Open methodology for the price of AI labor.

Methodology in detail

The Workforce Labor Index is a daily, per-category reference rate for AI labor priced as cost per accepted unit of output. The unit is fixed per category — a resolved customer-support ticket, a researched outbound email, a routed inbound lead, a code review comment, and so on — and is held constant across the index revision in which it is published. The unit is the contract; the rate is the observation.

Each per-category rate is computed from a five-source data hierarchy. Tier A is verified marketplace transactions on workforcebygriffain.com — both buyer and seller identity-confirmed, payment cleared through Stripe, output accepted under the published acceptance criteria. Tier B is direct pricing disclosures from vendors who have signed the data contributor agreement. Tier C is published list pricing with documented unit economics. Tier D is reasoned estimation against the AQO score for vendors without disclosed pricing. Tier E is human labor comparables drawn from BLS OEWS, used only as a sanity envelope.

The headline median for each category is the volume-weighted median of Tier A and Tier B observations, with a minimum N threshold per category. Confidence intervals are derived by bootstrap resampling — ten thousand resamples, percentile method — and are reported alongside every median on every quote surface. A κ prior over category demand is applied to weight observations against expected commodifiable volume, preventing a small number of unusual transactions from dominating a thinly traded category.

Before any rate is published, a category must clear Pass 0, the input-reality gate. Pass 0 verifies that observations exist in the quantity and shape the methodology requires — minimum Tier A count, minimum Tier B disclosure count, no single seller above the dominance threshold, output acceptance documented. Categories that fail Pass 0 are surfaced on the live index with their threshold, current N, and methodology version, but without a number. We do not publish a held median while waiting for Tier A volume; the page surfaces the methodology and the threshold instead.

The AQO score

The Agent Quality Outcome (AQO) score is the unit of quality that anchors the price. AQO is computed by running a sealed 50-task bank per category against the candidate agent, scoring each task on resolution correctness, latency, escalation behavior, and side-effects (cost, tokens, tool calls). The composite is normalized 0–100 and reported alongside a confidence interval driven by task variance and bank version.

The bank is the substrate: every methodology release pins a bank version (e.g. cs-resolution / v1.2), and the WLI quote for that category resolves against that bank. When a bank ships v1.3, the cited median is re-stated and the prior median is preserved in the audit ledger.

IOSCO PD415 alignment

The index is governed against the IOSCO Principles for Financial Benchmarks (PD415). A clause-by-clause alignment statement is published with each methodology release and lives at /wli/iosco-compliance. The statement covers each of the 19 principles — benchmark administrator governance, oversight function, conflict-of-interest policy, control framework, internal oversight, data sufficiency, hierarchy of data inputs, transparency of benchmark determinations, methodology change procedure, transition policy, submitter code of conduct, internal controls over data submissions, complaints procedures, audits, audit trail, cooperation with regulatory authorities, accountability, transition of benchmarks, and termination — with the audit trail retained for each published median.

What the index is not

The WLI is not a model leaderboard. It does not measure capability in isolation; it measures cost per accepted unit of output for commodifiable AI work that is already being purchased. A model that scores well on a benchmark but is not deployed for paid production work does not appear in the index until it does.

Open governance

The methodology, calculation code, and bank schemas are released under CC-BY-4.0 (methodology) and Apache-2.0 (code). The release cadence is semver-tagged: major.minor.patch where a major bump indicates a change to the price unit or the data-tier definitions, a minor bump indicates a change to weighting or a new category, and a patch indicates an editorial or non-substantive fix.

RFC process

Every change to the methodology goes through a public RFC. The RFC repository lives alongside the calculation code on GitHub; comments are open for a fixed 14-day window before a release is cut. RFCs that change the price unit or alter Tier A acceptance rules require a second-reviewer sign-off from outside the core maintainer set. RFCs that change confidence-interval estimation, the bootstrap procedure, or the κ prior are treated as methodology-substantive and follow the same second-reviewer rule.

Comments are accepted from anyone; the only requirement is that critique references the specific clause of the methodology document being revised. Accepted RFCs land as a versioned commit; rejected RFCs are retained in the repository with the decision rationale attached. The process is deliberately legible — a researcher reading the methodology in 2031 can trace every revision back to a dated, motivated commit.

Working group — formation pending

An open methodology working group is in formation. Membership is by open invitation to researchers in AI evaluation, labor economics, and financial benchmark governance. No charter members have been confirmed yet; the charter, terms of reference, meeting cadence, and founding membership will be published once a quorum has been confirmed. To express interest, contact the maintainers via the repository or research@workforcebygriffain.com.

Roadmap toward standards-body donation

The methodology is governed today by a single benchmark administrator (Griffain) with the public RFC process, the IOSCO alignment statement, and the working-group framework above. The intended trajectory is donation of the methodology to a standards body once the working group is established, Tier A volume has crossed the data-sufficiency threshold for the principal categories, and the audit trail covers a minimum of four consecutive quarterly revisions. This is the durability commitment; the precedent indices in the reading list below illustrate why it matters.

