Pricing the Human Layer: Instrument Fit, Allocation Rules, and Settlement Design for AI Training Data

1Background

1.1 The industrial moment

Between 2022 and 2026, training data became a licensed industrial input. News publishers signed multi-year corpus deals with model builders; stock-media catalogs were licensed wholesale; per-crawl payment rails appeared on the open web; and human-generated data began commanding a price premium of ≥10× over synthetic substitutes in observed transactions [12]. Meanwhile the labor layer of that input — annotation, expert demonstration, validation — stayed on piece rates, and AI-generated content began flooding the same marketplaces that price human work: on one major streaming service, AI tracks reached 44% of daily uploads, with 85% of their streams judged fraudulent [11]. The question of how the humans in the data supply chain should be paid stopped being philosophical and became a pricing problem with live money attached.

1.2 The fairness gap: ownership rhetoric without machinery

The market pays a music catalog with a subscription, an annotator with a piece rate, a benchmark author with a one-time fee, and a stock photographer with a per-license royalty — and almost never explains why. Contributor pass-through of licensing revenue appears in fewer than ~6% of the deals we could verify, and where it exists it rides per-item lineage infrastructure (storefront metering, content fingerprinting) that the AI data supply chain mostly lacks. Platforms that promise contributors "ownership" of the datasets they build inherit the confusion at scale: which contributions can even carry a royalty? For how long? Split how — among roles, samples, and merged campaigns? What stops every rule from being gamed the week it ships? Equity-like compensation requires machinery — instrument screening, attribution, settlement, governance — and none of it is standardized.

1.3 Prior art: three literatures that don't touch

Data valuation supplies principled per-sample values — Data Shapley and its successors [1] — but the values are rank-unstable on deep networks (cross-run stability ≈ 0.04, versus ≈ 0.86 for Banzhaf values [2]), which disqualifies naive Shapley as a payment basis. Royalty economics (music, publishing, patent pools) supplies the closest working analogs: extreme revenue skew, advance-and-earn-out structures, collective licensing, and a long catalog of gaming precedents. Mechanism design and elicitation supply the tools — optimal auctions [6][7], incentivized preference measurement [4][5], monopsony evidence in online labor markets [8], and the data-externality results that break naive willingness-to-accept pricing [9]. None of the three answers the prior question: which instrument fits which data artifact. That mapping is this program's core contribution, with the rest of the machinery built behind it.

1.4 This program

The results below come from a research program run with register discipline: falsifiable hypotheses in a public register, pre-registered experiments with bars written before runs, adversarial red-teaming of every allocation rule, and honest reporting of failures (two invalidated runs and one failed rubric version are part of the record). Humanbased — a full-stack data-production marketplace whose campaign, validation, lineage, and settlement layers make these estimands observable — serves as the application laboratory; §10 situates it among comparable data-marketplace operating systems. Four contributions:

• A market census of 40+ real data deals, blind-scored into a nine-class instrument taxonomy (§2).
• A validated instrument-fit screen mapping a data artifact's value structure to its efficient compensation instrument, including the finding that evaluation data is structurally not a royalty asset (§3–4).
• A red-teamed four-layer allocation rule family for the royalty-suitable cells, with quantified settlement behavior, menu design, agent-participation bounds, and the economics of zero-upfront campaigns under tradeable ownership (§5–9).
• A landscape placement and governance protocol under which every production number is either a measurement with a replay command or a labeled policy with a falsification condition (§10–11).

2A census of how data is actually paid for

Before 2023, almost none of this market existed. Frontier models were trained on scraped public data at no cost to the source. Two forces ended that: a wave of copyright litigation (The New York Times, Getty, authors' and music-label class actions) turned unlicensed training into a balance-sheet risk, and the exhaustion of easy public text — against a ≥ 10× price premium for verified-human data — made the remaining supply worth paying for. So between 2023 and 2026 a licensing market was built essentially from scratch, and it did not converge on one way to pay. We cataloged 42 real deals and products spanning the pipeline; before distilling them to instruments (Figure 1), the raw industrial landscape is seven distinct regimes:

Reported figures are public deal disclosures and press, 2023–2026; several are cross-checked in the package's literature cards. They anchor magnitude, not precision.

