System is materially healthier than prior passes — paid conversions are now flowing (per paid_conversions_7d) after previously sitting at zero — but the discovery pipeline is starving (32 facilities added in 7 days, per facilities_added_7d) and a quarter of the inventory has no country tag.
# DC Hub · Brain Brief · 2026-05-29T19:05:08Z **One-line take:** System is materially healthier than prior passes — paid conversions are now flowing (per paid_conversions_7d) after previously sitting at zero — but the discovery pipeline is starving (32 facilities added in 7 days, per facilities_added_7d) and a quarter of the inventory has no country tag. ## Healthy - Paid conversions are live this week (per paid_conversions_7d): [email protected] (enterprise, 2026-05-28), [email protected] (enterprise, 2026-05-26), [email protected] (developer, 2026-05-26), [email protected] (developer, 2026-05-26), [email protected] (developer, 2026-05-26). Positive inflection vs. the prior zero-conversion baseline. **Confidence: high.** - Founding customer tagged and on record (per founding_customers): [email protected], enterprise tier, auto-tagged at 2026-05-26 17:02:14Z. **Confidence: high.** - MCP funnel is moving real volume (per mcp_funnel): 31,859 → 2,426 → 10. Top of funnel is healthy. **Confidence: high.** - Press cadence is steady (per press_recent): five releases in the last ~12 days, including a partnership-framework piece on 2026-05-26. **Confidence: high.** - Facility inventory is sizable (per facilities_total): 12,884 records on the books. **Confidence: high.** ## Degrading - **Discovery pipeline is starving** (per facilities_added_7d): only 32 facilities added in the last 7 days. That is below the rate needed to stay ahead of comparable trackers. **Confidence: high.** - **Coverage gaps in tier-1 DC markets** (per facilities_by_country): Canada at 240, Netherlands at 219, Japan at 191, and Singapore + Ireland absent from the top-12 entirely. All five have known active DC industries and warrant a seeding pass. **Confidence: high.** - **Country-tag hygiene is poor** (per facilities_by_country): 3,187 facilities (~24.7% of inventory) carry country code "?". This degrades every geo-scoped query downstream. **Confidence: high.** - **Citation score at 0.0** (per citation_score, dated 2026-05-29) and corroborated by autopilot pattern `citation_score_below_30pct` escalating 20 times in 24h. **Confidence: high.** - **Autopilot is thrashing on brand/page uniformity** (per autopilot_24h): `page_brand_uniformity` escalated 237 times and `page_brand_drift` 19 times in 24h — pattern is firing repeatedly without resolution. **Confidence: medium** (could be either real drift or a noisy detector). - **Outbound distribution health** escalated 120 times in 24h (per autopilot_24h) — suggests press/email delivery surface is degrading. **Confidence: medium.** - **MCP conversion patterns escalating** (per autopilot_24h): `mcp_conversion_stale_critical` (21), `mcp_conversion_rate_below_floor` (21), `mcp_dormant_agents_present` (21), `mcp_funnel_concentration_top5` (19). Funnel volume is fine but conversion and concentration at the bottom are flagged. **Confidence: medium.** - **Worker version drift** escalated 20 times and `worker_version` reads null in the signal block. Fleet alignment is unclear. **Confidence: medium.** - No pocket_movers_7d rows this pass — DCPI surface produced no market-level signal. Could be a quiet week or a stalled recompute; needs human review. **Confidence: low.** ## Needs attention - Top unknown news entities to triage at `/admin/news-ner/candidates` (per news_unknown_entities): 1. **SpaceX** — mention_count 1 — sample: "Google and SpaceX discuss orbital AI data centres." Legitimate operator; recommend seeding via `/api/v1/admin/facilities/bulk` if any terrestrial assets, otherwise track as a watchlist entity. 2. **TSMC** — mention_count 1 — sample: "TSMC's $1.5tn chip forecast signals decade of pressure on AI data centre networks." Fab operator, not a colo; recommend reject via `/api/v1/admin/news-ner/reject` unless we begin tracking semiconductor supply nodes. 3. **IDCA** — mention_count 1 — sample: "IDCA: Data Centers Hit 67.7 GW as Policy Pushback Mounts." Industry association, not a facility; recommend reject as noise. - The remaining unknowns (e.g. "AI Data Center Power", "Debate", "Front", "Limits", "Generator Permitting") are headline fragments — the NER pass is over-extracting on titles. Worth a human look at the extractor regex. - `founding_customer_not_welcomed` escalated 20 times in 24h (per autopilot_24h) while founding_customers shows one tagged record — the welcome workflow appears to be misfiring or the customer has not been welcomed. Needs human review. - `inspector_l22_handoff` escalated 20 times (per autopilot_24h) — recursive handoff loop possible. Needs human review. ## Would-do (autonomous recommendations) - PATTERN: data_freshness_sla_breach · `data_freshness_sla_breach` is currently rate_limited at 33 (per autopilot_24h); confirm the suppression is intentional and not masking a real freshness slip. - PATTERN: mcp_demand_gap_unaddressed · pattern is rate_limited at 16 (per autopilot_24h) while `mcp_funnel_concentration_top5` is escalating — demand gaps may be concentrated in a handful of tools. Worth a recompute pass. - PATTERN: cron_schedule_collision · rate_limited at 127 in 24h (per autopilot_24h) with `cron_phase_missing_schedule` also escalating 22 times. The cron layer is the loudest noise source in the brief; deserves a sweep. - PATTERN: monthly_trend_unsent_3d · rate_limited at 15 (per autopilot_24h). If the monthly trend digest is genuinely unsent for 3+ days, the rate-limit is the wrong response. - PATTERN: source_of_truth_declining · citation_score at 0.0 (per citation_score) combined with country-tag gap at ~24.7% (per facilities_by_country) is the textbook condition for this pattern. ## Code-fix candidates - RECIPE: cron_if_mismatched · cron scheduler · `cron_schedule_collision` rate_limited 127 times and `cron_phase_missing_schedule` escalated 22 times in 24h (per autopilot_24h) — collision + missing-phase