Skip to content

PLAN.md

Latest commit

 

History

History
244 lines (172 loc) · 18.8 KB

PLAN.md

File metadata and controls

244 lines (172 loc) · 18.8 KB
Raw

Kuber Map for FIBO — Project Plan (v2)

A curated, learner-first knowledge map for the Financial Industry Business Ontology (FIBO) that doubles as audit-ready grounding for financial AI agents — with a provenance-tagged cross-domain bridge layer contributed back to EDM Council.

Status: Shipped & publicly launched — Phases 0–3 complete (5 use cases; value proven, +45.3pt); Phase 4 in progress (map published; EDM Council contribution packaged, engagement is the user's call). · Owner: Sanjeev Azad · Last updated: 2026-07-05

Execution status lives in BACKLOG.md (what's done / what's next). Snapshot: all 10 FIBO domains + Commons extracted (3,104 concepts, 19 cross-domain bridges); 5 curated use cases with context packs + an MCP endpoint; value proven (grounded vs ungrounded: +45.3pt accuracy, 97.0% auditable, 0% hallucination; 263 Q, gpt-4o-mini, corroborated on gpt-4o); desktop map + mobile PWA; public FOSS launch (GitHub Pages, 9-page wiki, CI, community scaffolding). Next: EDM Council engagement (user's call) + more use cases. Fresh checkout: make setup && make fibo && make commons && make all.

v2 changelog. This revision followed a three-lens adversarial audit (readiness, ROI, uniqueness). Major changes: (1) re-sequenced roadmap — agent-grounding is now Phase 1, not Phase 3; (2) corrected two overstated claims — FIBO cross-domain linking and OKF typed-frontmatter (§3, §6); (3) added an OWL-restriction extraction spike and level-of-detail rendering as hard Phase-1 requirements; (4) reframed the value prop and novelty around pedagogical curation + intra-FIBO bridges, treating OKF and agent-grounding as enablers, not the thesis. See §12 for the audit trail.

1. The sharpest framing (what we're actually building)

We turn FIBO into audit-ready grounding for financial AI agents — so they stop hallucinating on financial semantics, and you can prove to a regulator exactly which concepts justified every decision.

the interactive map is the showcase and the curation surface, not the product. The product is a curated FIBO subgraph + cross-domain bridges, exportable as per-use-case grounding context for financial AI agents, with provenance/citation for audit.

Value ranking (drives the roadmap):

  1. Agent-grounding (~70%) — the only pillar touching a pain with a budget: financial AI agents hallucinate on domain semantics and can't show a regulator why they answered. → Phase 1.
  2. Cross-domain bridges (~25%) — near-zero standalone value, but the moat: they make multi-domain use cases coherent and are hard to copy (curation, not code). → the connective tissue that makes pillar 1 work.
  3. The map (~5%) — a fork of Bodhi's Cytoscape UI; its jobs are (a) sales/showcase and (b) the human review surface for curating bridges. Not a business on its own.

The old v1 ordering (map → bridges → agents) was inverted relative to value and has been corrected.

2. Background (the three ingredients)

2.1 Bodhi Map — the UI we reuse

Hand-curated Cytoscape graph (128 concepts, 12 clusters, 6 typed edges). MIT-licensed, owned by us (© 2026 Sanjeev Azad). Fork it wholesale for the rendering/curation surface.

  • Corrected structure (verified against the repo): build tool is scripts/okf.js (JavaScript), app entry is app.html. There is no js/graph.js and no Python build_data.py in Bodhi. → We extend scripts/okf.js for the OKF→data.js build and keep the Python work confined to FIBO extraction (§5). Runtime is buildless; regenerating data.js is a build step.

2.2 Google OKF v0.1 — an export format, not the thesis

Markdown+YAML bundle spec (Google Cloud, June 2026). Required frontmatter type; recommended title/description/resource/tags/timestamp; reserved index.md/log.md.

  • Correction (load-bearing): OKF types relationships by prose around body markdown links, not by frontmatter. A relations: frontmatter block is spec-legal (unknown keys tolerated) but inert to any generic OKF consumer — only our own build script reads it. So the "agents consume typed OKF slices" claim is false for generic agents; they see untyped links like plain markdown.
  • Consequence: treat OKF as one export format among several (also emit JSON / Cypher / an MCP endpoint). If OKF stalls, the curated FIBO+bridges graph still stands. Don't bet the value prop on OKF.

