SOX reference architecture for enterprise LLM deployment at scale — Reference Architecture

Reference architecture · published 2026-10-16 · regulated enterprise · SOX · layout walkthrough · implementation executed in CI before publish

The control that resolves this is a Tenant-Fair Inference Tier, evaluated before the request reaches a model provider rather than audited afterwards. In regulated enterprise the binding constraint is SOX, and the question it actually asks is can an AI-influenced number in a filing be traced and reproduced. The shape below holds p95 ≤ 450ms over gRPC with mTLS, and ships with a standard-library implementation that runs as written.

The control gap

A shared inference tier holds under average load and breaches its p95 under a neighbour's batch job. The SLA is written on a percentile, the dashboards show a mean, and admission control is a single global rate limit that protects the provider rather than any individual tenant.

What SOX specifically requires

SOX cares about the integrity of financial reporting. If a model output touches a number that reaches a filing, the control question is whether that output can be reproduced exactly and attributed to a reviewed, approved version.

The artifact that satisfies it is reproducible output records pinned to a model version. The way teams most often fail it: a model upgrade changing a reported figure with no record of which version produced the original.

The parts you have to build

1. Per-tenant token bucket — Sustained rate plus burst, so one tenant cannot spend another's budget.
2. Reservation ledger — Guaranteed in-flight slots per tenant; reservations must sum below the ceiling.
3. Global concurrency ceiling — Bounds the shared tier so it degrades predictably instead of collapsing.
4. Percentile window — Streaming p95/p99 per tenant — the number the contract is written on.
5. Queue with deadline — Requests carry a deadline and are shed rather than served late.
6. Shed path — 429 with an accurate Retry-After so clients back off instead of hot-looping.

Implementation artifact — Per-tenant admission control and percentile tracking

Standard-library Python, no external dependencies, runs as written. Executed by the publish gate: if it raises, this page does not ship. Wire it behind gRPC with mTLS; the failure mode it guards against is a half-open HTTP/2 stream silently dropping the deadline. The output you hand the assessor is the reporting-period reproducibility record.

"""Per-tenant admission control for a shared inference tier.

Why this shape: a global rate limit protects the provider and nobody else. The
failure enterprises actually hit is the noisy-neighbour one — tenant A's batch
job saturates the queue and tenant B breaches the p95 latency SLA that is
written into their contract. You cannot fix that downstream of admission.

Two controls, both required:
  * a per-tenant token bucket, so one tenant cannot consume another's budget
  * a global concurrency ceiling with per-tenant reservations, so the shared
    tier degrades predictably instead of collapsing
"""
import heapq
import time
from dataclasses import dataclass, field


@dataclass
class TokenBucket:
    rate: float            # sustained requests per second
    burst: float           # instantaneous allowance
    tokens: float = None
    updated: float = field(default_factory=time.monotonic)

    def __post_init__(self):
        if self.rate <= 0 or self.burst <= 0:
            raise ValueError("rate and burst must be positive")
        if self.tokens is None:
            self.tokens = self.burst

    def take(self, n: float = 1.0, now: float = None) -> bool:
        now = time.monotonic() if now is None else now
        elapsed = max(0.0, now - self.updated)
        self.tokens = min(self.burst, self.tokens + elapsed * self.rate)
        self.updated = now
        if self.tokens >= n:
            self.tokens -= n
            return True
        return False

    def retry_after(self, n: float = 1.0) -> float:
        """Seconds until n tokens exist. Return this in a 429 Retry-After so
        clients back off correctly instead of hot-looping."""
        deficit = max(0.0, n - self.tokens)
        return deficit / self.rate


class Admission:
    """Per-tenant buckets plus a global concurrency ceiling with reservations.

    reserved: tenant -> guaranteed in-flight slots. The sum of reservations
    must stay below `ceiling`, or the guarantee is fiction.
    """

    def __init__(self, ceiling: int, reserved: dict = None):
        self.ceiling = ceiling
        self.reserved = dict(reserved or {})
        if sum(self.reserved.values()) > ceiling:
            raise ValueError(
                f"reservations {sum(self.reserved.values())} exceed ceiling "
                f"{ceiling}: at least one tenant's SLA cannot be honoured")
        self.buckets: dict = {}
        self.inflight: dict = {}

    def configure(self, tenant: str, rate: float, burst: float):
        self.buckets[tenant] = TokenBucket(rate=rate, burst=burst)

    def admit(self, tenant: str, cost: float = 1.0) -> tuple:
        """-> (admitted, reason, retry_after_seconds)"""
        bucket = self.buckets.get(tenant)
        if bucket is None:
            return False, "tenant_not_provisioned", None

        used = sum(self.inflight.values())
        mine = self.inflight.get(tenant, 0)
        my_reservation = self.reserved.get(tenant, 0)