Blind-scored against the instrument taxonomy, those 42 products collapse to the seven consensus classes in Figure 1.

What the landscape actually says

Read as industrial practice rather than a chart, three facts stand out — and together they define the gap this report exists to close:

• The money flows to intermediaries, not to the humans who made the data. Publisher deals pay corporations; expert-labor vendors pay a piece rate and keep the ongoing relationship. Verifiable pass-through of downstream revenue to the individual contributor appears in under ~6% of the deals we could trace — and only where per-item lineage already exists (stock, voice, Content ID). For AI training data specifically, individual human ownership is essentially absent.
• Royalty is the exception, and it is infrastructure-bound. Recurring pass-through royalty lives almost entirely in creator marketplaces that already own per-item metering and a storefront. The deals that dominate the dollars — publisher and book licensing — converge on lump-sum and subscription precisely because their attribution stops at the corpus. Instrument follows infrastructure, not fairness.
• The scarce layer is priced like temp labor. The RLHF / expert-data economy is the fastest-growing segment and supplies the expert human judgment models increasingly depend on — yet it pays by the hour or the task and attaches no downstream stake at all.

So "how data is paid for" in 2026 is not one practice but three disconnected regimes — corporate licensing, per-item creator royalty, and piece-rate labor — with no shared logic for which instrument fits what data, and almost no path for an individual contributor to hold an ownership stake in what they helped build. The rest of this report supplies that missing logic (§§3–5) and the machinery to give contributors a stake where it is efficient to (§§6–11).

3The instrument-fit screen — and how it earned its numbers

The census is not noise: instrument choice tracks five measurable properties of the artifact's value structure. Each is scored 0/1/2; fourteen mechanical decision rules map the vector to an instrument class.

Two structural laws fall out of the rules and survived every test: the granularity law (royalty granularity can never exceed attribution granularity — per-sample royalties on corpus-level attribution pay noise) and the term law (persistence bounds the royalty term — nothing that burns on exposure can carry one).

The rubric earned its accuracy the honest way: version 0 failed its own pre-registered bar (58–71% against the census, below the 80% target). Four anchor gaps were diagnosed, fixed, and re-tested on a fresh 16-product holdout — never by re-scoring the set it failed on. Figure 2 traces the full arc.

4Evaluation data is not a royalty asset

The screen's first named application was a hypothesis posed by the program's owner: evaluation and benchmark data should not carry royalties. The evidence is now graded A on the retrospective side. Every evaluation vendor in the census (6/6) prices by fee or subscription; none by royalty. Structurally, eval sets hard-fail three dimensions at once: they realize value at a decision gate (R = 0), burn on contamination — exposure inflates measured scores ≈10 points within 1–2 years, destroying the asset — (P = 0), and resist per-item attribution of a pass/fail judgment (A ≤ 1). By the granularity and term laws, there is no royalty term to price. The efficient rails are fee-for-service for sets, freshness-SLA retainers for continuously refreshed suites, and audit retainers for red-team work. The prospective half of the test — twin fee-vs-revshare term sheets of equal expected value offered to real buyers in procurement auctions — is designed and awaits execution.

5Where royalty is right: the four-layer allocation stack

For artifacts that clear the screen, ownership is computed — never negotiated per-dispute — by a deterministic projection through four layers, each with its own default rule and evidence-gated upgrades:

What each layer actually decides

The projection runs top-down: L1 sizes the three stakes, L2 splits the contributor stake across roles, L3 splits each role across the people in it, and L4 handles the case where these samples are later bundled with another campaign's to power one licence. Each layer has a deliberately boring, deterministic default; the "upgrades" (measured grades, usage-metering, Shapley values) are permitted only when a pre-registered evidence gate is met — otherwise the default stands.

Worked example: 1,000 samples, then a 700-sample bundle

Take the canonical campaign — 1,000 accepted samples built from four contribution types (human supply, human label, supply-validation on 10% of samples, label-validation on 20%). The owner paid an upfront commission; the signed policy sets owner 40%, platform 10%, contributors 50%.

