Test the decision, not only the screen
Each happy path is paired with a misleading-input, boundary, timeout, ambiguity, or stale-state case.
Independent release record / July 15, 2026
Evidence before
confidence.
01 / Method
How the work
was verified.
Compare at the source
Live facts are checked against authoritative public endpoints or filings. Calculations are recomputed independently from visible inputs.
Keep estimates in their lane
Reported facts, derived metrics, scenario assumptions, and hypotheses receive different labels instead of sharing one confidence claim.
Fail partially and honestly
When one provider is unavailable, useful verified context survives where possible and unavailable fields do not become invented numbers.
Make the release reproducible
Pure-model tests, route tests, strict types, a production build, and browser workflows form one release gate.
02 / Project matrix
What passed—and
what it cannot claim.
Verified with boundary
HeatSignal
- Checked
- Heat-index arithmetic, EPA color bands, ambiguous geocoding, partial AQI behavior
- Evidence boundary
- Weather and CAMS AQI are model estimates; they are not a local regulatory monitor.
Source-matched
Voltline
- Checked
- EIA state price, customers, sales, hourly grid values, ZIP formulas and utility records
- Evidence boundary
- ZIP use is modeled from state and housing evidence; it is not metered ZIP consumption or a tariff quote.
Source-matched
FieldLens
- Checked
- VIN check digit, NHTSA decode, recall count, campaign dates and partial-service recovery
- Evidence boundary
- Campaigns match year/make/model and still require VIN-specific confirmation from the manufacturer or NHTSA.
Verified with context
SceneCraft
- Checked
- Exact URL, metadata, indexability, DNS/RDAP evidence and partial-audit states
- Evidence boundary
- SEO guidance is diagnostic; a missing sitemap or robots file is not automatically a ranking failure.
Live-source guarded
PlateScout
- Checked
- Location resolution, OSM coordinates/tags, distance ranking, truncation and provider recovery
- Evidence boundary
- Cuisine tags are not menu proof. Availability, hours and dietary suitability must be confirmed with the venue.
Round-trip tested
PaperPatch
- Checked
- Coordinate conversion, PDF save/reopen, edit history, replacement guard and keyboard undo/redo
- Evidence boundary
- A visual cover-up is not secure redaction; a visual signature is not identity verification.
Formula-verified
Automation ROI
- Checked
- Normal, zero, negative, fractional, break-even, range and realized-value scenarios
- Evidence boundary
- Recovered time is capacity, not guaranteed cash. The realization factor and pilot evidence remain owner assumptions.
Fixture-consistent
CareSignal
- Checked
- All six synthetic scores, count validation, stale handoff protection and CRM-preview identity
- Evidence boundary
- Fixture consistency is not predictive accuracy, conversion lift or a fairness audit on real organizations.
Deterministic QA
RoutePilot
- Checked
- Routing combinations, phrase boundaries, negation, estimate ranges and invalid inputs
- Evidence boundary
- The cost range is a planning model, not a calibrated quote or a claim about customer value.
Abstention tested
AccountBrief
- Checked
- Known companies, domains, ambiguous places, people, products, creative works and source freshness
- Evidence boundary
- The system must abstain when identity evidence is weak; hypotheses remain separate from sourced company facts.
03 / Primary references
Traceable, current,
and close to the claim.
How to explain the engineering
“I treated every project as a decision system. I separated source facts from calculations and estimates, tested realistic and adversarial inputs, preserved useful partial results when a provider failed, and made every limitation visible beside the output. AI accelerated implementation and review; the product boundaries, source choices, acceptance criteria, and release decision stayed human-directed.”