The L3 Ch8 shop has the brand voice locked (Lesson 1), structured output flowing into ServiceTitan / Sera / HCP / FieldEdge / BuildOps (Lesson 2), and the dashboard reading green. The discipline holds in week one. By week six it has started to drift — the Avoca booking confirmation that used to skim cleanly now buries the kill reason in a paragraph, the Rilla scorecard that was sharp now reads generic, the Birdeye AI Employee review response that was on-brand has begun committing to remedies. Drift is the default. The Cardinal Rule's 30-second verify pass catches per-artifact fabrications; it does not catch slow voice degradation, judgment drift, or quality regression across a workflow. The named workflow that catches drift is the weekly AI output audit: sample 10 AI-generated artifacts per week per workflow, score against the four dimensions (accuracy, voice, judgment, format conformance), build the audit log, and feed the patterns back into the system-prompt library. This lesson is the build. The artifact selection rubric (which 10), the four-dimension scoring rubric, the audit log schema, the per-workflow cadence, the Friday huddle that turns the audit into action, and the quarterly trend review that surfaces drift before it costs revenue. The L3 service manager runs the audit weekly; the discipline holds; the L3 Ch8 stack produces consistent output across quarters rather than degrading across weeks.

Why Per-Artifact Verify Is Not Enough

The Cardinal Rule's 30-second verify pass is built for the per-artifact decision — does this Avoca booking confirmation match the customer record, does this Rilla scorecard match what actually happened in the ride-along, does this AI-drafted financing summary match the Wisetack portal output. The pass catches fabrications at the moment the artifact ships. It does not catch trend-level degradation: the AI's voice slowly chatbot-izing across three weeks, the AI's judgment quietly relaxing on commitment language across a month, the AI's format quietly drifting on enum values across a quarter. The per-artifact verifier is looking at this artifact; the trend-level shift happens across many artifacts and is invisible to any individual verify pass.

The 2026 ServiceTitan State of AI report's 12% AI-embedded number is, at the L3 stack level, the audit-discipline number. Shops that deploy AI without weekly audits drift toward generic output, commitment-language slip, voice degradation, and structural fabrications the per-artifact verify cannot catch. The drift compounds — customers receive AI artifacts that no longer match the brand voice; review responses commit to remedies the shop did not authorize; coaching scorecards produce generic talk-track suggestions that no longer match the shop's rebuttal library; AI-drafted estimate data carries slowly-fabricated SEER ratings the verify pass started to skim because most of the time it was right. By the time the dashboard's headline metrics surface the degradation, three months of customer artifacts have shipped at degraded quality. Re-installing the discipline costs more than maintaining it.

The weekly audit is the trend-surveillance layer. The L3 service manager pulls 10 artifacts per workflow, scores against the rubric, identifies patterns the per-artifact verify missed, queues system-prompt updates for Monday, and documents the audit log for the quarterly governance review. The discipline is small — 90 minutes per week for a 5-workflow L3 stack — and the leverage is large: the audit catches drift in week three rather than month six, and the shop's AI artifacts maintain consistent quality across quarters rather than degrading.

The Audit Sample: Which 10 Artifacts

The audit samples 10 artifacts per workflow per week. The sample is not 10 random artifacts; it is a stratified sample designed to surface drift across the workflow's failure modes. Five strata, two artifacts per stratum.

Stratum one: high-volume routine. Two artifacts from the workflow's most common operating mode — a typical Avoca booking confirmation, a standard Rilla scorecard, a routine AI-drafted call summary. These artifacts represent the workflow's bulk; drift in this stratum is the most expensive because it affects the largest customer-facing surface.

Stratum two: edge cases. Two artifacts from the workflow's unusual operating mode — an Avoca booking that involved a name correction, a Rilla scorecard from a ride-along that ran long, a call summary where the customer was upset. Edge cases surface AI's robustness; drift in edge cases is the leading indicator of drift in high-volume routine.

Stratum three: high-stakes. Two artifacts where the customer commitment was material — an Avoca booking for a $14K-plus replacement, an AI-drafted financing summary, a Birdeye AI Employee review response on a negative review, a Rilla scorecard from a Comfort Advisor's $42K commercial proposal. High-stakes artifacts carry the largest exposure if drift slipped past the per-artifact verify.

