Skip to content

ADR-0025: TPI Management Quality — seed-backed, identity map via staircase

Status: Accepted (2026-05-24) Supersedes: obs-1 design note "report-parse extraction approach" framing (corrected) Superseded by:


Context

obs-1 (scraper sufficiency review) identified TPI as a Tier-1 source to complete the E-pillar positive signal set alongside NZBA, SBTi, TNFD, and IFRS-S2-MANDATE. The design note for TPI-SECTOR was produced in a chat session on 2026-05-24 and handed to CC-on-VM for implementation.

Two decisions needed recording: (1) instrument choice (MQ not Carbon Performance), and (2) the scoring map and its engine-implementation.


Decision 1: MQ only, not Carbon Performance

TPI publishes two instruments:

  • Management Quality (MQ): climate governance maturity, 0–5 ordinal staircase. Sector-agnostic framework applied to all assessed companies.
  • Carbon Performance (CP): emissions-pathway alignment, sector-specific benchmarks (power, steel, cement, etc.). Deduction character — below-pathway = adverse signal.

This build uses MQ only, consistent with obs-1's "management-quality grade" framing. CP is a different signal (sector-pathway alignment with a deduction character) and would require its own rule wiring and sector-pathway logic. Noted as a possible future addition (NF-E5-IND.3 is a pre-existing placeholder).


Decision 2: Seed-backed, not report-parse

The obs-1 handoff described TPI as needing a "report-parse extraction approach." This was a misread of the source. TPI publishes MQ grades via an open-access online tool — structured per-company values, not locked in prose PDFs. The per-sector PDFs are Carbon Performance methodology notes, not where MQ grades live.

TPI-SECTOR is therefore built as a seed-backed list (data/seed/tpi-mq.json), in the same pattern as NZBA-MEMBERS, TNFD-ADOPTERS, and IFRS-S2-MANDATE. Refresh cadence: annual, after TPI publishes updated MQ results (typically Q1). is_seed_backed = 1.


Decision 3: Identity map via staircase

The MQ scale is an ordinal 0–5 staircase. The methodology intent is: score_value = mq_level × 20 (L0→0, L1→20, L2→40, L3→60, L4→80, L5→100).

The scoring engine uses raw_score = base_score + Σ(points_i × boolean_i). To implement the identity map without engine modifications:

Staircase encoding: 5 binary sub-rules per rule (sub_rule_id 6–10), each worth 1 pt:

Sub-rule Description Fires when
.6 TPI MQ: Level ≥ 1 (awareness) mq_level ≥ 1
.7 TPI MQ: Level ≥ 2 (building capacity) mq_level ≥ 2
.8 TPI MQ: Level ≥ 3 (integrating operations) mq_level ≥ 3
.9 TPI MQ: Level ≥ 4 (strategic assessment) mq_level ≥ 4
.10 TPI MQ: Level ≥ 5 (transition planning) mq_level ≥ 5

pointsSum = mq_levelraw_score = clamp(mq_level, 0, 5)score_value = mq_level × 20

The staircase also reflects TPI's own methodology: a company must satisfy all lower-level indicators before advancing. Each sub-rule firing is semantically correct, not artificial.


Decision 4: Wired to E1 (financials) and UN-E1 (all sectors)

TPI "widens" E1 and UN-E1 — it does not light a dead rule. E1.6–E1.10 are financials-only (applicable_sectors='40'). UN-E1.6–UN-E1.10 are universal (applicable_sectors='ALL').

Institutions absent from the seed emit no signals (absence = not assessed by TPI, not a gap). No signal → Case A (no live evidence) → OMIT sentinel in scoring → no score movement. Absent institutions are not penalised.


Coverage finding (corrects obs-1 expectation)

The obs-1 design note predicted "low-to-mid teens" in-universe matches, assuming TPI covers heavy emitters only. TPI's 2024 expansion to 2,000 companies across 24 sectors (including software, pharma, medical devices, consumer services, and financial services) produced:

47 of 79 in-universe institutions confirmed in TPI (2026-05-24 seed).

Distribution: L5=11, L4=12, L3=21, L2=2, L1=1. Most companies (21/47) are at L3 ("par"), consistent with TPI's own reported ~57% at Level 3.


Ceiling-interaction note

With 10 sub-rules per rule (5 original + 5 TPI staircase), companies with both existing binary positive signals (SBTi, IFRS-S2-MANDATE) and TPI L3+ can reach score_ceiling=5 (100 pts).

Example: SBTi near-term (1) + SBTi net-zero (1) + TPI L3 staircase (3) = 5 pts → 100.

Assessment: this is correct and intended. Multiple independent strong sources exhaust the positive-signal space — reaching the ceiling is the right outcome. TPI L3 is "par" (ordinary good practice), but combined with verified science-based targets, it represents a company that has both formal commitment and governance integration. The ceiling interaction is more pronounced for non-financials (UN-E1) because CORP-WEBSITE sub-rules (1-2) remain deferred, leaving more of the ceiling achievable via SBTi + TPI alone.


Re-measure result (obs-1 Tier-1 complete)

With TPI-SECTOR joining TNFD-ADOPTERS and IFRS-S2-MANDATE, obs-1 Tier-1 is complete.

Run N Mean ESG Above 50
run_id=42 (TNFD + IFRS-S2-MANDATE, pre-TPI) 79 36.2 unknown
run_id=43 (all Tier-1 sources) 79 56.3 38/79

The mean ESG jumped from 36.2 → 56.3 (+20 pts), primarily from TPI's coverage of non-financial companies that previously had few positive E-pillar signals. The structural deduction-vs-affirmative imbalance is reduced but not eliminated — S and G pillars remain affirmative-light. The baseline-anchor decision (ADR-0012 + ADR-0023 re-open) is now unblocked with this measured distribution as input.