How the registry works

Use-of-proceeds is the atomic unit

The registry is built around use-of-proceeds — what a bond's capital can be deployed to finance. A use-of-proceeds is describable and verifiable: "Construction of solar PV generation", "Restoration of coastal wetlands", "Installation of early-warning systems for extreme heat events".

Each use-of-proceeds gets a stable identifier in the ACT-U-XXXXXX family. An economic activity layer (ACT-A-XXXXXX) sits beneath, ISIC/NACE-aligned, providing producer-side interoperability. One economic activity can have multiple use-of-proceeds framings (e.g. manufacture vs installation vs operation of solar PV).

The operational test: can you write "the use of proceeds is X" and have it be verifiable? If yes, X is a use-of-proceeds. If no, X is a hierarchy category (ACT-H-).

Why the ACT-U / ACT-A split exists

"Activity" is an overloaded word in sustainable finance. Most taxonomies publish lists of eligible activities, but inspection of those lists shows something revealing: the descriptions are almost always use-of-proceeds descriptions (verb-phrases describing what bond proceeds can pay for), not the NACE/ISIC-style economic-activity classification the word "activity" technically denotes.

The EU Taxonomy's "Electricity generation using solar photovoltaic technology" reads as a verb-phrase; "construction, operation, and maintenance of …" is how many entries are worded. That is a use-of-proceeds, not an economic activity. The EU Taxonomy then additionally supplies NACE codes alongside each entry — which is the real economic-activity classification layer. So the EU bundles both layers under the single word "activity" and distinguishes them only via the NACE codes.

ACT-ID makes the distinction explicit:

• ACT-U — use-of-proceeds. What the bond's capital can be deployed to finance. Verb-phrase, financeable, verifiable. This is what bond frameworks list, what SPO providers verify, and what investors aggregate in their portfolios.
• ACT-A — economic activity. The producer-side NACE/ISIC sector classification. Useful for cross-referencing industrial classification systems and for answering "which economic activities does this labelled bond market touch?" It sits beneath the use-of-proceeds as a classification dimension, not as its own financeable thing.

Collapsing these into one entity (as most taxonomies do) makes three downstream things harder than they need to be: comparing taxonomies whose granularity differs, cross-walking to industry statistics, and resolving the many-to-one relationship where one economic activity hosts multiple use-of-proceeds framings (e.g. manufacturing solar modules vs installing them vs operating the plant are three distinct use-of-proceeds on the same NACE sector).

How existing frameworks name their atomic unit

Because the word "activity" is overloaded, this table maps the language of major frameworks onto the ACT-ID model:

When a taxonomy publishes an "activities" list, 9 times out of 10 the entries are use-of-proceeds in ACT-ID's terminology. When a bond framework references the EU Taxonomy, it's the use-of-proceeds layer that gets reproduced; the NACE codes are metadata, not the thing being financed.

How the ACT-A layer gets populated (and where it's intentionally sparse)

Only the EU Taxonomy publishes NACE codes for each of its activities in its source material. To make the ACT-A cross-reference useful across the rest of the registry, the v0.3 build assigned NACE codes to the subset of ACT-Us where NACE is genuinely the right classification dimension — national climate taxonomies, and the industrial / infrastructure subset of CBRT. Each ACT-U's economic_activity_authoring_method field records which route produced the link:

• source_native — NACE code supplied in the source taxonomy material (119 EU UoPs). Authoritative.
• registry_inferred — NACE code inferred by the registry via LLM (Claude Sonnet 4.6) on 2026-04-21, with confidence level and one-sentence rationale recorded in economic_activity_inferences. 1,089 UoPs across 22 national climate taxonomies plus the mappable CBRT subset.
• NULL — no NACE link, intentionally. Covers NEI accessibility indicators, CBRT social-development and community- resilience UoPs, policy-development and capacity-building activities, ecosystem-restoration UoPs that don't map to an industrial class, and any other UoP where "what industry is this?" is the wrong question to ask. 386 UoPs remain NULL by design.

Supersedence rule: if a taxonomy authority later publishes authoritative NACE codes for its activities, those override any inferred values; the inferred record stays in the audit table as a historical reference. source_native beats registry_inferred always.

The full audit trail for the 2026-04-21 inference pass — target set, LLM prompt, per-UoP rationales, and the SQL actually executed on production — is preserved in the repo at db/consolidation-audit/act-a-nace-backfill/.

Objectives carry structural properties

Environmental, social, and governance objectives are first-class records (ACT-O-XXXXXX) — not interchangeable string labels. Each objective carries properties that drive the rules: objective_domain, geography_dependency, accounting_basis, primary_measurement_unit, is_globally_fungible.

Climate mitigation is globally fungible, location-agnostic, flow-based. Biodiversity is ecosystem- or site-specific, non-fungible, stock-based. Resilience objectives are state-change-based and forward-looking. The schema enforces these differences — a biodiversity binding cannot be recorded without a location; a mitigation binding can.

