H1
Allowlisted, confidence-gated AI briefs can stay closer to primary sources than unfettered AI news rewrites.
Experiment.
A living system, not a press release. Illumora tests whether gated AI journalism about AI — paired with literacy practice — can stay honest under public scrutiny.
The question
How good can fail-closed AI journalism get — when the subject is AI itself?
Peers prove that AI can fill a wire about AI. Illumora asks a harder question: can an allowlisted, confidence-gated desk — with a public methodology, corrections log, and provenance — earn enough trust to teach? The product stack (Studio, Learn, Atlas, Craft) is the other half of the experiment: read, then practice.
We do not accept payment or placement from labs we cover. Prefer the primary source when our wording and the source diverge.
Hypotheses
What we are testing.
H2
Public methodology + corrections + provenance improve trust relative to opaque AI news feeds.
H3
Pairing Pulse with Learn / Studio practice raises AI literacy better than news-only consumption — coach, not crutch.
Live metrics
Public numbers.
Metrics refresh on each page load. Editorial floors are code constants; Pulse and Learn counts are live from the public database. Gate-fail time series stay ops-side until we ship sanitized rollups.
Pulse published
01
65
Live count of published briefs on the public wall.
Corrections log
02
0
Explicit corrections and retractions — empty is still a signal.
Learn modules
03
4
Hung literacy exhibits in the Learn gallery.
Draft floor
04
78%
Auto-publish requires draft confidence ≥ 0.78 (classify ≥ 0.72).
Wire sample
46
Recent ranked briefs on the speed lane
Deep / Policy sample
19
Analytical spine — not a breaking race
Multi-source sample
1
Select leads synthesizing multiple allowlisted primaries
Floors in code: classify 72% · draft 78% · updated 2026-07-22T11:10Z
Update cadence
How often this page changes.
- Continuous
Pulse ingest, classify, draft, and gate decisions. Metrics above recompute on load. - Weekly
This Week package + Pulse+ digests. Experiment narrative reviewed when floors or lanes change. - On record
Corrections append to the public log. Provenance JSON for every published brief.
Where AI is used
Disclosure matrix.
Same matrix as Methodology— exported in the outreach PDF and JSON dump.
| Surface | AI role | Human role | Disclosure |
|---|---|---|---|
| Pulse draft | Drafts title, dek, body from allowlisted source text only | Ops Trust Desk on hold / retract; gates fail closed | On every brief + methodology |
| Ingest / classify | Routes items to Pulse, Atlas, Learn, or discard | Allowlist + authority tiers | Methodology |
| Social captions | Compresses Pulse/Learn/Studio into IG/X/LI copy | Excellence gates; caption claims must match body | Not on-post (yet) |
| Studio / Learn | Generates practice prompts and lesson assist | Technique registry + product UX | Product is model-assisted by design |
| Never AI | Inventing sources, quoting people we did not fetch, paid placement, inventing model versions for social | — | Hard prohibition |
Machine-readable
Provenance on every brief.
Each published Pulse story exposes /api/pulse/{slug}/provenance — sources, confidence, lane, synthesis mode, and an optional HMAC-SHA256 signature when PULSE_PROVENANCE_SECRET is set.
Outreach pack
PDF + JSON for partners.
One branded dossier and one dump: hypotheses, floors, AI-use matrix, corrections excerpt, and canonical URLs — for labs and foundations exploring the experiment.
Grant and research contacts are tracked internally for a later pass. For inquiries now: support@illumora.io.