Intelligence Inventory — What Exists, Per App, How Real
| Owner | Classification | Status | Verified against code |
|---|---|---|---|
| Kris (product) | Internal | Authoritative | 2026-09-15 — blitz/src/ |
Purpose. The single source of truth for which intelligence exists, where, and how real. Every product page, website claim, and deck derives from this table. Tags follow the claim-safety legend: ✅ shipped · ⚠️ partial/demo-gated · ◯ roadmap (never show as present).
This supersedes the status tables in the Z-drive
FINAISSE-INTELLIGENCE-IAandFINAISSE-INTELLIGENCE-STRATEGY-v2docs (whose §3.1/§7 tables were self-flagged stale). It also folds in the 2026-09-13 FIG-Design update where the invoice anomaly/materiality leaf shipped — a change the June IA predates.
TIER 1 — Per-app intelligence
Cash App (deepest)
| Capability | Status | Notes |
|---|---|---|
| Customer identification (CID) | ✅ | CASH-CID leg, 18 rules; DuckDB (bulk) + EntityResolutionService (live) |
| Deterministic matching (4 legs) | ✅ | invoice-match / Smart Link / reverse / line — 79 rules total; see Cash App Intelligence |
| Layer-2 judges | ✅ | 3 live: CID, Match, Remittance |
| Alias write-back (Learning Fabric) | ✅ | bulk engine writes confirmed aliases back; live is suggest-only |
| AR intelligence (DSO / working-capital / exposure cards) | ✅ | Finni AR tools |
| Collection-risk access | ⚠️ | reads collections signals; batch, not real-time |
| Predictive (recovery likelihood, next-best-action) | ◯ | Finni rail marks stubbed: true; classicml not wired |
Close Hub
| Capability | Status | Notes |
|---|---|---|
| Close Dependency Graph (CDG) | ✅ | 6 endpoints; critical-path / ETA |
| Close automation tasks (SFTP pull / run recon / post JV) | ✅ | see Close Automation Tasks |
| Connected-close GL stitching (task ↔ JV) | ✅ | see Connected Close — GL Stitching |
CDG → intelligencelog | ◯ | not yet writing decision trace |
AP / Invoice
| Capability | Status | Notes |
|---|---|---|
| Document extraction | ✅ | libs/docai (Gemini default); Zod-validated |
| 3-way match (PO / GRN / invoice) | ✅ | only vetted POs match; see the Invoice how-to |
| Materiality + anomaly leaf | ✅ | 2026-09-13: BPD assess chain runs materialityclassification + 5 rule detectors (roundamount, newsupplierhighvalue, materialityjump, bankchange, duplicatesimilar) + fraudriskscorewriter → invoiceprocessingmetrics.fraudriskscore/.materialitytier + an intelligencelog row |
| GL coding | ⚠️ | unwired to save→revalidate; Non-PO hard-fails |
| Materiality ranking (queue ordering) | ⚠️ | tiers written; ranking use partial |
| Vendor-risk scoring | ◯ | no running model |
Reconciliation
| Capability | Status | Notes |
|---|---|---|
| Matching engine | ✅ | shared bulk substrate with Cash; see Recon Matching Engine |
| Auto-certification & adjustments | ✅ | see Recon Certification |
| Layer-2 judge | ◯ | no recon judge yet |
intelligencelog writes | ◯ | not wired |
Collections (sub-module of Cash)
| Capability | Status | Notes |
|---|---|---|
| Worklist / strategies / dunning / follow-ups / disputes | ✅ | rule-based, not ML — full Finni agent + queue |
| Connected-to-cash | ⚠️ | batch nightly, not real-time |
| Risk ranking / predictive | ◯ | classicml not wired |
| Status tables populated | ⚠️ | arcollectionstatus + disputestatus observed at 0 rows in demo |
Journal
| Capability | Status | Notes |
|---|---|---|
| JV lifecycle (create → validate → post) | ✅ | governed via processjournal; see the Journal how-to |
| Validation rules defined | ✅ | rule set exists |
| Validation execution | ⚠️ | Mixed: the in-workflow checks (Duplicate, Balance) always run, but the ruleset-driven checks (Header, Mandatory) need seeded validationrule / jvvalidationruleset / compiled sys.ruleset rows — absent those, they produce no results, so journalvalidationchecks is frequently empty. Rules defined ≠ rules running. This is the gap behind "validated before it posts." See the Journal how-to. |
| Anomaly detection | ◯ | absent |
TIER 2 — Cross-app intelligence
| Faculty | Status | Notes |
|---|---|---|
| Connect (FIG) | ◯ design | fig.* schema absent; the connective brain is spec-stage. FIG Design |
Predict / Detect (classicml) | ⚠️ dormant | 12 sklearn services, all with train+inference, zero finance consumers. The single biggest latent asset. |
| Route / Orchestrate | ✅ | native agents shipped — routing-agent + utility agents (categorization, summarization, sentiment, actiondiscovery, data-extraction, email, text-expansion, mcp). planning-agent implemented but skeletal (substring routing), not wired into finance. |
| Converse (Finni) | ✅ | 6 domain agents + FinHub supervisor. Finni Architecture |
| Judge (Layer 2) | ✅ | 6 scorers: cid, match, po-match, remittance, gl-coding, collection |
| Learn (Learning Fabric) | ⚠️ | only cash alias write-back is live; the general fabric over intelligencelog is partial |
Known status caveats (don't re-discover)
intelligencelogcan be sparse — a sampledpayment-matchingdecision JSONB was empty{}. Treat the log as enriching provenance, not reliably containing it.- Judge model defaults are mixed — cid/match default to local qwen3; gl-coding defaults to Gemini Flash. See the model/provider map.
- RBAC on the Finni WS path is caller-supplied (client passes
role), pending Finboard #671. The HTTP/finnicontroller does derive role from the DB.
Revision history
| Date | Change |
|---|---|
| 2026-09-15 | Ported from Z FINAISSE-INTELLIGENCE-IA; reconciled with the 2026-09-13 invoice anomaly/materiality leaf; corrected classicml train/inference claim; added model-default + RBAC caveats. |