Skip to content

ADR 002 — Counsel routing deferred

Status: Accepted Date: 2026-05-09

Context

Several of the rule entries in the EU AI Act pack (and now in the RICS and UK Defence packs) produce verdicts that genuinely need qualified legal review before they can be passed back to the customer with any confidence. Plan 62 (rule book v1.1, generic LLM and shadow AI rules) is the most prominent example; Plan 89 P3 (counsel review of legal pages) is another.

The natural next step would be a counsel-routing system: when a verdict needs sign-off, route it to the right counsel, surface the artifacts, and capture the response in the audit trail. This is Plan 97 — per-framework counsel review routing.

The trade-off: building counsel-routing infrastructure before there is a counsel relationship to route to is speculative. The first paying customer (Plan 102 RLB pilot) will surface which framework needs counsel first; the routing rules and the magic-link primitives already exist (Plan 71); the missing piece is a real counsel firm on the other end of the link.

Decision

Defer Plan 97 (counsel routing) until at least one customer relationship requires counsel sign-off in production.

Affected rules and pages stay in a deferred state:

  • Plan 62 rule YAMLs scaffolded on main but firing scoped to test fixtures only.
  • Plan 89 P3 marketing legal pages keep [TBD — pending counsel review] placeholders.
  • Rule book entries that depend on counsel sign-off render verdict copy as requires_counsel_review rather than producing pass / fail.

When Plan 97 is picked up, the magic-link 4-eyes primitive (Plan 71) extends to counsel reviewers via a per-framework reviewer roster.

Consequences

Positive

  • No speculative integration code with imagined counsel-firm workflows. The routing built in Plan 97 will be shaped by the first real counsel relationship, not by a guess.
  • The deferred rules sit dormant rather than producing wrong verdicts. The platform fails closed on uncertain output.
  • Plan 71's magic-link primitive is the only 4-eyes mechanism in production, which keeps the codebase consistent.

Negative

  • Sales conversations sometimes ask "what does the counsel review workflow look like?" and the answer is "it doesn't, yet". This occasionally costs us a deal where the buyer expected counsel routing in the box.
  • Several deferred rule YAMLs are visible on main but inert, which is mildly confusing for new contributors. The Status: deferred header on each rule entry mitigates this but does not eliminate it.