        # A tenant inside its reservation is admitted even when the shared pool
        # is full — that is what makes the reservation worth signing.
        if used >= self.ceiling and mine >= my_reservation:
            return False, "global_ceiling", 0.05

        if not bucket.take(cost):
            return False, "tenant_rate_limit", round(bucket.retry_after(cost), 3)

        self.inflight[tenant] = mine + 1
        return True, "ok", None

    def release(self, tenant: str):
        self.inflight[tenant] = max(0, self.inflight.get(tenant, 0) - 1)


class LatencyWindow:
    """Streaming p95/p99 over a bounded window.

    An average hides exactly the breach you are contractually exposed to: an
    SLA is written on a percentile, so the number you alert on must be one.
    """

    def __init__(self, size: int = 2048):
        self.size = size
        self.samples: list = []

    def observe(self, ms: float):
        self.samples.append(ms)
        if len(self.samples) > self.size:
            self.samples = self.samples[-self.size:]

    def quantile(self, q: float) -> float:
        if not self.samples:
            return 0.0
        ordered = sorted(self.samples)
        # Nearest-rank: the smallest sample at or above the q-th position.
        idx = min(len(ordered) - 1, max(0, int(round(q * len(ordered) + 0.5)) - 1))
        return ordered[idx]

    def breach(self, slo_ms: float, q: float = 0.95) -> bool:
        return self.quantile(q) > slo_ms


if __name__ == "__main__":
    adm = Admission(ceiling=10, reserved={"bank-eu": 4})
    adm.configure("bank-eu", rate=5, burst=5)
    adm.configure("retail-us", rate=50, burst=50)
    for _ in range(9):
        adm.admit("retail-us")
    print("noisy neighbour holds", adm.inflight)
    print("reserved tenant still admitted:", adm.admit("bank-eu"))
    print("burst exhausted ->", [adm.admit("bank-eu") for _ in range(5)][-1])

    win = LatencyWindow()
    for ms in [120, 140, 135, 128, 900, 131, 139, 142, 150, 2200]:
        win.observe(ms)
    print(f"p95={win.quantile(0.95):.0f}ms breach(400ms)={win.breach(400)}")

The SOX control, as code

Segregation of duties over model releases. This is the artifact an assessor asks for under SOX specifically — it is not the same control as the one above, and a system that implements one does not thereby implement the other. It executes under the same publish gate.

"""Segregation of duties over model releases, with an immutable approval chain.

SOX cares about change management over systems affecting financial reporting.
The control that fails is the one where the person who wrote the change also
approved and deployed it, and where the approval record can be edited after the
fact. Both are structural, so both are enforced here rather than reviewed.
"""
import hashlib
import json


class ApprovalChain:
    """Append-only. Each link commits to the previous digest, so an edited
    record breaks verification rather than quietly rewriting history."""

    def __init__(self):
        self.links: list[dict] = []

    def _head(self) -> str:
        return self.links[-1]["digest"] if self.links else "genesis"

    def append(self, action: str, actor: str, change_id: str) -> dict:
        body = {"action": action, "actor": actor, "change_id": change_id,
                "prev": self._head()}
        digest = hashlib.sha256(
            json.dumps(body, sort_keys=True).encode()).hexdigest()
        link = dict(body, digest=digest)
        self.links.append(link)
        return link

    def verify(self) -> bool:
        prev = "genesis"
        for link in self.links:
            body = {k: link[k] for k in ("action", "actor", "change_id", "prev")}
            if link["prev"] != prev:
                return False
            if hashlib.sha256(
                    json.dumps(body, sort_keys=True).encode()).hexdigest() != link["digest"]:
                return False
            prev = link["digest"]
        return True

    def actors(self, change_id: str, action: str) -> set[str]:
        return {l["actor"] for l in self.links
                if l["change_id"] == change_id and l["action"] == action}


class ReleaseGate:
    def __init__(self, chain: ApprovalChain):
        self.chain = chain

    def author(self, change_id: str, actor: str) -> None:
        self.chain.append("author", actor, change_id)

