What has been measured, what remains human judgment, and which launch gates are still open.
Reliability gate
100%Retrieval evaluation
75/75Questions passed
90%Required gate
14/7Consecutive daily passing runs
Aggregate workflow funnel · rolling 30 days
1195 application open events
107 cross workflow or guided example action events 9% relative to application open events; not a user conversion rate
976 completed Policy Desk questions 82% relative to application open events
1 Policy Desk cited source link opens 0% relative to completed questions; briefing and record page source clicks excluded
Operational metrics as of Oct 13, 2026, 7:55 AM EDT. Counts are events, not unique people or a sequenced cohort; one visit can produce several events.
Last production run: 2026-10-13 · scheduled_production_database_source_retrieval. The 75-case fixture runs daily inside the production runtime. The reliability gate requires at least 90% on the current run and seven consecutive daily passing receipts. Retrieval does not by itself prove that every generated sentence is correct.
Run date
Result
Mode
2026-10-13
75/75 · 100%
scheduled production database source retrieval
2026-10-12
75/75 · 100%
production database source retrieval
2026-10-11
74/75 · 99%
production database source retrieval
2026-10-10
75/75 · 100%
scheduled production database source retrieval
2026-10-09
74/75 · 99%
production database source retrieval
2026-10-08
75/75 · 100%
scheduled production database source retrieval
2026-10-07
74/75 · 99%
production database source retrieval
Briefing gate
14/14Automated structure checks
113/113Bullets with source links
0/14Briefing editions retained as reviewed
14 daysRequired clean run
The automated check tests length, completeness, duplicate bullets, quiet day behavior, obvious overclaim language, and whether each identified policy claim resolves to a source link. It cannot certify source fidelity or editorial judgment. The two week launch gate remains open until an authenticated reviewer logs each edition in the organizational audit record. No organizational review is recorded yet.
—Answers reviewed in this browserReading browser-local review state
20Fixed answer cases
5Human scored criteria per answer
The 75-question automated run tests retrieval. This smaller gate reviews final answer citation entailment, completeness, evidence lane discipline, and abstention quality. It remains open until the exported human review record is retained and published.
Scope: product wide aggregate. Browser local tool exports count Food Label drafts and historical exports from the retired Project Evaluator. They are distinct from server generated CSV exports reported on Data Health; neither measure represents unique people. Policy Desk: 976 completed questions · 0 follow ups · 0 helpful and 0 needs work ratings · $19.25 estimated model cost. As of Oct 13, 2026, 7:55 AM EDT.
Independent challenge sets
Evaluation benchmark registry
Benchmarks are frozen, reviewed comparison sets. Disagreement triggers investigation; it never silently imports a competitor record or establishes legal authority. Current page snapshot: dcp-88df31923b7a6a4b · current page build: 6accf78722ce669b506b70c5. A registry entry without a comparable run shows no benchmark numbers.
Policy Desk citation evaluation2026-09-13
75 maintained questions: 52 expected records and 23 expected abstentions · retrieval evaluation
InfraPrism
Cases tested
75
Cases retrieved
52
False negatives
0
False positives
0
Run date and time
Fixture version
sha256:dbd71fca4f89
Corpus snapshot
dcp-d73f0fc8d0f3f614
Build identifier
git:aff8a10
Snapshot changed: this retained run does not describe the current page snapshot. Re run before treating it as current.
Result: All 52 expected records were retrieved; all 23 unsupported or unsafe cases produced the expected abstention.
Known disagreement: Final-answer claims are not measured by this retrieval run; use the separate 20-answer faithfulness review.
A retrieval pass does not validate every generated sentence or prove corpus completeness.
UNC data-center policy benchmark2026-09-13
188 reviewed records · recall audit
University of North Carolina research snapshot
Cases tested
188
Cases retrieved
140
False negatives
48
False positives
Not applicable
Run date and time
Fixture version
sha256:f93e3e761ddd
Corpus snapshot
dcp-d73f0fc8d0f3f614
Build identifier
git:aff8a10
Snapshot changed: this retained run does not describe the current page snapshot. Re run before treating it as current.
Result: Recall-only identifier match. False positives are not defined because tracker records absent from the UNC set are not necessarily incorrect.
Known disagreement: Forty-eight unmatched UNC records remain audit leads that require primary-source adjudication before any coverage change.
Used to identify possible recall gaps; records are not automatically ingested or treated as authority.
DCMap statewide action baseline2026-09-13
22 reviewed statewide actions · recall audit
DCMap
Cases tested
22
Cases retrieved
22
False negatives
0
False positives
Not applicable
Run date and time
Fixture version
sha256:0463968a88e4
Corpus snapshot
dcp-d73f0fc8d0f3f614
Build identifier
git:aff8a10
Snapshot changed: this retained run does not describe the current page snapshot. Re run before treating it as current.
Result: All 22 records were retrieved; 20 of 22 action dates aligned within the fixture window. False positives are not defined for this recall-only set.
Known disagreement: Oregon HB 4084 and HB 3546 are record matches but their stored action dates fall outside the benchmark event window.
DCMap intentionally emphasizes restrictive actions and is not a balanced policy census.
Two statewide policies and ten selected-market local leads · recall audit
GatherGov
Cases tested
12
Cases retrieved
2
False negatives
10
False positives
Not applicable
Run date and time
Fixture version
sha256:adf0988032d3
Corpus snapshot
dcp-d73f0fc8d0f3f614
Build identifier
git:aff8a10
Snapshot changed: this retained run does not describe the current page snapshot. Re run before treating it as current.
Result: Both statewide policies were primary-linked; none of the ten bounded local discovery leads were retrieved. False positives are not defined for this recall-only set.
Known disagreement: Ten local leads remain missing because nationwide local surveillance is outside scope and the public article does not expose its underlying records.
The public article does not expose its 632-row dataset or primary records for most local claims; local matches remain discovery leads and are not silently ingested.
No comparable run published. No comparable retrieval run has been published for this parser-boundary fixture. The registry does not present test assertions as benchmark numbers.
Parser fixtures prove known examples, not complete state coverage or future page compatibility.
51 jurisdictions with anchors or evidenced empty review · recall audit
InfraPrism reviewed evidence
No comparable run published. No single confusion-matrix run is published because this fixture records jurisdiction-specific anchors and reviewed empty states.
A reviewed empty result is time-bounded and never proof that no relevant activity exists.
Retrieval stays at or above 90% for seven consecutive daily production runs.
Twenty fixed final answers pass human faithfulness review.
Fourteen consecutive briefing editions pass substantive human editorial review.
Adoption hypotheses
At least five beta testers return without repeated prompting.
Several testers would be disappointed to lose access.
At least three explicitly commit to the founding membership.
Do not add a paywall yet. Passing a quality gate does not prove recurring demand.
Next decision review: September 18, 2026, or sooner if all three quality gates close and the beta reaches the stated adoption thresholds. If neither happens, retain the no paywall decision and set a new dated review.