Every Alignment Score in the Workbench rolls up to one root question: for this supplier’s capabilities, in this market, how much will actually get absorbed into real enterprise work that produces real human and economic outcomes — and where?
The seven layers below are a MECE decomposition of that question. The three gates are the kill-criteria. The four-tier confidence ladder is what prevents fiction from being scored as fact. This page is the canon. It is auditable.
L1 → L2 → L3 (Gate 1) → L4 → L5 → L6 (Gate 2) → deploy → L7 (observe + feedback). Gates kill orthogonally — an enterprise can be mechanically ready and socially fractured, or socially aligned and mechanically broken. Either kills a deployment.
What work exists in the market, at what volume, in what shape. Lightcast (1.28B postings · 594M profiles · 28.8K skills) joined with TAG's proprietary depth (104M placements/yr) and public stats (BLS · ILOSTAT · OECD). The labor market is a joined dataset — no single vendor's product.
The Unit of Potential is the latent optimal configuration of resources — human and AI — to achieve a business outcome, grounded in business constraints. Work-first, human-first, constraint-first. The contrarian inversion of the model-first AI strategy industry.
Can this enterprise absorb AI transformation at all? Three composites: System Readiness × Organizational Readiness × Environmental Context. If blocked, stop. The gate that kills the 80% of pursuits without the organizational capacity regardless of UoP design.
For passing enterprises, who is best at which slice of the work? Honest multi-vendor allocation with human-retention gates. OpenAI · Anthropic · Google · Mercor · Cognition · OSS · the deploying entity's people. Truth as the only product surface. (Renamed from Neutral Shelf in v1.0.4 — the surface answers the buyer-side question 'how should I allocate vendors across this work?' more directly.)
Real roles, real geographies, real redeployment paths. Sourced from TAG's 104M annual placement substrate when available. Augment / reskill / retain / displace ratios at role granularity. Zero-elimination as a constraint, not a slogan.
Will the internal coalition actually commit to deploying this specific UoP? Procensus (Semantic-RBO consensus engine) detects factions, surfaces the irreducible disagreements, and audits stakeholder-level position before deployment. The gate that kills the 20% of those passing Gate 1 where the specific UoP will never land.
What really happened — outcome + human telemetry — feeding back to sharpen L3–L6 priors. Fusebox AgentFuze gives per-UoP real-time ROI; TAG's HR substrate gives the workforce truth. The flywheel: every In-Production UoP makes the next UoP's score more accurate.
Gate 0 · Agent Capability. A third gate upstream of L3. Can the agent-vendor frontier actually do the work in this archetype’s slices today, at acceptable reliability and cost? Five sub-dimensions: Task Capability (GDPval), Agentic Horizon (METR HCAST), End-to-end Project Deliverability (Remote Labor Index — max 2.5%), Exposure Structure (Eloundou α/β/γ), and Bottleneck Reduction (Frey-Osborne). Currency rule: capability scores age on a 90-day half-life — the frontier moves too fast for older numbers to carry load.
A UoP is a chain. The weakest link breaks it. Weighted sums hide chain-breakers. The min() rule forces every score decision to raise the binding constraint rather than optimize the average.
Can the frontier do this archetype's work today? Five sub-dimensions composited with currency-decay rules. Aging on a 90-day half-life because the frontier moves too fast.
System Readiness × Organizational Readiness × Environmental Context. Nine dimensions across the three composites — they fail orthogonally. Worked cases prove no two collapse to one signal.
Will the internal coalition commit? Procensus engine, faction detection, audit trails. Positions surfaced as data, not aggregated to a meaningless mean.
Every cap is visible on the cell. We do not silently downweight; we show the ceiling and the reason. Operators close the gap to lift the score, not by overriding the math.
Of 3,316 currently scored cells, 3,315 are Estimated. One — TAG × Harvey · Legal Drafting — is Measured. The progression along this ladder is the product.
Scored from external priors only — Lightcast firmographics, posting density, public benchmarks. No org-side enrichment. Composite capped ≤ 60.
Enriched with org context — sponsor identified, governance posture mapped, workflow dimension scoped. Composite uncapped, but no live measurement yet.
