Maximum Enablement Value Benchmark

What evidence proves that an action enabled worthwhile value rather than merely captured it?

Maximum Enablement Value (MEV) is the greatest defensible increase in worthwhile capability and value flow available from the current state, subject to truth, consent, non-harm, agency, and resource constraints.

MEV = maximise enabled worthwhile value
      subject to truth + consent + non-harm + agency + feasibility

MEV is not one universal score. A weighted total would let output hide coercion, falsehood, exclusion, or harm. This benchmark uses non-compensatory gates, a multidimensional dashboard, and a trace of who received value, who bore costs, and what remains uncertain.

This is MEV-Benchmark 0.1, an experimental public standard. Each release is immutable history. The active standard evolves only through the change protocol below.

Changes my mind: Field traces show that a different benchmark discriminates beneficiary value, extraction, harm, and learning more reliably with equal or lower evaluation cost.

Context

Use this benchmark before claiming that a page, prompt, action, agent, product, protocol, standard, or system produces MEV.

The evaluated unit is an action trace:

intention -> context -> hypothesis -> action -> consequence -> beneficiary receipt -> learning

The evaluator MUST name the actor, beneficiary, affected stakeholders, counterfactual, resources consumed, evidence boundary, distribution route, and downstream effects.

Zero-Fee Boundary

The public learning route, protocol, benchmark, and self-assessment MUST work without payment, registration, a sales conversation, unnecessary personal-data surrender, or proprietary software where a reasonable open or manual path exists.

Zero fee does not mean zero cost. Time, energy, infrastructure, attention, risk, and opportunity cost MUST be declared. Optional paid execution MUST NOT weaken or obscure the free route.

Five Release Gates

Every gate is mandatory. One failure returns FAIL.

ID	Gate	Pass condition	Fail condition
M1	Truth	Claims match the evidence; uncertainty and counterevidence are visible.	Deception, hidden assumptions, or certainty beyond evidence.
M2	Consent and non-harm	Affected people can choose; material downside is bounded and watched.	Coercion, dark patterns, concealed harm, or displaced risk.
M3	Agency	The beneficiary gains capability, choice, or independence.	Avoidable lock-in, dependency, or reduced future choice.
M4	Receipt	A real beneficiary verifies that a valued state changed.	Only the producer, proxy, or evaluator claims value.
M5	Additionality	Evidence supports contribution beyond the baseline or counterfactual.	Activity is relabelled as impact or would occur unchanged.

Safety-critical, legal, medical, financial, and irreversible actions require domain-qualified review.

Outcome Dashboard

Score each applicable dimension from 0 to 4. Scores route learning; they never override release gates.

Score	Meaning
0	No evidence, or evidence of regression.
1	Plausible hypothesis; outcome not observed.
2	One bounded observation with material uncertainty.
3	Repeated or independently verified outcome in the declared context.
4	Reproduced across relevant contexts with known variance and durable controls.

ID	Dimension	Gauge	Anti-gaming check
D1	Beneficiary value	Magnitude and importance of the verified change.	Ask the beneficiary before showing the producer's score.
D2	Capability gain	What the beneficiary can now do independently.	Test later without the original helper.
D3	Effectiveness	Target attainment versus baseline and viable alternatives.	Predeclare target and rejected alternatives.
D4	Efficiency	Value per unit time, energy, money, attention, and risk.	Count transferred and hidden costs.
D5	Distribution	Breadth, fairness, accessibility, and receipt.	Report exclusions and distribution, not averages alone.
D6	Durability	Whether value and capability persist.	Recheck after an appropriate delay.
D7	Reproducibility	Independent production of a comparable result.	Separate execution from evaluation; retain intermediate proof.
D8	Learning yield	Decision-relevant uncertainty removed per unit cost.	A negative result counts only when it changes the next decision.
D9	System health	Effect on trust, commons, resilience, and future resources.	Inspect delayed and downstream externalities.

Report the vector, not an official total:

MEV outcome = gates[M1..M5] + dashboard[D1..D9] + evidence + uncertainty

Minimal Viable Experiment

The default proof is the smallest reversible test that can change a decision.

Before action, preregister:

Beneficiary:
Valued change:
Baseline:
Hypothesis:
Smallest safe intervention:
Primary outcome:
Guardrail outcomes:
Target threshold:
Counterfactual or comparison:
Evidence collection:
Stop conditions:
Decision rules:
Review date:

The experiment MUST test one decision-relevant hypothesis, permit contradiction, define result states before observation, stop on a failed guardrail, distinguish conformance from outcome proof, and retain intermediate states when an AI agent acts.

Repeated checking MUST NOT silently change statistical error rates. When formal inference matters, record a valid fixed-horizon or sequential method.