Stratum four: caught at verify. Two artifacts that the per-artifact verify pass caught a fabrication on — pulled from the "AI Caught a Hallucination" bulletin board. The audit reviews these to confirm the verify caught what it should have caught and the remediation actually corrected the system prompt. Without the post-catch audit, fixes are believed-but-not-verified; with it, the remediation discipline closes.

Stratum five: random. Two artifacts pulled at random from the past week's full set. Random sampling catches what the stratified sample's failure-mode bias would miss; the audit's random stratum is the safety net that surfaces unknown failure modes.

The Four-Dimension Scoring Rubric

Each sampled artifact is scored across four dimensions on a five-point scale. The rubric is the same across every workflow in the L3 stack; consistency across workflows is what allows the L3 service manager to compare drift patterns across Avoca, Rilla, Birdeye AI Employee, Hatch, ResponsiBid, and the AI-drafted estimate data workflow. Four dimensions, five points each, twenty points total per artifact, 200 points per workflow per week.

Dimension One: Accuracy

Does the artifact match source-of-truth on the numbers, names, parts, warranty terms, and financing/regulatory language? Five-point scale: 5 = every fact verified against source-of-truth and correct; 4 = minor non-critical drift (a name spelling off, a non-material number wrong); 3 = material fact wrong but caught at verify; 2 = material fact wrong and shipped to customer; 1 = systematic fabrication pattern. Anything below 4 surfaces in the Friday huddle; below 3 triggers immediate system-prompt update queued for Monday; below 2 triggers an "AI Caught a Hallucination" entry and a quarterly governance review item.

Dimension Two: Voice

Does the artifact match the L3 Ch8 Lesson 1 brand-voice system prompt? Trades English, named tools, shop-specific phrasing, owner-approved tone, no chatbot tells ("I'd be happy to assist," "Please don't hesitate," "As an AI"). Five-point scale: 5 = indistinguishable from shop-staff-written; 4 = minor chatbot tell or generic phrase that does not break the brand; 3 = noticeably AI-flavored but still on-brand; 2 = generic chatbot voice; 1 = the customer would recognize as AI-generated. Voice drift is slow; the audit catches it in week three rather than month six.

Dimension Three: Judgment

Did the AI make the right judgment call given the artifact's context? Did the AI escalate the Birdeye AI Employee review response that contained a commitment ("we'll refund the full amount") rather than auto-posting? Did Avoca correctly recognize the after-hours emergency vs. routine call? Did Rilla flag the ride-along moment that was actually a coaching opportunity? Did the AI-drafted estimate include the warranty-term-out-of-range exception flag? Five-point scale: 5 = AI made the same judgment a senior team member would; 4 = minor judgment drift acceptable for the artifact's stakes; 3 = judgment call AI got wrong but verify caught; 2 = wrong judgment shipped to customer with material exposure; 1 = systematic judgment failure pattern. Judgment drift produces commitment-language slips, escalation failures, and FTC / Reg Z / FCRA exposure; the audit's judgment dimension is the regulatory canary.

Dimension Four: Format Conformance

Does the artifact conform to the structured-output schema from Lesson 2? JSON parseable, enums in the allowed set, required fields present, nulls properly encoded, FSM-platform-conforming. Format conformance is largely machine-checkable at the integration layer; the audit's role is to catch the format drift the integration's validation missed because the validation rules drifted with the schema. Five-point scale: 5 = schema-perfect, integration ingests cleanly; 4 = minor format tells that did not break ingest; 3 = format issue that required manual fix at ingest; 2 = format issue that broke the ingest pipeline; 1 = systematic format failure pattern across the sample. Format drift produces FSM data quality degradation; the audit catches it before the dashboard's diagnostic supplements signal the downstream effect.

The Audit Log Schema

Every audit produces a row in the audit log. The log is the artifact that survives the quarterly L3 governance review, the PE board diligence pull, and the Authority Brands / Wrench Group / Nexstar QBR. The schema is consistent across every workflow in the L3 stack; the L3 service manager's library has the canonical schema documented and version-controlled.