L1 — the cap table (basis points of the whole dataset): owner 4,000 · platform 1,000 · contributor pool 5,000. Inside the invariant (platform exactly 1,000, contributors ≥ 3,000, owner ≤ 6,000), so it is admissible without adjustment.

L2 — split the 5,000 contributor bps across the four roles (spend-anchored, declared at publish; illustrative shares):

L3 — equal within role gives the per-task bps in the last column (validators earn more per task because there are fewer validation tasks over the same role budget). A contributor who supplied 40 samples, labeled 20, and ran 10 label-validations therefore holds 40 × 2.0 + 20 × 2.0 + 10 × 3.0 = 150 bps — 1.5% of the entire dataset — computed, signed, and byte-replayable, with no negotiation anywhere in the chain.

L4 — the cross-campaign case (the question that motivated this whole stack). Later, 200 of this campaign's samples are bundled with 500 from a different campaign to power one royalty-paying developer. At equal quality, the count × quality default splits that developer's royalty 200 / 700 = 28.6% to this campaign's slice, 71.4% to the other; within the 28.6%, each contributor is paid pro-rata to the ownership they already hold over those 200 samples (their L1–L3 bps). Note what the default deliberately does not do: it will not silently pay the 500-slice a premium for "driving demand," nor the 200-slice a premium for "completeness." Those arguments are real, but they enter through exactly one audited channel — a capped, pre-declared coverage bonus or a co-signed policy at bundle creation — never as ex-post discretion (EXP-008 shows the default tracks exact Shapley within ~5 points except in true completeness-threshold bundles, which route to the quote upgrade).

6Settlement under real-world skew: design for the modal dollar

Sections 2–5 asked whether data can carry a royalty and how to split one. This section asks the question a contributor actually cares about: what does a royalty feel like to receive? The answer reshapes everything after it — because on real datasets, royalty income is not a wage, it is a lottery ticket, and a system that pays people has to be built for the ticket most people hold, not the one a few win.

Why the average is the wrong number to design around. Dataset revenue follows a power law: a few licensed datasets earn almost everything and the long tail earns almost nothing — the same shape as music streams, book advances, and patent portfolios. Under that shape the average payout and the typical payout live in different worlds. Calibrating revenue to the empirically observed skew (Pareto α ≈ 1.02 — the top 1% of assets take ≈ 80% of revenue), we ran 4,000 dataset revenue paths through the full settlement machinery: the unglamorous plumbing of actually paying people over time — monthly accrual, a $25 floor below which it isn't worth cutting a check, a trailing-twelve-month (TTM) test that closes dormant accounts, and a terminal sweep that pays out whatever is left at the end of the five-year term.

The four numbers below are the result. The mean contributor entitlement is $25; the median contributor earns about $1. The $25 average describes almost no one — it is dragged up by a small lucky tail — so designing the settlement system around it would be designing for a person who does not exist. That is what the section title means by design for the modal dollar: build the mechanics, and the messaging, around the roughly one dollar the typical contributor actually receives.

How to read Figure 4. Each dot is a percentile of contributor outcomes, placed on a logarithmic dollar axis (each gridline is 10× the last). The one thing to see is the gap between the solid dot — the median, the typical contributor at $0.99 — and the hollow ring, the mean at $25. The average sits out past the 90th percentile: it describes the winners, not the field. When the "average payout" lands beyond nine of every ten people, the average is the wrong target and the median is the honest one.

