A decision you can inspect.

Cover

Tier-one support routing plan

Filing

The plan

Route a bounded subset of low-complexity support requests through an AI response workflow, with human approval before sending, to test whether response time improves without reducing answer quality.

Stakes: customer trust and operator workload. Reversibility: high during the bounded pilot. Decision horizon: 30 days.

Charges

What must be true

• CHG-01The workflow can identify eligible low-complexity requests with at least 95% precision.
• CHG-02Human reviewers can evaluate proposed responses without increasing total handling time.
• CHG-03Source citations remain attached from retrieval through final approval.

Discovery — Evidence

Provenance by Charge

• E-001 / CHG-01200 synthetic request records; classification run dated 2026-07-22; SHA-256 prefix 91a7c4e2. Supports a bounded test only.
• E-002 / CHG-02Internal timing protocol v0.1; no completed reviewer-time measurement. Evidence gap remains open.
• E-003 / CHG-03Ten synthetic traces with source IDs preserved; run dated 2026-07-23; SHA-256 prefix 0dd821af.

Prosecution

Six-Month Failure Autopsy

The plan failed because the eligibility classifier looked strong on clean synthetic records but degraded on ambiguous real requests. Reviewers then spent more time correcting context than they saved drafting responses. Source links remained present, yet reviewers treated their presence as proof of relevance.

Missing control: no real-request precision sample. Unverified assumption: review time will fall. Failure mode: apparent speed at classification masks total-work increase.

Defense

Strongest controlled case for proceeding

Do not authorize broad routing. Run two independent experiments: first, a shadow-mode eligibility test on 100 de-identified requests; second, a timed reviewer study on 30 synthetic responses. Require source relevance confirmation, not source presence alone.

Verdict

Split into experiments.

The evidence supports testing the components separately. It does not support deploying the combined workflow or claiming a response-time improvement.

Sentencing — Tripwires

Conditions that govern the Verdict

1. Stop the eligibility experiment if precision is below 95% after 50 real-request classifications.
2. Stop the reviewer experiment if median total handling time rises by more than 10%.
3. Refile the plan if any approved response lacks a reviewer-confirmed relevant source.
4. Return for a new Verdict on day 30; silence does not extend authorization.

Evidence trail

Role and time record

• 09:00Example Operations Team filed the plan.
• 09:14Server-side capsule extracted three Charges.
• 09:31Prosecution recorded the failure mode and evidence gaps.
• 09:46Defense proposed two bounded experiments.
• 10:02Verdict returned; four Tripwires sealed in Packet DTP-SAMPLE-001.