Per-artifact row fields: audit_date (ISO 8601), workflow_name (Avoca, Rilla, Birdeye AI Employee, Hatch, ResponsiBid, etc.), artifact_id (reference to FSM object or integration log), stratum (high-volume routine / edge case / high-stakes / caught-at-verify / random), accuracy_score (1-5), voice_score (1-5), judgment_score (1-5), format_score (1-5), total_score (sum, max 20), notes (free text, short), pattern_flag (boolean — does this drift represent a pattern across multiple artifacts this week), system_prompt_update_queued (boolean), reviewer_id (L3 service manager).

The log is reviewed at three cadences. Friday: this week's 50 artifact rows (10 per workflow × 5 workflows) reviewed in the 30-minute Friday huddle, system-prompt updates queued for Monday. Monthly: the L3 service manager pulls a trailing-4-week view of total_score by workflow and notes whether scores are stable, drifting up, or drifting down. Quarterly: the L3 governance review pulls a trailing-13-week view, identifies the workflows with score trends, decides the system-prompt or vendor-configuration interventions, and documents the quarterly governance memo.

The Named Workflow: The Friday 90-Minute Audit

The L3 service manager's Friday audit takes 90 minutes for a 5-workflow L3 stack and produces the audit log entry, the system-prompt update queue, and the Friday huddle agenda. The workflow has a name — the Friday audit — and a fixed cadence: 1:30 p.m. Friday afternoon, before the L3 Ch2 Friday 5 p.m. dashboard email goes out, after the morning's tech audit (L3 Ch7 weekly CSR audit) wraps. The cadence is the discipline; missing the cadence is the first warning sign of L3 Ch8 stack drift.

Step One: Pull (15 min)

The L3 service manager pulls the 50 artifacts (10 per workflow × 5 workflows). Pulls are scripted where possible — a Zapier or n8n flow that samples randomly from the past week's structured-output logs, stratifies per the rubric, and produces a 50-artifact link list. For workflows where automated pull is not yet built, the manual pull takes 5-10 minutes. The artifacts land in a shared review folder; the audit log spreadsheet is pre-populated with the 50 rows ready for scoring.

Step Two: Score (45 min)

The L3 service manager scores each artifact against the four-dimension rubric. 50 artifacts × ~50 seconds per artifact = 45 minutes. The scoring discipline is to read the artifact, cross-reference source-of-truth if needed (manufacturer spec sheet, Wisetack portal output, customer record, brand-voice prompt for voice scoring), and assign scores. Notes are short — "name spelling drift," "chatbot tell in greeting," "missed escalation on commitment." Pattern flags are set when the same drift appears across 3+ artifacts in the week.

Step Three: Pattern Analysis (15 min)

The L3 service manager reviews the scored sheet, identifies the patterns (which workflows are drifting on which dimensions), queues the system-prompt updates for Monday, and writes the Friday huddle agenda. Three-to-five pattern entries per week is typical; more than five suggests a workflow-level issue that needs deeper attention.

Step Four: Friday Huddle (15 min)

The L3 service manager runs a 15-minute Friday huddle with the team — CSR, dispatcher, advisor, ops manager, marketing manager — sharing the patterns and walking through the system-prompt updates queued for Monday. The huddle is the team-facing communication that surfaces drift before customers see it; it also surfaces the team's intuitions about what is going wrong, which often catch failures the L3 service manager's audit missed. The huddle's 15 minutes is non-negotiable; without it the audit becomes back-office paperwork.

The Monday System-Prompt Update Cycle

The Friday audit produces a system-prompt update queue. Monday morning the L3 service manager works through the queue. Each update follows the same shape: identify the drift pattern, write the system-prompt revision, test against the past week's drifted artifacts to confirm the revision would have caught them, deploy the revision to the vendor's prompt configuration, log the update in the L3 prompt library's version history. Three to five updates per week is typical; the update cycle takes 30-60 minutes Monday morning.

