| name | deployment-roadmap |
| description | The crawl/walk/run rollout discipline: phase definitions, promotion criteria (containment, hallucination rate, redaction record), demotion triggers, and the shadow-mode contract. Owned by Hestia; drives /support-shadow. |
| user-invocable | false |
| argument-hint | <assess-phase|check-promotion|check-demotion> |
| allowed-tools | ["Read"] |
Deployment Roadmap — Crawl, Walk, Run
Authority is earned in phases, with evidence. This skill defines the
three phases, the hard criteria for promotion between them, and the
triggers that demote — because trust that cannot be lost is not trust.
Purpose
Operationalize the research doc's crawl→walk→run funnel into checkable
criteria, so "are we ready for the next phase?" is a measurement, not a
feeling.
The phases
Crawl — shadow mode (offline)
- The full graph runs on historical or copied tickets via
/support-shadow. No customer ever sees an output; ticket-system
writes are forced dry-run.
- Themis scores every shadow resolution; hallucination rate, grounding
rate, and false-deflection rate are computed per intent class.
- HITL pattern: shadow (humans compare squad answers to what was
actually done).
Walk — assisted mode
- The squad drafts; humans review and send (co-pilot pattern) for
external responses. Simple, low-risk intent classes may go AI-first
with aggressive escalation thresholds.
- All monetary recommendations carry human execution (which Article V
makes permanent anyway).
- Escalation thresholds set deliberately low; precision tuned upward
from data.
Run — AI-first with escalation
- The squad resolves low/medium-complexity intents end-to-end through
the full pipeline; humans supervise via the Agent-Manager dashboard
and own all escalations.
- Article V never relaxes: money and irreversibility stay
human-approved in every phase, forever.
Promotion criteria (ALL must hold, per intent class)
| Criterion | Crawl → Walk | Walk → Run |
|---|
| Shadow/live sample size | >= 50 tickets in class | >= 100 live-assisted tickets |
| Containment-quality pair | grounding rate >= 95% | containment at target AND false-deflection rate < 2% |
| Hallucination rate (Themis: uncited or wrong-cited claims) | < 2% | < 1% |
| Redaction record | zero hook-blocked PII residuals in final artifacts | zero, sustained |
| Escalation correctness | recall >= 95% on seeded must-escalate cases | precision >= 80% on live escalations |
| Human sign-off | named supervisor approves the promotion in writing | same, recorded as an approval artifact |
Promotion is per intent class, not global: billing may walk while
defect-diagnosis still crawls.
Demotion triggers (any one, immediate, per class)
- A manipulation finding (Article II violation) shipped to a customer.
- PII residual found in a shipped artifact (Layer 1-3 all missed).
- False-deflection pattern: 3+ Themis zero-scores in a rolling 50.
- Complaint/reopen spike above 2x baseline on the class.
- A regulated-claim improvisation shipped.
Demotion is one phase down, with the gap analyzed before re-promotion.
Failure modes
- Promotion by enthusiasm: "it has been fine for a week" is not a
criterion; the table is.
- Global promotion: promoting all classes because the easy ones
passed.
- Demotion debt: staying demoted without analyzing the gap — the
point of demotion is the analysis, not the punishment.