Why this matters for the rest of the paper. Three later results rest on this one chart, which is why it sits where it does — the reality check between "royalty is admissible" (§§2–5) and "so here is how to make it humane" (§§7–9):

• It is why the menu in §7 exists. If the typical royalty is a near-worthless lottery ticket, a risk-averse contributor rationally prefers cash now — so the platform offers a buyout, priced at the κ ≈ 0.54 that is computed right here (the fourth tile: the population values $1 of expected royalty at 54¢ of certainty). Without this section, κ is a number from nowhere.
• It sets honesty constraints on the whole system. You cannot advertise wage-like income when the median is a dollar; and you cannot close sub-$1 accounts with a naive rule, because a TTM < $1 clause would fire on 93% of contributors while destroying almost no real value — catastrophic optics for nearly zero saving.
• It tells you which plumbing is safety-critical. The terminal sweep is not a cleanup step — it is the only payment 95% of contributors ever receive. Get the sweep wrong and almost everyone gets zero, regardless of how fair the cap table was.

Sections 7–9 — the menu, the fair buyout price, and tradeable ownership — are all machinery for making this lottery humane. This section is why that machinery is necessary rather than optional.

7Menus: let contributors choose their risk, honestly

The operator's problem, stated plainly: when the screen admits royalty, minimize the developer's upfront payment while granting contributors a fair ownership share — the certain cash should be the minimum that still attracts enough contributors. Expected recruits = participation likelihood × n(outreach); n varies by campaign, but the likelihood is computable from measured risk preferences, and this section computes it. Two instruments control it.

The first is the buyout exchange rate κ (kappa): how many dollars of certain cash the platform offers today per $1 of expected future royalty a contributor gives up — at κ = 0.54, a $25-expected entitlement trades for a guaranteed $13.75. Given the skew of §6, risk-averse contributors rationally prefer cash now over a lottery ticket, and calibrated to measured risk-preference distributions (Holt–Laury CRRA bands) the population certainty-equivalent of $1 of expected royalty is ≈$0.54 — so κ = 0.54 is the fair buyout price, not a discount imposed on contributors. We simulated menu self-selection across 16 preference cells at three exchange rates:

The recruitment frontier: how little upfront is enough?

The second instrument is the upfront floor φ — the certain cash paid per engagement, expressed as a fraction of the persona's market rate. Holding the ownership share fair (royalty expected value equal to the forgone wage bill, ρ = 1), participation likelihood becomes a computable curve in φ — and how much royalty substitutes for cash depends entirely on how contributors can exit the lottery:

Using the two dials — an operator's playbook

κ (how the buyout is priced) and φ (how much certain cash a campaign posts) are the two dials a campaign operator actually sets. The results above collapse into a short procedure:

1. Confirm the screen admits royalty first. If METH-1 says fee-or-retainer (evaluation data, burn-on-exposure work), these dials don't apply — stop here.
2. Set the ownership share fair, and hold it fixed. Fix royalty generosity so the entitlement's expected value equals the wage bill the campaign is not paying upfront (contributors at or above the 30% pool floor). The share is the fairness lever; the upfront is the cost lever — never trade one for the other.
3. Price the buyout at the fair κ ≈ 0.54, never below. Take-up is nearly flat under the fair rate (Figure 5), so underpricing the buyout recruits no one extra and simply transfers value away from contributors. Offer it as a corner leg alongside the full-royalty corner — and never a linearly-priced middle, which draws 0%.
4. Pick the participation likelihood you need, then read the minimum φ off the frontier. Expected recruits = likelihood × outreach. If outreach is cheap to widen, deliberately accept a lower likelihood and a lower upfront; if the persona pool is tiny (scarce specialists), buy a high likelihood and post more cash.
5. Buy the liquidity discount only if you can pay its price — transparency. Tradeable ownership is the single biggest lever (φ from 0.39 to 0.17 at the 50% line), but it only works with public per-asset revenue lineage; without it the secondary market prices at the prior and dies of adverse selection (§9). No lineage means you are on the κ-menu curve, not the tradeable one — plan φ accordingly.
6. Check the revenue back-stop before publishing. The promised expected value has to be fundable: a fair royalty on the 50-surgeon corpus promises $300k, which needs ≥ $600k of plausible lifetime licensing revenue. If the demand can't back it, lower the royalty generosity and move up the φ curve — never advertise expected value the demand side can't pay (DC-1's demand-assurance gate enforces this).
7. Keep φ above its own floor. Minimizing certain cash is the goal; driving it to zero is a different regime — it reclassifies contributors as investors and trips the S-03 floor and the zero-upfront admissibility rule (§9).

