← Back to Daily Builds

Daily build note · August 14, 2026

AI Interaction Disclosure Evidence Desk

A deterministic evidence desk that compares structured voice, web-chat, and messaging disclosure observations with a human-approved policy, ranks missing or inconsistent evidence, manages...

Combine Or Alternate Between The Strongest Themes Above Build note published Public demo coming soon

What shipped

A complete standard-library Python 3.11+ CLI now runs directly from this build root without installation, credentials, environment variables, network access, or external services. The vertical slice is complete: create the fictional three-channel desk, validate and hash its closed inputs, generate a deterministic 33-file pending packet, bind a human-authored decision proposal to current source and finding hashes, regenerate, and independently validate any honest decision state or require evidence_accepted.

The sample has one fictional service workflow, voice/web-chat/messaging channels, eight touchpoints, twelve observations, four linear retests, 37 synthetic evidence files, 17 historical/current findings, and six current blockers. It visibly covers clean evidence, missing evidence, late timing, exact-copy mismatch, placement mismatch, a supplied accessibility-evidence gap, open blockers, and successful retests.

Architecture

  • core.py owns strict UTF-8/JSON/CSV parsing, exact nested contracts, safety scans, evidence path and closure checks, hashing, snapshots, and atomic directory primitives.
  • engine.py is a filesystem-independent exact-match rule engine with stable finding IDs, linear retest resolution, canonical queue order, source/finding binding, and the four decision states.
  • sample.py stages and validates the new fictional fixture before replacing an exact safe target.
  • render.py builds internal and client-safe views independently, produces all expected bytes deterministically, indexes every non-index artifact, and validates the output tree against independent recomputation.
  • validation.py coordinates immutable-input generation, staged validation, atomic output replacement, atomic decision recording, stale-decision rejection, and acceptance-gated validation.
  • tests/ provides 24 adversarial and end-to-end tests plus a sample-only human-decision proposal helper.

The design favors complete byte recomputation over trusting generated metadata. It deliberately keeps approval separate from factual resolution: a current approve can produce evidence_accepted, but the underlying findings remain open until a later conforming linked retest.

Suggested next steps

The next-day test is one paid **$2,500 USD upfront** bounded review for one real workflow and up to three channels. Qualify that the buyer already owns an approved policy/applicability decision and can provide sanitized structured observations without a connector. Measure whether intake is feasible, whether the exception/retest packet improves a 45-minute review, and whether the final receipt is useful to the buyer. Do not add a dashboard or integration until at least two paid engagements expose the same repeatable source format.