Phase 1 quality register

Phase 1 QA Register

Research date 2026-09-15. This register exists because a platform that claims to publish confidence must publish its own. Every weakness below is disclosed, not hidden.

Corpus

Verification status

Citation tiers. A (primary) 1037 (55%) · B (reputable secondary) 793 (42%) · C (aggregator/vendor) 66 (3%). No Tier-C source is the sole support for any claim.

Per-sector quality and re-run priority

Sectors 13–25 were researched after the session's 200-call web-search pool was exhausted. Their Tier-A verification is sound — in several cases higher than average, because agents fell back to primary sources — but source discovery was narrow, so trends that only search would have surfaced are likely missing. Priority = ranking weight × (1 − triangulation rate) × speed-of-change multiplier.

Aggregate effect of degraded discovery. Full-search sectors: 77% triangulated, evidence factor 0.70. Degraded sectors: 62% triangulated, evidence factor 0.65. The gap is real but modest — roughly 15 percentage points of triangulation. No cross-sector conclusion in the analysis documents rests on degraded sectors alone.

Data repairs applied (2026-09-15)

Full audit trail in analysis/normalisation-log.json. Nothing was silently corrected.

Known limitations

1. Six canonical trend-type terms have zero instances — design, education, funding, privacy, startups, venture_capital. Three are industry labels rather than kinds of change; the vocabulary needs pruning in Phase 2.
2. Four source types are entirely unpopulated — patent, search_trends, social, community. Zero social and search is defensible. Zero patent is not: it is the longest-lead detection signal, free and bulk-downloadable. This is the single largest gap in the detection surface.
3. Source registry reliability is graded 450 high / 152 medium / 1 low. That describes a filtered registry, not the information environment. Phase 2 must not misread it as evidence that sources are generally reliable.
4. strategic_importance is under-discriminating (mean 4.01, sd 0.82) — analysts used it least carefully. Rubric needs tightening.
5. companies[] and products[] are free text, matching entity records only ~16% of the time. Entity resolution is required before the graph features in the product spec can work.
6. Non-English and non-US coverage is uneven. Several sectors recorded it as an explicit gap; 41% of registered sources are US-only.
7. The Sept 16 2026 FOMC decision and the Google ad-tech remedies opinion both fall one day after the research date and are unresolved throughout.

Structural gaps the research itself established

These are findings, not failures — the absence is the information:

• No frontier AI lab discloses a gross margin (sector 01)
• No operator anywhere publishes robotaxi unit economics (sector 10)
• No independent audited DPI benchmark exists for venture capital (sector 11)
• No US system operator publishes metered data-centre-only load (sector 05)
• No agentic-commerce transaction volume is published by any participant (sector 12)
• No AI music licence has disclosed any terms (sector 24)
• No public audited CPM index exists for any advertising channel (sector 16)
• No methodologically transparent count of North American data-centre capacity exists — CBRE and JLL differ ~9x (sector 18)

More research from the library · Explore trend records