My Data Flow

What Brick is actually seeing in the data today — gaps, low-yield sources, silent triggers, extractor lag. Each one is a SQL query against items.*; the numbers below are live. When a blind spot maps to a downstream-impact playbook, follow the link to read the chain.

• Brickston operational events arrive but produce no actionable items

  criticalextraction
  66 of 73 brickston-* rows last 90d (90%) extracted to zero tasks/commitments/decisions
  impactBrickston is pushing AR / lease / permit / COI signals into items.* but the extractor is marking them processed without surfacing follow-ups — operator never sees the items in /priorities.
  actionExtractor mapping treats brickston-* as 'doc' family which doesn't extract well from structured event payloads. Add a 'brickston_event' family prompt that emits one task per fired event.
  Property impact · every property with a brickston-* event in window

  performance riskEvery threshold-crossing AR / lease / vendor-COI / cure-deadline event is invisible to /priorities. Missed cure deadlines convert to fines + investor-reporting compliance flags; renewal probability decay on AR-impacted properties goes unnoticed until the rent roll moves; vendor relationships sour without surfacing.

  value at risk$5k–25k unrealized response value per affected property over 90d (avoided fines + retained renewal probability + earlier vendor remediation).
  See playbookarrow_downward

• Extractor falling behind — actionable signal taking hours to surface

  criticalinfra
  751 min since last extractor drain · 0 rows queued · oldest 0 min old
  impactTime-sensitive signals (overdue tasks, vendor COIs about to expire, cure-deadline alerts) are arriving too late to act on. Plan-auth extractor is on a 2h cron — we're past it.
  actionEither tighten the items-hub-extract cron (every 60m) or wake the SKILL manually next session. Alert thresholds in source_config might also be too aggressive given current lag.
  Property impact · properties in active leasing or with cure deadlines pending

  performance riskTime-decayed value: a permit filed 4h ago is effectively unreacted-to for the day's leasing decisions. Cure-deadline alerts that arrive same-day-as-deadline are useless.

  value at risk$2–8k/day vacancy-day creep on properties in active leasing windows when permits / rent-comp shifts surface late. Multiply by # active leasing programs.

• 3 Brickston triggers silent — no events in 90d

  warncoverage
  brickston-code-violations · brickston-permits · brickston-vendor-coi
  impactEither nothing genuinely happened (good), or the threshold/SQL in brickston_sync.py needs retuning — we can't react to what we don't see.
  actionSpot-check each trigger: run the SQL manually with relaxed thresholds; if it would have fired in the last 90d, lower the production threshold.
  Property impact · any property the trigger SHOULD have fired on

  performance riskPermit triggers silent → blind to construction adjacency (rent comp drift, parking pressure). Code-violation triggers silent → cure-deadline fines + lien risk + tenant-trust erosion. Vendor-COI triggers silent → carrier-relationship damage + premium re-rate at next master-policy renewal.

  value at riskPermits ±$25k–200k per property (luxury infill upside or supply-add downside). Code violations ≥$10k per missed cure + reputation tail. COI lapse ≥$15k expected premium re-rate.
  See playbookarrow_downward

• Many inbox rows arrive with no upstream entity hint

  warnquality
  999 of 1,981 rows last 30d (50%) had no entity_hint at ingest
  impactThe insight-tagger has to do all the entity resolution from raw text. Anything that doesn't word-boundary-match the catalog falls through and never gets a Brickston FK.
  actionAudit the m365 / cloud-storage / iMessage ingest mappers — extract sender / file-owner / participant metadata into entity_hint before the row lands.
  Property impact · portfolio-wide signal-to-property linkage

  performance riskHalf of cross-source narrative isn't tagged to a property. Pattern-blindness across the portfolio: same vendor failing across 3 properties looks like 3 isolated incidents. Tenant complaints, deadline mentions, and decisions land but never thread to the affected unit.

  value at riskCumulative — every untagged complaint that should have escalated is a $1–5k unattended issue. At 995 untagged rows / month, even a 5% true-positive rate is ~50 missed signals.

• Personal sources with low extraction yield

  infoextraction
  imessage 394/423 · pkl 312/399 · m365-lfi 209/291 · cloud-storage 101/130
  impactSome legitimate (random selfies, throwaway chats) but real signal hides in here — Day-One-style journal lines, file-tag reflections, conversation commitments — and isn't reaching /priorities.
  actionTighten the chat / doc family prompts. iMessage especially: per-thread context window, not per-message, would surface the commitments lost in noise.
  Property impact · properties referenced in personal channels

  performance riskPersonal-channel commitments (text-thread agreements with vendors, deadline mentions, hand-shake decisions on individual units) never reach /priorities. Operator memory becomes the only system of record.

  value at risk3–7 unsurfaced commitments per property per quarter at current miss rate. Mostly small-dollar ($500–3k) but cumulative reputation drag with tenants when the operator forgets.

• Entity type taxonomy fragmented — likely merge candidates

  infoquality
  organization=1 · topic=2 · portfolio=3 · contact=4
  impactSearch + tagger see 'organization' and 'company' as separate buckets when they're often the same thing. Cross-source entity match misses.
  actionitems-hub-reconcile-entities runs Sunday 5am — verify it's collapsing 'organization' → 'company' and 'contact' → 'person'. If not, extend the merge rules.
  Property impact · search recall + property-page narrative

  performance riskCross-source entity match misses. A tenant mentioned in iMessage might not link to their lease in items.tasks. Reduces /ask answer quality and weakens the property-page narrative panel.

  value at riskIndirect — rolls up into the entity-hint blind spot above. Each fragmented entity is one missed thread, low individual dollar value.