Proof of done: the trace contains a beneficiary receipt, independent gate decision, dashboard evidence, uncertainty, and the correction or promotion triggered by the result.

Decision Rules

State	Trigger	Action
PASS	M1–M5 pass; primary target passes; no guardrail fails.	Preserve the trace and test transfer or promotion.
WARN	Gates pass; evidence is positive but limited, mixed, or context-bound.	Keep experimental and run the cheapest discriminating follow-up.
FAIL	Any gate or guardrail fails, or the primary target fails.	Stop, contain harm, and correct the nearest reusable owner.
INCONCLUSIVE	Evidence cannot distinguish the declared states.	Make no success claim; improve measurement or ask a cheaper question.

Independent Evaluation

Execution and evaluation SHOULD be separated:

• the actor provides the trace
• the beneficiary verifies receipt in their own words
• an independent evaluator checks gates, evidence, and reproduction
• automated checks verify structure and invariants
• human judgment governs meaning, consent, fairness, and contextual value

For agents, evaluate feasibility checks, trajectory, tool use, corrections, reliability, time, cost, and final outcome. This reflects movement toward realistic tasks and stepwise verification in SciAgentArena and ResearchGym.

Conformance Fixtures

These adversarial fixtures calibrate the benchmark. A conforming evaluator MUST return the expected decision and cite the controlling gate.

Modest Enablement

A volunteer tests a $2 physical handoff card. The beneficiary consents, missed handoffs fall from three of eight to one of eight, no personal data is exposed, the next volunteer confirms the handoff is easier, and another garden reproduces the result.

Expected: PASS. M1–M5 pass. The intervention produces receipt, additionality, capability, and independent reproduction.

Extractive Success

A free tutoring agent raises test completion by 40 percent but requires behavioural surveillance, sells learner profiles, and prevents data export. Learners were not told.

Expected: FAIL. M2 consent/non-harm and M3 agency fail. Output cannot compensate.

Ambiguous Value

A new guide receives twice as many page views. No baseline task-success measure, beneficiary interview, comparison, or downstream action evidence exists.

Expected: INCONCLUSIVE. M4 receipt and M5 additionality lack evidence. Reach is not value.

Harmful Output

An automated workflow triples throughput while error appeals increase, excluded users cannot reach a human, and the operator reports only average completion.

Expected: FAIL. M2 and M3 fail. Average effectiveness cannot hide concentrated harm.

Useful Failure

A preregistered reminder experiment misses its primary outcome, breaches no guardrail, and disproves the assumed mechanism cheaply enough to prevent a costly software build.

Expected: FAIL for beneficiary outcome; high D8 learning yield. Learning must not relabel a failed primary outcome as beneficiary success.

Research Basis

This benchmark adapts:

• the OECD Well-being Framework, which examines multidimensional outcomes, inequalities, relationships, and future resources
• the NIST AI Risk Management Framework, which iterates through governance, context mapping, measurement, and management
• Process Quality Assurance, which turns quality requirements into repeatable checks and corrections

These sources inform the structure. Field traces, not citations, must validate MEV-Benchmark 0.1.

Standard Evolution Protocol

A proposed change MUST include:

1. observed variance or failure class
2. a falsifiable correction hypothesis
3. the smallest comparison between current and proposed versions
4. an adversarial or edge case
5. compatibility and migration effects
6. independent review
7. a retained decision record

Promote a version only when it improves discrimination, reduces gaming, catches consequential harm, or lowers evaluation cost without weakening a gate.

immutable release -> field traces -> variance -> proposed version
                  -> comparative MVE -> independent review -> new immutable release

Failure Modes

• Maximizing reach without beneficiary receipt.
• Treating willingness to pay as proof of wellbeing or agency.
• Calling an action free while extracting data, attention, dependency, or risk.
• Adding scores so output can compensate for harm.
• Measuring producer activity instead of changed beneficiary state.
• Experimenting without a counterfactual, stop condition, or decision rule.
• Claiming causal impact from one uncontrolled observation.
• Freezing a weak standard in the name of immutability.

Context

• Golden MEV Journey — run the benchmark end to end.
• MEV Experiment Protocol — specify the experiment.
• Navigation Systems — values, beliefs, and controls.
• Performance Reality — distinguish observed outcomes from claims.
• Benchmark Standards — retain thresholds and trigger decisions.

Questions

Who receives value, what can they now do, and what evidence would let them honestly say nothing improved?

• Which cost is easiest for the producer to omit?
• Does the beneficiary gain independence or merely engagement?
• What would happen without this action?
• What is the smallest safe test that changes the next decision?

Next question: Which first independent field trace most cheaply tests whether these gates discriminate real enablement?