2.3 FIBO — the subject matter

OWL ontology, RDF/XML (also TTL/JSON-LD via the ontology-publisher). ~2,446 classes in the current Production release. 10 domains: FND, BE, FBC, SEC, DER, IND, LOAN, BP, CAE, MD. MIT-licensed — redistribution/derivatives OK with attribution.

  • Maturity is annotated via fibo-fnd-utl-av:hasMaturityLevel (Release / Provisional / Informative) at the ontology-file level, not per class — must be propagated down to classes. Several domains (BP, CAE, MD, parts of DER/IND/SEC) are Provisional and churn quarterly.
  • Relationships are not flat triples: object-property relations (secured-by, hasUnderlier, party-to) live in owl:Restriction blank-node axioms on rdfs:subClassOf. The extractor must walk restrictions, not just triples (§5, spike).

3. Corrected theses (what survives scrutiny)

3.1 Cross-domain linking — corrected claim

v1 said FIBO is "deliberately thin on links across domains." That's overstated. FIBO ships FND/Relations/Relations.rdf, and FND is owl:imported by BE/FBC/SEC/IND — FND-mediated cross-domain linking is FIBO's backbone.

The real, defensible gap (this is the contribution):

  • Lateral, peer-to-peer bridges between business domains that don't route through FND (e.g. LOAN↔SEC for securitization, LOAN↔BE for reporting entities).
  • Learner-facing surfacing of cross-domain relations that exist in the axioms but are buried in restriction blank nodes and invisible in the official Viewer.

We must measure actual cross-domain imports first, then scope bridges to genuine lateral gaps — or the contribution won't survive EDM Council review.

3.2 Agent grounding — corrected claim

Grounding is real and valuable, but not novel (OG-RAG, Wipro, FinKario, Springer alignment work already do FIBO-grounded RAG). Our defensible delta is narrow and practical: per-use-case curated context packs with cross-domain bridges + provenance citations for audit — better than naive "dump FIBO in a vector DB" because vector similarity misses multi-hop cross-domain paths, which is exactly what curated bridges provide.

4. Uniqueness — where the moat actually is

No single pillar is novel: interactive FIBO viz is commoditized (official Viewer, GraphDB/Graphwise, WebVOWL); FIBO-grounds-agents is a crowded field; "OKF bundle → Cytoscape graph" is already an off-the-shelf tool (okf-skills). The defensible angle is the intersection, led by the two least-fashionable parts:

Pedagogical curation of FIBO (a teachable map with a hand-picked "core" default — nobody does this; every competitor renders FIBO exhaustively for experts) + a provenance-tagged intra-FIBO cross-domain bridge layer offered back to EDM Council.

OKF and agent-grounding are on-trend enablers, not the novelty claim. Lead with curation + bridges.

⚠️ Strategic: the EDM Council OKG (Open Knowledge Graph) Innovation Lab is both the best distribution/legitimacy path and the party most likely to build the cross-domain layer first. Engage them early (FIBO Community of Practice / EDMConnect) rather than building in a vacuum.

5. Architecture

bodhi-fibo/
├── fibo-source/            # git submodule → edmcouncil/fibo, PINNED to a Production tag
├── etl/                    # Python (rdflib / owlready2) — FIBO extraction ONLY
│   ├── extract.py          #   OWL → intermediate JSON. MUST walk owl:Restriction blank nodes,
│   │                       #   not just triples. Propagate per-ontology hasMaturityLevel → classes.
│   ├── freq.py             #   object-property frequency analysis → which edges to promote
│   └── validate.py         #   link integrity, orphans, provenance, IRI→path round-trip
├── knowledge/              # GENERATED OKF bundle (interchange + one export format)
│   ├── index.md
│   ├── FND/ LOAN/ FBC/ BE/ … (added one domain at a time, §7)
│   ├── bridges/            # CURATED intra-FIBO cross-domain layer (§8) — the moat
│   └── use-cases/          # AGENT grounding packs (§6) — the product
├── scripts/okf.js          # EXTENDED from Bodhi: OKF bundle → js/data.js (typed edges via our convention)
├── export/                 # per-use-case context packs: OKF slice + JSON + MCP endpoint (§6)
├── okf.config.js           # controlled vocab: edge types, cluster colors, use-case lenses
├── js/data.js              # GENERATED
├── app.html + js/          # forked Bodhi UI (Cytoscape + fcose) + LOD/lazy-load (§9)
└── fonts/ vendor/          # vendored locally (offline)

Pipeline: fibo-source (OWL) → extract.py (restriction-aware) → intermediate JSON → OKF bundle (+ hand-authored bridges/, use-cases/) → scripts/okf.js → js/data.js → Cytoscape; and → export/ context packs for agents.