Stakeholder-validated through Procensus. L6 alignment cleared. Deployment plan committed. Awaiting telemetry.
Live telemetry from Fusebox AgentFuze + TAG placement substrate. Realized Potential measured against Estimated Ceiling. Independent of the deploying entity. Audit-grade.
Why this ladder matters operationally. A CFO who acts on an Estimated number and doesn’t know it is Estimated is the credibility-killing event. Every cell shows its tier. Every claim in the product carries its receipt.
The data layer is not a marketing badge. It is the substrate every score is anchored to. See the live data catalog →
| Source | Role in the cascade | Coverage | Refresh |
|---|---|---|---|
Lightcast Delta Share | L1 Substrate | 1.28B postings (16-yr) · 594M profiles · 28,795 skills · 7.3M companies | Daily |
TAG (proprietary depth) | L1 + L5 + L7 | 100K enterprise clients · 104M placements/yr · 300M candidate interactions/yr | Continuous |
O*NET | L1 ontology | 19,265 DWAs · 1,016 occupations · KSA + tasks | Per release |
BLS · OEWS · ILOSTAT · OECD · Eurostat | L1 cross-country | 249 countries · 396 metros · CPS + OEWS p10–p90 wages | Per release |
Eloundou et al. 2023 (OpenAI) | Gate 0 Exposure Structure | 100% of O*NET occupations · α/β/γ exposure | Static (research) |
GDPval (OpenAI 2025) | Gate 0 Task Capability | 1,320 tasks · 44 occupations × 9 GDP sectors | Quarterly |
METR HCAST Time Horizon 1.1 | Gate 0 Agentic Horizon | Doubling cadence · 7mo → ~3mo (2024–) | Quarterly |
Remote Labor Index (Scale AI + CAIS, Oct 2025) | Gate 0 End-to-end Project Deliverability | 240 real Upwork projects · max 2.5% frontier automation — the anti-hype anchor | Annual |
Acemoglu 2024 · Frey-Osborne 2013 · Dingel-Neiman 2020 | Gate 0 cross-checks | Macro cap · bottleneck residual · teleworkability admissibility | Static (research) |
WEF Future of Jobs 2025 | L3 Workforce Readiness | 1,000+ CHROs surveyed · 59% reskilling need by 2030 | Annual |
Fusebox AgentFuze telemetry | L7 Live ROI | Per-UoP real-time outcome + workforce actuals | Continuous |
Anthropic Economic Index | Gate 0 calibration | Task-level usage data | Quarterly |
Every weight, every threshold, every ceiling rule changes through the methodology board on a quarterly cadence. The full changelog is auditable. Six deployed UoPs see their composite shift ±2-5 points when weights revise; tier classifications hold within Ready / Gated.
| Date | Version | Change |
|---|---|---|
| 2026-04-22 | v1.0.3 | Category Captain → Neutral Shelf rename on external surfaces (internal canon preserved per Greg) |
| 2026-04-21 | v1.0.2 | GLG number reconciliation: 1B+/105M/1,900 → 3.5B postings / 1.65B profiles / 14,200 absorption-scored enterprises + TAG depth layer |
| 2026-04-21 | v1.0.1 | Uncomfortable-number research v2: SWE-bench Verified dropped (OpenAI retirement Feb 23); SWE-bench Pro + Terminal-Bench 2.0 added |
| 2026-04-20 | v1.0.0 | Methodology v1 locked; 7-layer cascade with MECE decomposition; archetype library at 7; Anchor Participant pricing $10-25M |
| 2026-03-30 | v0.9 | Density penalty recalibrated L1=0, L2=0.10, L3=0.22-0.38 per Greg Mar 9 architecture |
| 2026-03-15 | v0.8 | H2A Calibration Standard v0.1 locked: profession × reliability × density × output-type formula |
| 2026-02-20 | v0.7 | Mercor APEX-Agents added as capability ceiling anchor (24% Pass@1) |
Every Alignment Score in the Workbench links back to this page. Every confidence tier shown on a cell is defined here. Every Reason chip on a Gated cell traces to a ceiling rule above.