8Agents, laundering, and the price of manipulation-resistance

Two adversarial questions dominate the taxonomy's top stakes: what happens when AI agents can do the work, and what happens when contributors game the counting rules. Both now have quantitative answers.

Splitting merged bundles — and the one regime that breaks the default

When slices from multiple campaigns merge into one licensed dataset, the default splits royalty by count × quality. Tested against exact Shapley values over six synthetic bundle economies (the motivating case: a 200-sample slice joined with a 500-sample slice), the default is far more robust than designed intuition suggested — and its one failure is precisely diagnosable:

9Zero-upfront campaigns and tradeable ownership

The hardest stress test: a campaign that pays no upfront at all — contributors work purely for royalty entitlements. Under the L1 curve this is permitted (the owner's share falls to its 10% base; contributors hold 80%), but the economics reclassify the contributor from worker to investor paying with labor: at the observed skew the modal dataset earns ≈ nothing, only the risk-tolerant tail (~34% of the preference population) rationally participates, and wage-law plus securities-analysis surfaces activate. Demand-side risk — will any buyer pay? — dominates every other risk, and most of the mitigating machinery already exists in the platform designs we audited: a three-rung license ladder (evaluation / production / redistribution) for price discrimination with a sample-and-sandbox discovery tier; verified-organization gating with the first royalty period held in escrow before a campaign is visible; and metered usage modes with hash-chained receipts. The one missing instrument is a demand-assurance launch gate: escrowed, refundable anchor-buyer pre-commitments that must fill before contributor labor starts — an assurance contract that converts unobservable demand into a binary pre-labor observable.

Does making the ownership tradeable — liquidity before value accrues — change the picture? We priced it (EXP-009; four pre-registered bars, all pass):

Running a zero-upfront campaign — an operator's playbook

Zero-upfront is not the far end of §7's cost dial — it is a different regime with its own admission rules. The procedure:

1. Reach for zero-upfront when the dataset is otherwise unbuildable, not to save money. The recruitment frontier already cuts certain cash 86% at fair terms with tradeable ownership; the last 14 points buy you a regime change — contributors become investors, with the wage-law and securities surface that implies. The legitimate use-cases: long-tail datasets no single buyer funds at market rates (the $34k rare-disease vignette), and genuinely speculative assets where demand is unknowable upfront.
2. Gate who may enter, not just the terms (DC-1). Investor-classed opt-in that shows the real outcome distribution — median ≈ $1, p90, the sweep mechanics — so people are recruited with the truth; never for subsistence-class task categories; owner must be a verified organization with the royalty pool escrowed for multiple periods, because with no upfront the escrow is the only skin-in-the-game signal left.
3. Convert demand risk into a pre-labor observable. The demand-assurance launch gate: escrowed, refundable anchor-buyer pre-commitments must reach the declared threshold before contributor work begins — an assurance contract that turns "will anyone buy this?" into a binary answered while it's still free to walk away. Keep the sample-tier listing live from day one so demand discovery runs during supply, not after.
4. Pick the right supply pool. Zero-upfront selects the risk-tolerant tail (~34% of the population; liquidity widens it). The best-fit personas are those with low opportunity cost and high expertise — semi-retired and emeritus experts, portfolio moonlighters — never workers for whom the floor is the rent.
5. Sequence liquidity in the DC-2 order. The κ buyout menu ships first as the liquidity floor. Trading unlocks only with: public per-asset revenue lineage live (or the market lemons out), transfer gating and payout-safety-window locks, beneficial-owner surveillance on trades, no market-price display on recruitment surfaces (marks are manipulable; promises of mechanics are not), and the accrued-royalty question settled before the first trade (accruals sweep to the seller; units stay fungible).
6. Write the comms for the modal outcome, not the mean. Platform-computed p10/p50/p90 scenarios on the listing — never owner-supplied, never wage-like language. The median contributor's entitlement is ≈ $1; the pitch is equity in a long shot plus a liquid exit, and saying anything else borrows trust the settlement data will not repay.