    def approve(self, change_id: str, actor: str) -> None:
        if actor in self.chain.actors(change_id, "author"):
            raise PermissionError(
                f"{actor} authored {change_id}; approval requires a second person")
        self.chain.append("approve", actor, change_id)

    def deploy(self, change_id: str, actor: str) -> dict:
        if not self.chain.actors(change_id, "approve"):
            raise PermissionError(f"{change_id} has no approval")
        if actor in self.chain.actors(change_id, "author"):
            raise PermissionError(f"{actor} authored {change_id}; cannot deploy it")
        return self.chain.append("deploy", actor, change_id)


if __name__ == "__main__":
    chain = ApprovalChain()
    gate = ReleaseGate(chain)

    gate.author("MODEL-2026-114", "r.khare")
    try:
        gate.approve("MODEL-2026-114", "r.khare")
        raise AssertionError("self-approval was permitted")
    except PermissionError:
        pass

    gate.approve("MODEL-2026-114", "a.controller")
    try:
        gate.deploy("MODEL-2026-114", "r.khare")
        raise AssertionError("author deployed their own change")
    except PermissionError:
        pass

    link = gate.deploy("MODEL-2026-114", "o.release")
    assert chain.verify(), "chain must verify after a legitimate sequence"
    assert link["action"] == "deploy"

    try:
        gate.deploy("MODEL-2026-999", "o.release")
        raise AssertionError("deployed an unapproved change")
    except PermissionError:
        pass

    # Tampering must be detectable, not merely discouraged.
    chain.links[1]["actor"] = "r.khare"
    assert not chain.verify(), "edited approval record still verified"

    print("SOX segregation of duties + approval chain: OK")
    print(f"  links {len(chain.links)}, verifies_after_tamper={chain.verify()}")

The blueprint

Tenant-Fair Inference Tier — the request path, with the decision points that must exist before the first model call:

flowchart TB
    subgraph Tenants
      TA[Tenant A - batch]
      TB[Tenant B - interactive p95 400ms]
    end
    TA --> AD{Admission}
    TB --> AD
    AD -->|bucket empty| S429[429 + Retry-After]
    AD -->|ceiling full,<br/>outside reservation| S429
    AD -->|admit| Q[Deadline queue]
    Q -->|deadline passed| SH[Shed - do not serve late]
    Q --> W[Worker pool]
    W --> P[Model / vLLM]
    P --> OBS[Per-tenant p95 / p99 window]
    OBS -->|breach| ALERT[Page + autoscale signal]

Deployment parameters

Every value below is a parameter you can hold a vendor or an internal team to. Adjectives are not acceptance criteria.

Risk register with acceptance criteria

Each row carries an acceptance parameter rather than an adjective, because "low latency" cannot be signed off. Retention here is bound by per the data class retention schedule, enforced automatically.

Questions to answer before approval

• Reproduce last quarter's figure exactly, using the model version that produced it.
  Version, prompt digest and output digest are recorded per reporting-relevant inference, and versions are pinned per period, so the figure is reproducible on demand.
• Which tenants have contractual percentile targets, and what reserves capacity for them?
  Any tenant with a contractual percentile gets an entry in the reservation ledger — guaranteed in-flight slots inside the global ceiling. Reservations are validated at configuration time and must sum below the ceiling.
• What happens to tenant B's p95 when tenant A sends 10x its normal volume — measured, not assumed?
  Tenant B stays inside its SLO, because tenant A is bounded by its own token bucket and cannot consume B's reserved slots. This is testable: saturate A and watch B's p95 window.
• Is the alert on a percentile or a mean?
  A percentile. The SLA is written on p95, so the alert is on p95. The mean is dashboard-only, because it hides exactly the breach you are contractually exposed to.
• What is shed under saturation, and how does a client learn when to retry?
  Requests past their deadline are shed rather than served late, and the shed path returns 429 with an accurate Retry-After computed from the bucket's refill rate so clients back off instead of hot-looping.
• Can a cache entry created for one tenant ever be served to another?
  No. Every cache key is namespaced by tenant id; keying on prompt text alone is the mechanism by which one tenant's completion is served to another.

See also

• Part of enterprise llm architecture at scale.
• PCI DSS 4.0 reference architecture for enterprise LLM deployment at scale covers the adjacent case.
• enterprise LLM deployment at scale in fintech covers the adjacent case.
• A different failure mode of the same component: HIPAA reference architecture for enterprise LLM deployment at scale.
• If this is the blocker in front of you: enterprise AI infrastructure review.