Taxonomy

Source: SOURCE-OF-TRUTH.md · synced 2026-06-12

Locked 2026-05-25. Three levels: sub_function → sub_vertical → tag. The first two are stable structural layers. The third (tags) is bootstrapped from data, not pre-defined — see §2 Step 3.1 note.

3a. Sub-Functions (top level — 5)

(Reference set has ~113 files total; ~107 are CVs after stripping images, screenshots, scanned-image PDFs. Final count after parser batch run.)

3b. Sub-Verticals (mid level — derived from Calibre’s JD folder structure)

The sub-vertical layer matters because rubrics fire at sub-vertical granularity. An LNG energy operator and a fertilizer agri operator both live under cargo-ops but score against very different criteria. Rubric configs key on (sub_function, sub_vertical).

3c. Tags (leaf level — bootstrapped from CV↔JD pilot match)

Approach: emergent, not designed. See §2 Step 3.1 note. The tag vocab will be the output of a lean pilot match — not a deliverable that has to land before the pilot runs.

Tentative draft tags below seed the bootstrap pilot (Claude can refine, drop, rename, add):

{
  "cargo-ops.agri":  ["biofuels", "coffee-cocoa", "edible-nuts", "fertilizers",
                      "grains-oilseeds", "oleochemicals", "rice", "sugar", "veg-oil"],
  "cargo-ops.energy": ["crude-oil", "fuel-oil", "lng", "lpg", "light-ends",
                       "middle-distillates", "petrochemicals", "thermal-coal"],
  "cargo-ops.metals": ["base-metals", "ferrous", "precious-metals"],
 
  "ship-ops.tanker":          ["chemical-tanker", "lng-carrier", "lpg-carrier", "crude-tanker"],
  "ship-ops.drybulk":         ["panamax", "capesize", "handysize", "supramax"],
  "ship-ops.container":       ["liner-ops", "feeder"],
  "ship-ops.offshore-support": ["psv", "ahts", "rig-support"],
  "ship-ops.broker":          ["voyage-charter", "time-charter", "sale-purchase"],
 
  "risk-analytics.credit":             ["counterparty", "settlement", "exposure"],
  "risk-analytics.market":             ["var", "p&l-attribution", "pricing-models"],
  "risk-analytics.operational":        ["controls", "incident-mgmt", "regulatory"],
  "risk-analytics.product-controller": ["valuation", "p&l-explain", "reserves"],
  "risk-analytics.trading-analytics":  ["quant", "model-validation", "backtesting"],
 
  "finance-fo.rm":                  ["onshore-hnw", "offshore-hnw", "uhnw"],
  "finance-fo.wealth-mgmt":         ["product-spec", "advisory", "discretionary"],
  "finance-fo.institutional-sales": ["funds", "asset-mgrs", "pension-funds"],
  "finance-fo.portfolio-mgmt":      ["equity", "fixed-income", "multi-asset"],
  "finance-fo.trading":             ["sales-trading", "execution"],
 
  "ops-back-office.aml":           ["transaction-monitoring", "kyc", "sanctions"],
  "ops-back-office.compliance":    ["mas-compliance", "advisory-compliance"],
  "ops-back-office.assistant-rm":  ["client-servicing", "documentation"],
  "ops-back-office.finance-ops":   ["reconciliation", "fund-accounting", "trade-support"]
}

These are seed candidates, not the locked vocab. Locked vocab is whatever the pilot match returns.

Relationship to the 4 grant verticals

CLAUDE.md and grant docs reference 4 verticals: Commodities & Renewables, Shipping & Maritime, Offshore Services & Start-ups, Technology & Gaming. These were the proposal-level marketing framing. The 5 sub-functions are the operational classification driven by Calibre’s actual data.

Mapping:

Action: when the Step 10 productisation surface ships, list 5 vertical “packs” on the marketing page: Commodities, Maritime, Risk & Analytics, Finance / Wealth, Banking Operations. The two original aspirational verticals (Offshore, Tech & Gaming) become “coming soon” placeholders. This keeps the grant’s “4+ verticals” claim defensible while reflecting real built coverage.

Superseded — original flat tag vocab proposal

The original taxonomy in plans/step-5-phase-2-cv-normaliser-design.md (formerly 2026-05-19-title-normaliser-design.md) was a flat 2-level vocab (cargo-ops:[energy:oil, metals:base, ...]). That doc is superseded by §3a/3b/3c here. The 3-level hierarchy + bootstrap approach is canonical. The original 2026-05-19 plan stays in the repo for audit traceability but its taxonomy section is no longer authoritative.