System-prompt updates are not free-form. The L3 service manager's library has the update template: drift_pattern (string), system_prompt_section (which section of the prompt is affected), prior_text (the text being revised), new_text (the revised text), test_artifacts (links to the prior-week artifacts the revision was tested against), test_outcome (did the revision catch the drift in the test artifacts), deploy_date (ISO 8601), deploy_status (in production / monitoring / rolled back). The library is version-controlled — Git, Notion, or a shared Google Doc with revision history — so quarterly governance can review the update trail and roll back any update that produced unintended consequences.

Quarterly Trend Review and Governance

The L3 governance review (90 minutes quarterly) pulls a trailing-13-week view of the audit log and analyzes trends. Three analyses per workflow: total_score trend (stable / improving / degrading), dimension drill-down (which of accuracy / voice / judgment / format is moving), pattern_flag rate (how many weeks per quarter had pattern flags). Three workflow-level decisions are made per quarter: keep the workflow as-is (stable scores), deepen intervention (degrading trend triggers system-prompt rewrite or vendor configuration update), retire the workflow (sustained degrading trend triggers vendor change or workflow retirement).

The quarterly governance memo documents the trend analysis, the decisions, and the rationale. The memo is the artifact the PE board, the Authority Brands franchisor, the Wrench Group portfolio reviewer, the Nexstar peer group, or the M&A diligence team reads. The memo's narrative is: this is what we audited, this is what we found, this is what we changed, this is the result. Documented, signed, auditable. The discipline is the leverage; the memo is the discipline made visible at the governance level.

How the Weekly Audit Integrates with the L3 Stack

The weekly AI output audit is the third leg of the L3 Ch8 four-lesson stack — brand voice (Lesson 1), structured output (Lesson 2), weekly audit (this lesson), documentation and audit trail (Lesson 4). The audit consumes the brand-voice and structured-output artifacts and produces the accuracy signal the documentation/audit trail records. Without the brand voice the audit has no voice baseline; without structured output the audit's format-conformance dimension has nothing to score against; without the audit the documentation/audit trail has no accuracy signal to log.

The audit also feeds the L3 Ch2 Friday dashboard. The five headline numbers (booking %, after-hours capture, cost per booked call, abandoned-call %, AI-recovered revenue) are workflow output metrics; the audit's total_score per workflow is the quality input metric. When the dashboard's headline metrics drift down, the audit log shows whether the cause is workflow quality degradation (audit total_score dropping) or external factors (call volume, marketing mix, seasonality). The audit is the leading indicator; the dashboard is the lagging indicator. The L3 service manager reads both.

The audit also closes the loop with the Cardinal Rule's "AI Caught a Hallucination" bulletin board. The board's per-catch entries feed the audit's stratum-four sample (artifacts caught at verify). The audit confirms the catches were appropriate and the remediations stuck. Together the board and the audit are the bottom-up (per-artifact catches) and top-down (per-workflow patterns) view of AI quality discipline.

Failure Modes and the Quarterly Rollback Rule

Three failure modes kill the weekly audit's value. The L3 service manager watches for each one. Failure Mode One: the audit becomes paperwork. Scores are entered but no system-prompt updates are queued; the huddle becomes a status update; nothing changes Monday. Detection: the system_prompt_update_queued boolean is False for 3+ consecutive weeks despite drift patterns surfacing. Fix: the L3 service manager re-grounds on the workflow — the audit exists to drive change, not to log scores.

Failure Mode Two: the sample is biased. The L3 service manager unconsciously samples the workflows they have time to review rather than the stratified sample. Detection: the stratum distribution in the audit log is not the rubric's 2-2-2-2-2; high-stakes and caught-at-verify strata are under-represented. Fix: scripted sampling that enforces the rubric's stratification.

Failure Mode Three: the rubric is gamed. The L3 service manager scores leniently to avoid Monday's update workload, or scores harshly to justify deeper intervention they wanted independently. Detection: score distribution skews tightly (4s and 5s only, or 2s and 3s only) rather than spreading across the five-point scale. Fix: quarterly governance reviews the score distribution and re-calibrates the rubric; the L3 service manager benchmarks against a peer L3 service manager (Nexstar, BDR, Service Champions, peer-group rounds) where possible.