The registry supports objective domains beyond environment and climate: social, governance, health, accessibility, economic_development, strategic, integrated, custom.

Criteria and thresholds are separate entities

Criteria (ACT-C-XXXXXX) define measurable requirements — what you measure and how you evaluate. Thresholds (ACT-TH-XXXXXX) define the specific values that satisfy a criterion. They are separate because a single criterion like "life-cycle carbon intensity" can be paired with different thresholds across taxonomies, time horizons, or contexts.

A criterion carries a canonical default threshold. A taxonomy's binding can reference the criterion and inherit the default, or supply an override_threshold_id when it varies. Thresholds carry their own lifecycle — applicability windows, conditions, source references — so "the EU 2030 threshold tightening" is modelled as a structured change, not a silent value update.

Bindings carry the judgements

Everything that varies by "how a specific taxonomy uses a specific use-of-proceeds" lives on a binding row, not on the canonical use-of-proceeds itself: the objective being served, the value-chain role (direct / enabling / transitional / upstream / downstream), the contribution sign (positive / neutral / negative / contingent), the benefit attribution (producer / operator / offtaker / investor / mixed), and geography at the precision the objective demands.

The same use-of-proceeds can have many bindings. Afforestation serves climate mitigation (carbon uptake), climate adaptation (resilient species selection), biodiversity (native species, soil conservation), and resilience (drought/pest tolerance) — simultaneously, each in a different taxonomy's language. One canonical record, many bindings.

Bindings carry a rank (preferred / normal / deprecated) so competing judgements — an expert reviewer dissenting from the taxonomy authority's position — are preserved as structured data rather than suppressed.

Mappings connect what shared identity does not

When two taxonomies bind to the same canonical use-of-proceeds, equivalence is implicit — no mapping needed. Mappings (ACT-M-XXXXXX) record the relationships that aren't captured by shared identity: partial overlap between distinct use-of-proceeds, cross-taxonomy alignment between hierarchy nodes, correspondences between differently-expressed criteria.

Each mapping carries a mapping_kind (activity_equivalence, activity_overlap, hierarchy_alignment, criterion_alignment) and a scope (equal, broader, narrower, partial_overlap). Status tracking (active, superseded_by_merge, deprecated) preserves legacy assertions as audit history when underlying records are later merged.

Taxonomies are frameworks over the activity list

Each taxonomy is a framework (ACT-T-XXXXXX) with a set of versions (ACT-T-XXXXXX-vN), a methodological_approach (rule-based / principle-based / green-list / traffic-light / transition-pathway / hybrid), a framework_purpose (sustainability_taxonomy / reporting_framework / internal_bank_policy / accessibility_framework / …), and a lifecycle_state (proposed / in-consultation / draft / published / implemented / under-revision / retired).

A hierarchy of nodes (ACT-H-XXXXXX) organises the use-of-proceeds for that specific taxonomy — different taxonomies can structure the same activities differently. Hierarchies are taxonomy-specific; the activities they organise are canonical and shared.

Institutions (ACT-I-XXXXXX) are first-class, linked to frameworks through role relationships (lead_developer, co_developer, convener, administrator, advisor, publisher). Most taxonomies are multi-institution — MAS + CBI for Singapore, ASEAN Taxonomy Board with member-state regulators, IPSF for the Common Ground Taxonomy.

Identifiers are hash-derived and opaque

Act IDs are not sequential numbers or semantic codes. Each suffix is derived from a SHA-256 hash of the record's canonical content at creation time, encoded in a URL-safe alphanumeric format. The identifier is intrinsically tied to the record.

Opacity is deliberate: an identifier that contained meaning (sector codes, version numbers) would break when the hierarchy changed or the version shifted. Hash-derived opaque identifiers survive every reclassification the publisher makes.

Versioning is via content hashes. Registration tiers (metadata_only / hierarchy_mapped / fully_bound) let the registry carry both depth — the 26 fully-bound taxonomies — and breadth, matching the jurisdiction-level coverage of dedicated trackers.

Lineage is metadata, not a separate entity

Version history is carried as structured metadata on each record — every canonical entity has derived_from (ancestors) and superseded_by (successors) fields. When two records are merged, redirects are installed so old identifiers continue to resolve.

A reference to ACT-U-W4G2UF in a 2026 bond framework resolves to the same semantic concept in 2030, even if the underlying record has been versioned, forked, or merged in the intervening years. Consumers who want the full lineage walk it through the metadata; consumers who just want the current record follow the redirect.

Binding-level provenance (authoring_method, reviewed_by, reviewed_at, source_document, confidence) is recorded inline on each binding plus a history table logging every change — so every judgement the registry makes is auditable back to its author, source, and date.