← Back to Daily Builds

Daily build note · August 4, 2026

AI Front Desk Pilot Receipt Desk

A local-first post-launch evidence desk that combines an AI receptionist policy pack, vendor-neutral call events, sanitized interaction summaries, and business outcomes to produce an...

Local Business Automations Build note published Public demo coming soon

What shipped

A runnable Python 3.11+ standard-library CLI with three commands:

``bash python3 -m front_desk_pilot_receipts sample-pilot --workspace /tmp/pilot python3 -m front_desk_pilot_receipts generate-receipts --workspace /tmp/pilot python3 -m front_desk_pilot_receipts validate-pilot --workspace /tmp/pilot ``

The sample creates exactly 20 deterministic calls, 20 linked structured summaries, a representative Launch Guard policy, and appointment/estimate/job/revenue outcomes. The evidence set includes a booked job, blocked price promise, missed emergency escalation, correctly honored opt-out, duplicate follow-up, unresolved high-intent lead, clean FAQ resolution, unsupported signal, missing disclosure evidence, and internal-only canary/private-path values.

Generation writes the weekly brief, ranked review queue, scorecard, exception report, outcome ledger, evidence-backed policy proposals, 20 internal JSON receipts, 20 separately rendered client-safe Markdown receipts, and a machine-readable index. All eight required rule codes are implemented; the honored sample opt-out correctly produces no opt-out failure, while the unit test proves OPTOUT_NOT_HONORED fires on corrupted evidence.

Architecture

  • Standard library only; no installation, environment variables, credentials, or network.
  • Strict canonical CSV/JSON contract with closed vocabularies and timezone enforcement.
  • Workspace-containment checks resolve paths before accepting them.
  • Stable SHA-256-derived receipt/finding IDs and fixed fixture metadata for byte-identical output.
  • Structured rule evaluation only; summary prose is never classified.
  • Decimal outcome amounts and separate confirmed/plausible/unattributed buckets.
  • One aggregate object feeds the brief, scorecard, and index; validation recalculates source truth.
  • Internal and client-safe receipts use distinct data views and renderers.
  • Policy SHA-256 detects mutation; policy proposals never write to policy input.
  • Validation checks artifacts, relative references, identities, totals, ordering, resolution evidence, and leakage.

The parent Launch Guard decision vocabulary was reused without copying its Node implementation. The secondary trace receipt build informed stable IDs, separate safe rendering, and defensive validation only; this build has its own phone-pilot domain model.

Verification

bash verify.sh runs 11 unit/end-to-end test groups and a fresh CLI smoke flow. It proves the positive workflow and failures for malformed JSON/CSV, duplicate calls, unknown dispositions, invalid policy decisions, offset-free timestamps, missing/escaping summaries, unlinked outcomes, deleted artifacts, broken index references, aggregate corruption, unsupported resolution claims, and canary/caller-token/private-path injection. It also proves policy immutability and byte-identical regeneration.

Suggested next steps

  1. Run a paid pilot with one real sanitized export and document mapping gaps.
  2. Add one explicit offline adapter only for that observed export format.
  3. Add reviewer-supplied resolution records and signatures as a new versioned input.
  4. Add branding/templates only after owners validate the packet's decision usefulness.
  5. Keep live writes and policy mutation out until evidence review has repeated demand and stronger authorization controls.