10The landscape: where this sits among data-marketplace operating systems

The methodology was built to generalize; the census makes the comparison concrete. Every existing "data marketplace OS" archetype implements one slice of the instrument space — usually the slice its lineage infrastructure can support:

Read against the screen, the pattern is exact: each archetype's instrument ceiling is set by its attribution and excludability infrastructure. Creator marketplaces reach royalty-class because they own per-item lineage and metered delivery (A = 2, E = 2); bulk licensors sit at subscriptions because payment attaches to the corpus (A = 0); labor networks stay at piece rates because nothing downstream is metered at all. A full-stack data-production OS — campaign definition, validation, per-contribution lineage, metered usage modes, deterministic cap tables, governed settlement — is what it takes to make the whole instrument menu available and let the screen, rather than infrastructure poverty, choose the instrument. That is the Humanbased wager, and the reason its economics had to be built as a science rather than borrowed from any single archetype: no existing system spans enough of the space to copy.

11Governance: how numbers earn the right to move money

What this section is for. Every earlier section produced numbers — κ = 0.54, the ownership curve, the agent pay bound, the slice split. A number in a research report and a number that decides a real person's payout are not the same thing, and this section is the discipline that separates them: what a number must survive before it is allowed to move money, and how it is allowed to change afterward. Without it, the fairness of everything above is unverifiable in practice — a contributor would have to simply trust that a rule wasn't quietly retuned to cut their share. Governance is what replaces that trust with proof.

Three failure modes make it necessary. If money-moving numbers changed on every recomputation: contributors could not plan (their deal would shift under them); whoever controlled the recomputation would control the money (a gaming surface); and no one could tell a fair rule from a self-serving one. The methodology therefore ships inside a two-speed governance lane — one speed for rules you need before the evidence is in, one for rules that must wait for it.

Two speeds: doctrine and science

Doctrine — the fast lane. Production-blocking rules you need now that rest on theory and precedent — the two you have already met, agents-as-SaaS (S-28) and pay-validated-task-units (S-14). These ship immediately as labeled policy, but each carries an explicit falsification condition (the exact evidence that would overturn it) and a review date. A doctrine is a conservative default with a self-destruct clause, not a permanent truth — honest about being provisional.

Science — the slow lane. Measured parameters (exchange rates, grade multipliers, elasticities) are not allowed to go live on belief. A parameter earns the right to move money only by surviving seven gates, in order:

1. Pre-registration — committed to a signed formula registry before it is computed on the deciding data.
2. Effect evidence — a real measured effect, its confidence interval clear of the null.
3. Replication — it holds on independent campaign cohorts, not one lucky sample.
4. Simulation stress — no gaming strategy can lift an attacker's share past 2× their true contribution.
5. Shadow run — scored silently alongside the live rule on ≥ 30 campaigns before it touches a payout.
6. Red-team window — an open period for anyone to break the exact candidate.
7. Versioned activation — signed, and switched on for new campaigns only.

Three protections that make it trustworthy

• Forward-only. A rule change binds new campaigns; campaigns already signed keep the rules they agreed to. No one's deal is ever rewritten retroactively.
• Corrections pay off-ledger. A genuine error is fixed by a separate make-whole payment, never by editing the immutable ownership record — fairness and an unfalsifiable ledger at once.
• The registry is a pre-registration ledger. Every money-moving number is committed and signed before use — the same idea as pre-registering a scientific experiment so the result can't be tuned after the fact. That is what makes the platform a cumulative experimental instrument rather than a policy of the week.

How it fits. Where §§2–9 produce the rules and §10 places them in the market, this is the meta-layer that governs all of them — the part that turns a research report into a system you could actually run without asking contributors to take fairness on faith. It is the operational backbone of the report's central promise: every number in production is either a measured estimate with a replay command, or a labeled policy choice with a falsification condition — never an arbitrary one.

