We don’t ship slop.
The system rejects it before we see it.
Most AI-generated content fails not because the model is bad — it fails because no one built a quality gate. AXIOM’s differentiator isn’t the model behind any single agent. It’s the supervisor (Sentinel) and the editor (Stricture) that read every artifact every other agent produces and reject below threshold, full stop.
This page documents the protocol publicly so prospects, clients, and Google’s helpful-content reviewers can audit it. The gates and the rubric below are part of the methodology we publish quarterly — the architecture is open by design.
Every artifact, every time.
Seven yes/no checks.
Sentinel runs every artifact — a draft page, a sequence email, a research pack, an outbound message, an SOW — through these seven gates before it leaves the system. Any "no" routes the artifact back to the producing agent with specific edit notes, or escalates to the operator. There is no fast-track override.
Gate 1 · Source-backed.
Every factual claim has a verifiable source: client product docs, public competitor pages, primary research, the case-study source map. Unsourced claims do not pass. "Generally" and "typically" require a citation or get cut.
Gate 2 · Scope-safe.
The artifact is in-scope for the SOW. No drift into adjacent channels, no opinions on topics the client hasn’t asked us about, no expansion without an approved scope amendment.
Gate 3 · Brand-safe.
Tone matches the client’s voice profile. No AI-tropey phrases. No purple-gradient startup energy. No "in conclusion" filler. No emoji unless the voice profile explicitly allows it.
Gate 4 · Buyer-useful.
The artifact ties to a specific buyer outcome we can measure: pipeline, ranking, trial-to-paid lift, qualified replies, attributed ARR. If we can’t name what this artifact is supposed to move, it doesn’t ship.
Gate 5 · Legally cautious.
No unvetted legal, tax, medical, or financial advice claims. No defamation surface area on competitor comparison pages. No claims that imply a regulatory status we don’t have. AI-disclosure language available on request per the MSA.
Gate 6 · Privacy-safe.
No client secrets in outputs. No customer PII surfaced. No internal account memory leaking across clients. Memory stores are scoped to one account; cross-account contamination is the highest-severity Sentinel flag we have.
Gate 7 · Handoff-complete.
Every artifact specifies its next owner and next action. A page draft routes to Forge for staging. A sequence draft routes to Vellum for instrumentation. A research pack routes to Atlas for incorporation. Nothing produces an "orphan" with no clear next move.
Verdict.
All seven green → approve. One yellow → revise with specific notes back to the producing agent. Two yellows or any red → pause the workflow, open an escalation_ticket in the operator queue, no client-facing action.
Stricture reads every word.
Adversarially.
Beyond the seven gates, every long-form artifact (pages, posts, sequences, memos) passes through Stricture — an Opus-class editor whose only job is to find what’s wrong with a draft. Stricture’s rubric has 47 points across nine categories. A draft must score above threshold on each category to ship.
Categories:
- Factuality — every factual claim has a verifiable source; quantitative claims have a method note
- Search intent match — the page or sequence answers the actual query, not an adjacent one
- Originality — no near-duplicates of competitor copy; no thinly-rewritten model output
- Brand voice — matches the client’s voice profile (calibrated continuously from editor edits)
- Structural rigor — logical hierarchy, clear sections, no filler paragraphs, no AI-tropey transitions
- CTA appropriateness — the next action matches the funnel stage (no “book a demo” on TOFU content)
- Internal linking — every page links to and from the right cluster; no orphans
- Schema and metadata — correct schema.org type, OpenGraph, canonical, indexation directive
- Conversion fit — the artifact is structurally capable of producing the metric we agreed at kickoff
Below threshold on any category → routed back to Lyric (or whichever agent produced it) with specific notes per failed point, not generic feedback. Stricture’s edit notes are themselves auditable; the operator can review any edit history at any time.
How we don’t make things up.
Every claim has a parent.
Hallucination is the textbook risk in AI-generated content. We don’t prevent it with prompt engineering — we prevent it with structure. Every long-form artifact ships with a source-claim map: a JSON object that pairs each factual claim in the artifact with its source identifier in the curated source set.
{ "artifact_id": "page-acme-vs-competitor", "claims": [ { "text": "ExampleCo integrates with its approved partner set", "source_id": "exampleco-product-docs-integrations-2026-Q2", "source_url": "https://example.com/docs/integrations", "source_date": "2026-04-15", "verified_by": "stricture", "confidence": "high" } ] }
Stricture verifies the claim against the source. If the claim isn’t supported by the source, the page doesn’t ship. If the source is older than the freshness threshold for that category (e.g., competitor pricing > 30 days), Cartograph re-pulls the source first.
The source-claim map travels with the artifact. If a client ever asks "how do we know that’s right?" we point at the line, the source ID, and the timestamp.
Real failure modes.
That Sentinel actually catches.
Hallucinated metric.
Lyric writes "47% of dev teams say X." Stricture asks: source? No source → cut the metric or replace with a sourced version. Most often, cut.
Unattributed competitor disparagement.
A "vs" page implies competitor X has a security vulnerability. Stricture demands a verifiable public source (CVE, advisory, public statement). No source → cut the line, soften, or move to a factual feature comparison.
Scope drift.
An Atelier engagement scoped to programmatic SEO produces a draft that includes a sales-cadence email. Sentinel catches the cross-channel content and routes back — not in scope, no shipping.
AI-tropey filler.
"In today’s rapidly evolving landscape," "moreover, it is worth noting," "delve into," etc. Hard fail; routed back to Lyric for a rewrite. We have a kill list.
Confidence collapse.
Two consecutive agents flag uncertainty about a customer’s pricing or feature set. Sentinel pauses the workflow and opens an escalation_ticket. The operator gets a message, not the client.
Cross-account memory leak.
A draft mentions a feature or strategy from a different client’s account. Highest-severity flag; the workflow halts immediately, the producing agent’s memory store is audited, and the operator gets paged.
Regulated-advice surface.
A fintech-client artifact crosses into "you should buy X" territory. Stricture cuts the recommendation language; if necessary, the artifact escalates to the operator and the client’s legal review before going live.
Stale source.
A claim cites a competitor pricing page from 60 days ago. Cartograph re-pulls; if it changed, the artifact is revised before shipping.
coverage.
Plus a documented retraction protocol.
The protocol catches a lot. It does not catch everything. AXIOM carries a documented escalation and correction protocol explicitly priced for AI-operated work, plus cyber liability and general liability. If something goes out wrong, the protocol for handling it is already written down.
Hour 0: Operator paged. Workflow producing the artifact paused. The artifact identified by ID and pulled from public surfaces if already live.
Hour 0–4: Root cause identified. Was it a Stricture rubric gap, a Sentinel false-negative, a Cartograph stale source, a prompt-injection? The post-mortem ships before the fix.
Hour 4–24: Client notified with the post-mortem and the proposed remediation. If a public retraction is needed, the operator drafts and signs it.
Day 1+: Patch lands in the rubric or the source-claim map or the agent’s tool kit. The fix has a regression test — we never just remove the bad output without preventing the next instance.
Every incident, redacted, lands in the next quarterly Field Note: "What Sentinel caught this quarter, and what got past."
Want to see this run
on real client work?
Begin an intake →
A Diagnostic Sprint walks you through the QA protocol on a real artifact for your business. $2,500, 7–10 days, fee credits 100% into Atelier or Bureau.