Two toolchains, cleanly split: Python does FIBO OWL extraction (where the hard axiom-walking lives); JS (scripts/okf.js, extended from Bodhi) does bundle→data.js. Don't reinvent the JS build in Python.

OKF file model (per FIBO class)

---
type: FIBO Class
title: Mortgage Loan                          # rdfs:label
description: A loan secured by real property.  # skos:definition, trimmed (many classes lack one — see coverage risk)
resource: https://spec.edmcouncil.org/fibo/ontology/LOAN/.../MortgageLoan
tags: [LOAN, Release]                           # domain + PROPAGATED maturity
core: true                                      # in curated default view?
relations:                                      # OUR convention (inert to generic OKF consumers — see §2.2)
  - {type: is-a,        target: /LOAN/Loan.md,             provenance: fibo}
  - {type: secured-by,  target: /FND/.../RealProperty.md,  provenance: fibo}
  - {type: reported-in, target: /BE/.../RegulatoryReport.md, provenance: curated}   # a cross-domain bridge
---
Definition + learner notes. Related concepts ALSO linked inline as body markdown so
generic OKF consumers still get the (untyped) graph: see [Loan](/LOAN/Loan.md).

provenance: fibo|curated is essential: it lets us style bridge edges distinctly, and export only curated edges as the EDM Council contribution.

6. Agent-grounding layer — the product (Phase 1)

Each use case pins to the FIBO subgraph that grounds it and exports a portable context pack.

---
type: Agentic Use Case
title: Loan Origination & Underwriting Agent
description: Assesses a mortgage application against borrower, collateral, and regulatory concepts.
tags: [use-case, lending, LOAN, FND, FBC, BE]
grounds_in:                                     # the curated subgraph (closure exported as a context pack)
  - /LOAN/.../MortgageLoan.md
  - /FND/.../RealProperty.md
  - /FBC/.../Borrower.md
  - /BE/.../RegulatoryReport.md
agent:
  workflow: [ingest-application, resolve-entities, value-collateral, check-eligibility, decision]
  guardrails: [FIBO-provenance-citation, human-in-loop-on-decline, audit-trail]
---

Exports (export/): the grounds_in closure as (a) an OKF slice, (b) plain JSON for RAG, (c) an MCP tool / retrieval endpoint an agent calls. Each answer carries the FIBO concept IRIs that grounded it → the audit trail.

Candidate use cases: Loan Origination (Phase 1), KYC/beneficial-ownership (BE+FND), regulatory reporting (BE+FBC+MD), derivatives risk (DER+SEC+IND+MD), corporate actions (CAE+SEC), settlement (BP+SEC+FBC).

7. Domain selector & incremental onboarding

UI control to pick one domain / a set / "All". Repeatable onboarding recipe per domain: scope extract.py → write knowledge/<DOMAIN>/ → author index.md + tag core: → add cluster color → author lateral bridges to loaded domains → rebuild.

Order (value + dependency): FND → LOAN → FBC → BE → SEC → DER → IND → MD → BP → CAE. FND first (everyone imports it); LOAN first "real" domain (concrete, high FinTech value, the Phase-1 use case).

8. Cross-domain bridge layer — the moat / the contribution

Curated bridges/ files asserting lateral cross-domain relations FIBO doesn't draw natively, each with a rationale: + citation.

---
type: Cross-Domain Relation
title: Mortgage Loan → Mortgage-Backed Security
description: Pooled mortgage loans serve as the underlier of an MBS.
tags: [bridge, LOAN, SEC]
relations:
  - {type: underlies, target: /SEC/.../MortgageBackedSecurity.md, provenance: curated}
source_class: /LOAN/.../MortgageLoan.md
rationale: Documented justification + citation, for EDM Council review.
---

Contribution caveat: EDM Council models relations as formal owl:Restriction axioms under a DCO-signed PR + working-group governance process. Informal rationale: strings won't merge as-is; budget to re-model accepted bridges in their axiom style. Engage the OKG Lab first (§4).

9. Rendering & scale (hard Phase-1 requirement, not deferred)

~2,446 classes + subClassOf + promoted property edges ≈ 5,000–10,000 elements — fcose chokes at that size. Level-of-detail + lazy per-domain loading is required in Phase 1, not a "may need later." Curated core view is the default landing (readable); "Show full ontology" loads per-domain on demand. Maturity/level filter reuses Bodhi's level control.