The quarterly rollback rule: if a workflow's audit total_score has been below 14 (out of 20) for 8+ consecutive weeks, the workflow rolls back. Vendor change, prompt rewrite, or workflow retirement — the L3 governance review decides which. The 8-week threshold is the L3 Ch8 stack's quality cliff: below that, the workflow is producing more drift than the verify discipline can absorb, and continuing to operate it produces customer-facing exposure the shop cannot defend.

The Weekly Audit and the L3 Shops That Survive AI

Per-artifact verify catches today's fabrications. The weekly audit catches this month's drift. The quarterly review catches this year's degradation. The L3 shop that runs all three operates AI as institutional infrastructure rather than a content factory. The shops that run only the per-artifact verify produce customer-facing artifacts that drift toward generic chatbot output across quarters; the verify catches the per-artifact failures but not the workflow-level decay.

The discipline is small. 90 minutes per week for the audit. 30-60 minutes Monday morning for system-prompt updates. 90 minutes per quarter for the governance review. Four hours per week and one full day per quarter — about 200 hours per year at the L3 stack scale. The leverage is large: the audit catches drift in week three rather than month six, the customer-facing artifacts maintain consistent quality across quarters, the L3 Ch2 dashboard's headline metrics hold against industry baseline, and the PE board's diligence pull surfaces a defensible operating-system artifact rather than a workflow stack that has quietly drifted into generic output.

Build the rubric. Stratify the sample. Score consistently. Document the log. Run the Friday huddle. Update the prompts Monday. Review the trends quarterly. The L3 service manager who runs the weekly audit makes the rest of the L3 Ch8 stack durable. The shops that skip the audit drift; the shops that run it compound metric movement quarter over quarter. The weekly AI output audit is the trend-surveillance layer of L3 Ch8 quality discipline. Build it.

Key Takeaways

• Per-artifact verify is not enough — The Cardinal Rule's 30-second pass catches today's fabrications; it does not catch this month's drift. The weekly audit is the trend-surveillance layer.
• 10 artifacts per workflow per week — Stratified sample across five strata (high-volume routine, edge cases, high-stakes, caught-at-verify, random), 2 artifacts per stratum.
• Four-dimension scoring rubric — Accuracy, voice, judgment, format conformance. Five-point scale each. 20 points per artifact, 200 points per workflow per week.
• The audit log schema — audit_date, workflow_name, artifact_id, stratum, four scores, total_score, notes, pattern_flag, system_prompt_update_queued, reviewer_id. Version-controlled across quarters.
• The Friday 90-minute audit — Pull (15 min), score (45 min), pattern analysis (15 min), Friday huddle (15 min). Named workflow at 1:30 p.m. Friday.
• Monday system-prompt update cycle — Work through the queue. 30-60 minutes. Test revisions against prior-week drifted artifacts. Deploy to vendor prompt config. Log in version-controlled library.
• Quarterly trend review — 90 minutes. Trailing-13-week view by workflow. Total_score trend, dimension drill-down, pattern_flag rate. Three workflow-level decisions per quarter (keep / deepen intervention / retire).
• The quarterly rollback rule — A workflow at total_score below 14 (of 20) for 8+ consecutive weeks rolls back. Vendor change, prompt rewrite, or workflow retirement decided at L3 governance.
• Three failure modes — Audit becomes paperwork (no updates queued), sample is biased (stratification not enforced), rubric is gamed (scores skew tightly). Each has a detection trigger and a fix.
• Integrates with the L3 Ch8 stack — Consumes brand voice (Lesson 1) and structured output (Lesson 2); produces accuracy signal documentation/audit trail (Lesson 4) records.
• Integrates with the L3 Ch2 dashboard — Audit total_score is the quality input metric; dashboard headline numbers are output metrics. Audit is leading indicator; dashboard is lagging.
• Integrates with the bulletin board — "AI Caught a Hallucination" feeds the stratum-four sample; the audit confirms catches were appropriate and remediations stuck.
• Discipline is small, leverage is large — ~200 hours per year at L3 stack scale; consistent quality across quarters, defensible PE/M&A diligence artifact, drift caught in week three rather than month six.