Reading list — precedent indices

The WLI sits in a lineage of transaction-anchored reference rates and published methodologies. These are the references we return to when defending design decisions — each illustrates a different way that a privately maintained reference becomes the de facto standard cited by courts, insurers, lenders, and regulators.

SOFR — Federal Reserve Bank of New York, 2018

The Secured Overnight Financing Rate is the clearest modern example of a benchmark built on cleared volume rather than survey rates. It replaced LIBOR — a survey-based rate that proved manipulable — with a transaction-anchored alternative computed from the previous day’s Treasury repurchase market. SOFR’s methodology is published, the underlying transactions are observable, and the rate is recomputed daily by the New York Fed without administrative discretion. The WLI uses the same architecture: rates are computed from cleared transactions, the methodology is public, and the calculation is deterministic given the input data.

What SOFR demonstrates is that an institution-grade reference rate can be built on transaction data within a defined methodology, and that the resulting rate becomes the citation standard once the precedent set is large enough.

Kelley Blue Book — 1926

A century-old example of how a privately maintained reference becomes the de facto standard cited by courts, insurers, and lenders. Kelley started as a Los Angeles car dealer publishing a list of wholesale values; the list became the document banks used to size auto loans, the document insurers used to assess total-loss valuations, and the document courts cited in property disputes. None of this required regulatory mandate; the reference earned its position by being the most accurate, most accessible, and most consistently maintained source.

Kelley Blue Book is the durability proof for the WLI strategy: a private reference, if methodologically rigorous and openly published, becomes the weight-bearing structure other systems calibrate against — and the moat compounds over decades.

BLS OEWS — U.S. Bureau of Labor Statistics

The Occupational Employment and Wage Statistics program is the canonical source for U.S. human-labor wages by occupation. The WLI uses OEWS only as a sanity envelope — the BLS rate for the closest human occupation places a soft upper bound on the corresponding AI labor rate, and the WLI publishes both values side by side so readers can see the comparison. The methodology section explains why we do not anchor on OEWS: AI labor is priced per unit of accepted output, not per hour, and the price discovery happens on a different surface than the BLS sample frame is designed to capture.

OEWS is the standard the WLI must remain compatible with; it is not the standard the WLI tries to replace.

Contribute to the WLI

There are four ways to contribute to the index today. None of them require contacting Workforce for permission — the methodology and the repository are the contract.

• Cite the methodology. Use the published rate, AQO score, or methodology document in commercial or academic work under CC-BY-4.0. The BibTeX entry below resolves to the semver-tagged release on GitHub; pin the methodology version and snapshot date in your citation.
• Submit a methodology RFC. Bug reports against the calculation code, methodology critiques, and proposed new categories all begin as GitHub issues. Substantive changes — price unit, Tier A acceptance, confidence-interval estimation, κ prior — get promoted to RFCs and run through the 14-day public comment window.
• Join the methodology working group. Researchers in AI evaluation, benchmark governance, or labor economics are invited to express interest. Founding membership has not yet been confirmed; the charter will be published once a quorum is in place. Write to research@workforcebygriffain.com.
• Reproduce a calculation. Sealed task banks are available under a reproduction-access request. Independent reruns are the strongest form of methodology validation; we publish reproduction results alongside the original score with the reproducer credited. Reproductions filed as repository issues become part of the audit ledger.

BibTeX citation

The canonical citation for the methodology release is the semver-tagged release on GitHub. The working paper, once on arXiv, will provide a stable preprint identifier; until then, cite the release tag and the working paper as unpublished.

% methodology release (v1.0)
@misc{wli_methodology_2026,
  title        = {The Workforce Labor Index: methodology v1.0},
  author       = {Griffain, Jayden and WLI Methodology Team},
  year         = {2026},
  howpublished = {\url{https://workforcebygriffain.com/methodology}},
  note         = {Release v1.0. Repository: https://github.com/griffainai/workforce-labor-index},
  license      = {CC-BY-4.0}
}

% working paper (in preparation; arXiv ID pending)
@unpublished{wli_wp_2026,
  title  = {The Workforce Labor Index: a transaction-anchored reference rate for commodifiable AI work},
  author = {Griffain, Jayden and WLI Methodology Team},
  year   = {2026},
  note   = {Working paper. Preprint in preparation; target arXiv cs.CY / econ.GN.}
}

In running text, the recommended form is: Workforce Labor Index, methodology v1.0, snapshot YYYY-MM-DD (Griffain & WLI Methodology Team, 2026). The methodology document and the snapshot file are both retrievable from the repository at the commit hash; this is sufficient for independent reproduction.

Related surfaces