12What this report cannot claim yet

Honest scope limits, in order of importance:

• Every behavioral parameter is a literature prior, not a platform measurement. κ ≈ 0.54, elasticities, risk-preference bands — all graded imports (48 parameters, graded A/B/C). The designed field instruments (choice experiments with real-stakes anchors, procurement auctions) replace priors with measurements only after the ethics protocol is ratified.
• The rubric's truth standard is partly circular. "Market-converged instrument" is a defensible benchmark except in exactly the power-asymmetric cells this platform exists to change (expert annotation labor pays zero royalty today — the thesis gap, not necessarily rubric error). Escaping the circularity requires the independent standard the field work provides.
• H-E1's prospective half hasn't run. The retro evidence is grade A; the twin term-sheet auction with real buyers is the falsifying test.
• All simulations are synthetic-parameter mechanism tests. Shapes, not magnitudes; the M6 harness re-bases on realized platform data as it accrues, and valuation ground truth needs ≥8 completed campaigns with full lineage.
• Raters are same-model-family; human inter-rater reliability is likely lower than the reported arc. External review is a stated exit gate.

13Hypothesis register — the scoreboard

What this section is for. This is the report's honesty ledger — one row per major claim, showing exactly how much of it is currently earned. Its whole job is to let you weight every finding above correctly: to tell "validated on 40 real deals" apart from "held up in one simulation," which this report has sitting side by side. For a body of work whose thesis is that fairness should be tested, not asserted, a table where claims can be — and one already was — refuted is the proof the method is real. Read it as: trust this much, no more, and here is how to check us.

How to read a row. Each badge carries two things — how far along the claim is, and how strong the evidence behind it is:

• Status — SUPPORTED = validated, earned belief · TESTING = a real signal exists but the decisive test hasn't run yet · OPEN = a claim deliberately not made yet because the data doesn't exist (an honest blank, not an omission).
• Evidence grade — A primary/strong · B reputable survey or market observation · C simulation or ballpark.
• Evidence type — retro = checked against deals that already happened · in-silico = holds in simulation only · field = a live experiment (none of these have run yet).

So "TESTING · retro grade A" means a strong real-world signal from existing deals with the live test still pending; "TESTING · in-silico C" means it holds in simulation and nothing more yet. Only H-S1 is SUPPORTED — and only after its first version failed (§3), which is what makes the pass credible.

How it fits. The OPEN rows are not loose ends — they are the same items §12 names and the human-dependency lists point to (external reviews, the real-buyer auction, the field studies). So this table doubles as the roadmap: what would move each claim, and roughly when. One SUPPORTED, five in TESTING, a bank of OPEN — that ratio, stated plainly, is the report's most honest sentence.

14Takeaways: what to remember, and what to build

If you read this to understand the problem

Five ideas are meant to outlive the specific numbers:

1. The compensation instrument is a property of the data, not a matter of taste. Five measurable dimensions predict which of nine instruments a real market converges on. Royalty is one of them — and a minority one. A benchmark, a bulk corpus, and an annotation task each want a different instrument, and the screen says which; reaching for royalty everywhere is the exact mistake it exists to prevent.
2. Royalty income is a lottery, not a wage. Dataset revenue is a power law, so the typical contributor earns ≈ $1 while a few earn thousands. Any royalty system therefore lives or dies on its settlement plumbing and the honesty of its messaging — not on its headline rate.
3. Menus and liquidity are how you make the lottery humane — and cheap to staff. Letting people trade a fairly-priced royalty for cash (κ ≈ 0.54) and making ownership tradeable cuts the certain cash a campaign must pay by up to 83% at equal recruitment. Liquidity is a recruiting instrument, not a feature.
4. Every anti-gaming choice has a measurable price, and paying it beats being farmed. Manipulation-resistant rules deliberately diverge from the theoretically-fair ones; the gap is quantifiable and worth it. Collapsing everyone to their beneficial owner is the single master control.
5. In production, fairness comes from governance, not good intentions. A number earns the right to move money through pre-registration and evidence, changes forward-only, and corrects off-ledger. That discipline — not the rate card — is what a contributor can actually verify.

References

Citation

Please cite this work as:

BibTeX: