11oy Sentinel ships inside a11oy.
A11OY = "https://szlholdings-a11oy.hf.space/api/a11oy"

# A customer regime = a set of axis floors + required gates, applied as an overlay.
regime = {
    "name": "acme-eu-ai-act-high-risk",
    "axis_floors": {"transparency": 0.95, "fairness": 0.95, "auditability": 0.95},
    "required_gates": ["dual-use", "STIX/TAXII", "receipt-hash"],
    "inherit": "doctrine-v11",         # never rewrites the kernel
}

# Project a candidate action through the a11oy gate; overlay decides PASS/FAIL on top of the base verdict.
action = {"kind": "model_output", "text": "…", "context": {"sector": "credit_scoring"}}
base = requests.post(f"{A11OY}/v1/mcp/call",
                     json={"tool": "a11oy_gate", "args": {"action": action}}, timeout=60).json()
print("base verdict:", base.get("verdict"))

The overlay never asks the gate to change; it reads the per-axis signals and applies stricter floors locally.

---

Full walkthrough

Step 1 — The σ-algebra of governance signals

Every governed action produces a vector of 13 trust-axis scores (the yuyay_v3 axes: soundness, calibration, robustness, provenance, consent, reversibility, transparency, fairness, containment, attestation, freshness, authority, auditability). These axes generate a σ-algebra (\Sigma) over the receipt bus: events are measurable sets like "fairness < 0.9 ∧ sector = credit." A regime is a measurable predicate on (\Sigma).

AXES = ["soundness","calibration","robustness","provenance","consent","reversibility",
        "transparency","fairness","containment","attestation","freshness","authority","auditability"]

Step 2 — Pull the doctrine pretrain set (axis priors)

The doctrine axis priors come from the locked kernel and the lake. Use the live Λ axis snapshot as the canonical prior (this is exactly what the gate uses):

import requests
axes = requests.get("https://szlholdings-killinchu.hf.space/api/killinchu/v1/lambda", timeout=60).json()
prior = {a["name"]: a["score"] for a in axes["axes"]}
print("Λ prior:", round(axes["lambda"], 6), "floor:", axes["lambda_floor"], "pass:", axes["pass"])
# => Λ prior: 0.922181 floor: 0.9 pass: True

Step 3 — Define the Λ-extension (overlay)

def lambda_extension(axis_scores: dict, regime: dict) -> dict:
    floors = {**{a: regime["inherit_floor"] for a in AXES}, **regime["axis_floors"]}
    failures = [a for a in AXES if axis_scores.get(a, 0.0) < floors[a]]
    # geometric mean (yuyay_v3 canonical aggregate) — same operator as the base gate
    import math
    geo = math.exp(sum(math.log(max(axis_scores.get(a, 1e-9), 1e-9)) for a in AXES) / len(AXES))
    return {"regime": regime["name"], "lambda_overlay": round(geo, 6),
            "pass": not failures, "failed_axes": failures, "inherits": "c7c0ba17"}

regime = {"name": "acme-eu-ai-act-high-risk", "inherit_floor": 0.90,
          "axis_floors": {"transparency": 0.95, "fairness": 0.95, "auditability": 0.95}}
print(lambda_extension(prior, regime))
# Fairness/transparency 0.90 < 0.95 → fails the tightened regime, as intended.

Step 4 — Bind the regime to a receipt

Wrap the overlay decision as a Khipu receipt so it is itself auditable (verify it with recipe 01). Submit candidate receipts back through the a11oy gate (/v1/mcp/call) for the chained, signed base verdict; store the overlay decision alongside.

Step 5 — Validate against your corpus

Replay your historical receipts through the overlay and report the confusion matrix vs. your manual compliance labels. Pair this with recipe 09 — PAC-Bayes confidence margin to put a tail bound on the regime's empirical risk.

---

See also

• 09 — PAC-Bayes confidence margin — bound the regime's risk.
• 12 — Doctrine ledger query — read the locked priors.
• 01 — Verify a receipt end-to-end
• Live: a11oy gate (hosts the a11oy Sentinel) · killinchu Λ axes

Cite this recipe

@misc{szl_cookbook_compliance_regime_2026,
  title        = {Fine-tune a customer compliance regime (SZL Cookbook recipe 03)},
  author       = {{SZL Holdings}},
  year         = {2026},
  howpublished = {\url{https://github.com/szl-holdings/szl-cookbook/blob/main/recipes/03-fine-tune-compliance-regime.md}},
  note         = {Λ = Conjecture 1; overlay never mutates the locked kernel c7c0ba17.}
}

---

Doctrine v11 LOCKED — 749/14/163 — kernel c7c0ba17 · Λ = Conjecture 1 · SLSA L1 (honest)