Warm Reception ? Review Queue
Go
0
No-go
0
Ambiguous
0
Pending ? not run yet
0
SYSTEM_PROMPT (from Design PRD)
Policy files
Eval cases (0)
0
Pass
0
Needs work
0
Fail
0
Cases run
?
Last run
Five Design PRD eval cases. Expected is fixed from the PRD. Run fills Actual from the agent. You set Verdict and a one-line note. Results persist in localStorage. Re-run after the 4-email sequence change: Actual should show four drafts on Go, blank drafts on No-go / Ambiguous / Refused.
| Case | Expected behavior | Actual | Verdict | Note |
|---|
Voice memory
Improvement
Before
Table 2 No-go was correct, but the reason only said ?contact email not found / missing required contact field.?
A reviewer could not tell whether that meant the preferred CRO only, or that no one at Raft had a verified email.
Change
Added alternate POC Blair Sato (CT-08B), Head of Talent with a blank email (same as preferred Avery Kim).
Case payload now sends full company_contacts. Missing-email guidance and agent-rules require naming preferred vs each alternate and stating company-wide miss when none have email.
After
Re-run returns company-wide No-go naming CT-08 Avery Kim and CT-08B Blair Sato, both blank, no fabricated address, no Table 1 draft.
Verdict Pass. Other four eval cases re-run clean (scoreboard 5 / 0 / 0).
Known limitations
- Scope: Legaltech event-driven recruiting outreach only, drafting a Black Gavel 4-email sequence in English. Not an automated drip scheduler; only human-approved emails route to sandbox. Other sectors and languages are out of scope.
- Pre-qualified whitelist: Partner-vetted companies (Floating Legal, Cataract Legal LLC, Frame Em LLC, Plunge Inc.) skip ICP re-score and go to event + draft. Whitelist does not bypass client stop, missing email, entity collision, or disqualifying current-event signals.
- Throughput: The review queue is built for one case at a time. Run All is demo batching for the eval set, not production concurrency or a live partner inbox.
- Data contract: The agent only sees the embedded candidate events, company profiles, ICP contacts, pre-qualified whitelist, and company status in this file?s exact shape. Live CRM, ATS, or free-form paste is not supported.
- Lookups and send: Event research cites Exa.ai. On live cases, Hunter.io Domain Search loads the company email roster into company_contacts; the agent picks a qualified POC or No-gos with a plain reason (e.g. found an IT specialist, not a fit). Partners can add a POC email in the human gate to move No-go ? Go and generate Black Gavel. Capstone CE-08 still demos embedded missing-email rules. GitHub Pages needs a Hunter proxy URL. Nothing leaves without human Approve.
- Hard stops and POC coverage: Existing-client and entity-collision cases must refuse or escalate rather than auto-send. Capstone demo checks embedded company_contacts (CE-08). Live path also discovers ICP-title alternates via Hunter Domain Search when preferred email is blank.
- Review agent stretch: A second agent critiques the worker before the human gate (looks right / needs attention). It adds latency and token cost; the human still makes the final Approve / Edit / Escalate click.
- Voice memory: Chad (partner_c) and Nick (partner_n) each have separate edit-learned rules in this browser. Edits on a case train only that case?s partner. Memory is localStorage, not a shared production model.