10. Roadmap (re-sequenced: value first)

Status (2026-07-05): Phases 0–3 ✅ complete and exceeded (5 use cases vs the 1 planned MLV; +45.3pt vs the +15–20pt target). Phase 4 🟡 in progress — map published (public, GitHub Pages); bridge contribution packaged (contrib/); formal owl:Restriction re-modeling + EDM Council engagement remain (outward-facing, user's call).

Phase 0 — Scaffold (can start today 🟢) Fork Bodhi (own it), pin FIBO submodule to a Production tag, stub okf.config.js, ETL skeleton.

Phase 0.5 — De-risking spikes (blockers before real ETL)

  1. OWL restriction/blank-node extraction spike on FND+LOAN — prove we can pull typed object-property relations, not just triples.
  2. Decide serialization (RDF/XML from GitHub vs pre-serialized TTL/JSON-LD) + pin release. Done (spike): RDF/XML via rdflib; pinned to commit ae12db3 — note FIBO has no Production git tags (only 2018Q3); releases roll on master, so we pin a SHA. See fibo-source.pin.
  3. Define IRI→path mapping + version-churn/diff strategy + per-class maturity propagation.
  4. Object-property frequency analysis → which edges to promote.
  5. Definition/label coverage stat → sizes the curation budget.

Phase 1 — Minimum Lovable Version: prove the value (the answer to "is this just fancy tech?")

  • One use case end-to-end: Loan Origination (LOAN+FND+FBC+BE). Hand-curate ~40–60 concepts + the 5–10 lateral bridges it needs (skip full extraction).
  • Ship the context pack + retrieval endpoint (the product), and the curated map (the showcase/curation UI) with LOD.
  • Value-proof eval: loan-underwriting agent on a 50–100 Q financial-semantics benchmark, grounded vs ungrounded — score accuracy, hallucination rate, and % of answers carrying a valid FIBO provenance citation. Target: ≥15–20pt accuracy lift + 100% auditable answers. If it doesn't move, we learned cheaply.

Phase 2 — Harden the moat Full restriction-aware ETL for FND+LOAN; more lateral bridges with rationales; second use case (KYC).

Phase 3 — Scale domains (§7 recipe) with bridges + use cases per domain.

Phase 4 — Contribute + publish Export curated bridges (provenance: curated), re-model in axiom style, DCO-signed PR to edmcouncil/fibo via OKG Lab; publish the map as top-of-funnel.

11. ROI — honest estimates

Claim Estimate Certainty
Engineer-time saved grounding one agent use case (vs hand-curating FIBO context + cross-domain links) ~80–150 senior-engineer hours / use case Low–medium (plausible, unbenchmarked) — the one number worth putting in a deck
Accuracy lift / hallucination drop from ontology grounding External: ~54% avg KG-RAG accuracy lift; hallucination 48–63%→1.7% (clinical) Medium, cross-domain analogy — must measure our own financial number; do not present clinical stats as ours
Regulatory concept→report lineage effort 6-figure consulting line; even 10–20% cut is real Speculative — requires building lineage (not yet scoped)
EDM Council contribution Credibility/marketing; ~$0 direct revenue High certainty it doesn't pay bills

Buyers with a checkbook: the AI/agent builder (primary) and the compliance/model-risk team (traceable lineage). Ontologists, modelers, and FinTech engineers are users/contributors/spectators, not customers.

12. Readiness verdict & risks

🟡 YELLOW. Phase 0 starts today (licenses MIT, FIBO parseable, Bodhi structure confirmed). Phase 1 ETL is gated on the Phase-0.5 spikes.

Must-resolve blockers: (1) OWL restriction/blank-node extraction; (2) reframe + measure the cross-domain thesis; (3) pin release + serialization + maturity propagation; (4) IRI→path + churn strategy; (5) LOD rendering.

Underweighted risks: curation is person-months and financial/legal definition rewrites carry accuracy/liability risk (real SME review, not rubber-stamp); Provisional domains churn quarterly and break core:/bridge references keyed to moved IRIs; EDM OKG Lab may build the bridge layer first; OKF ecosystem is 4 weeks old (hence: export format, not thesis).

Audit trail (v1→v2): three-lens review — Readiness (verified FIBO/OKF/Bodhi facts, found the restriction-extraction gap), ROI (found the inverted value ranking, defined the MLV eval), Uniqueness (found the true differentiator is curation + intra-FIBO bridges). All three independently flagged the same reframe.

13. Sources