Start with the same job
Compare the agent and person against one clear task and one acceptance standard.
POWERRead the standard
A NEW REFERENCE FOR USEFUL WORK
Intelligence,
measured
in work.
APx asks a simple question:
How much finished, approved work did the agents produce?
Compare the same task. Count only accepted results. Include all elapsed time.
APx / INSTRUMENT 0013D + TIME
ILLUSTRATIVE OUTPUT
8AP
APx 3
THE FOURTH
DIMENSION
SCROLL TO UNDERSTAND PROPOSED STANDARD / v0.1 / BY rUvDEMONSTRATION, NOT TELEMETRY
START HERE / APx IN SIMPLE TERMS
Think of APx as
horsepower for agents.
A car engine needs horsepower so we can compare useful pulling power. Agent systems need APx so we can compare useful work. APx counts finished work that passes review. Busy agents, long runtimes and large amounts of code do not count by themselves.
A QUALIFIED PERSON8 hours
to complete and approve one defined task
THE AGENT SYSTEM1 hour
to deliver the same approved result
AGENTIC POWER8 AP
Eight times the human reference rate, shown as APx 3
Count work that passes
Broken, repeated or unverified output earns no credit. Its time and cost still count.
Include the full workflow
Measure waiting, retries, review, correction and human help from start to finish.
One sentence: APx tells you how quickly an agent system produces approved work compared with a qualified person doing the same job.
01 / THE UNIT
OUTPUT, NOT ACTIVITYNot how busy.
How useful.
Ten agents running for an hour tells you how busy the system was. It does not tell you whether anything useful happened.
AP measures accepted work per hour against a qualified human reference. APx puts that multiplier on a simple doubling scale.
One human baseline
The same accepted output rate as a qualified human, on the same kind of task, under a declared protocol.
Eight times the output
Eight reference hours of accepted work delivered in one elapsed hour by the declared system.
A compact exponent
Two to the power of three equals eight. Each extra APx point doubles the output multiplier.
THE DEFINITIONAP = accepted human reference hours ÷ elapsed hoursAPx = log₂(AP)
A proposed relative measure, not an SI unit, intelligence score, or employment replacement claim. Zero accepted output means 0 AP; its exponent is undefined.
02 / FEEL THE SCALE
Every step.
Twice the work.
Move the exponent. Watch the output grow.
The exponent is a display scale. It does not prove exponential improvement over time.
APx 38AP
−2 / 0.25 AP0 / 1 AP3 / 8 AP6 / 64 AP
In one hour: the accepted work of 8 reference human hours. Cells are rounded; the numeric value is exact to the displayed precision.
03 / THE POWER LAB
SYNTHETIC SCENARIOPut a real job
in the equation.
Choose a kind of work. Change the assumptions. The math updates instantly. These are teaching examples, not industry averages.
Fix a reproducible bug, add regression tests, and pass maintainer review.
Unique accepted bug fixes. Failed work earns no credit.
A hypothetical baseline here. Real scores need observed human trials, including failures.
Includes retries, waiting, review, repair, and integration.
AGENTIC POWER
8AP
ExponentAPx 3
8 accepted outputs × 120 reference minutes = 16 human reference hours, delivered in 2 elapsed hours.
Reference work2eight hour days
Human labor leverage32×Separate from AP
Cost per reference hour$0.75All costs entered above
Evidence statusSyntheticNo measured cohort
Acceptance comes first
Working code, regression tests, security checks, maintainability, reviewer acceptance.
What this means in plain language
If a system sustains 20 AP for eight hours, it delivers 160 reference human hours: 20 working days. Sustained for 250 such days, that is 20 reference working years of accumulated output. Neither statement means 20 years of work in one day.
04 / A SPECTRUM, NOT A CROWN
12 STARTING DOMAINSOne language.
Different standards.
A code fix, a book chapter, and a clinical workflow are not interchangeable. Compare like with like. A high score in one area says nothing about untested areas.
SELECTED DOMAIN / SOFTWARE
Qualified repository maintainer
Fix a reproducible bug, add regression tests, and pass maintainer review.
ACCEPTANCE GATE
Working code, regression tests, security checks, maintainability, reviewer acceptance.
A passing test alone does not establish a mergeable change.
These are proposed task profiles, not validated benchmark suites or measured occupational coverage. O*NET mappings require a versioned crosswalk and domain expert review.
05 / ONLY WORK THAT SURVIVES
Show the work.
Then count it.
Evidence moves through an acceptance gate. Errors, repeated artifacts, and unsupported claims do not get added to the score. Their time and cost still count.
FIX 001
Pending···Correct + reviewed
FIX 002
Pending···Regression test failed
FIX 003
Pending···Correct + reviewed
FIX 001
Pending···Same deliverable again
FIX 004
Pending···Correct + reviewed
FIX 005
Pending···Missing reviewer evidence
SYNTHETIC EVIDENCE REPLAY0reference hours
0 of 6 receipts processed. This replay illustrates accounting; it is not a live agent run.
Signatures prove origin and integrity. They do not prove that the work is correct.
06 / A MEASURE YOU CAN REPLAY
SPECIFICATION 0.1Built for the
RuV execution stack.
Keep measurement independent from optimization. The system can improve its approach. It cannot rewrite the exam to improve its score.
Define
Freeze the job, input data, quality rubric, budget, tools, environment, and completion window.
Calibrate
Measure qualified humans with the same protocol. Include unsuccessful attempts in the reference rate.
Execute
Run declared candidates. Meter elapsed time, human help, compute use, failures, and all costs.
Adjudicate
Use independent review and task checks. Reject unsafe, repeated, or unsupported outputs.
Compare
Calculate AP and APx. Report uncertainty, task coverage, assistance mode, and evidence grade.
Improve
Tune only on development tasks. Evaluate a separate holdout. Retain rollback and human approval.
THE FULL SPECIFICATIONDefinition, methods, governance, and adoption
THE RESEARCH BRIEFEvidence, prior art, and limitations
Check the core arithmetic.
Six browser checks exercise this explainer. The package contains the full Node test suite.
07 / BUILT ON EVIDENCE
RESEARCHED 07 SEP 2026The field is moving.
The measure must hold.
APx combines established evaluation lessons into a proposed accounting layer. It does not replace domain benchmarks or claim to be an adopted standard.
OpenAI
Agent workdays are runtime
The 3.1 workday figure uses summed agent runtime and an eight hour day. It is not a human productivity multiplier.
APx DESIGN RULEAPx reports runtime separately from accepted output.
06 SEP 2026METR
Human time measures task difficulty
Time horizons describe task difficulty at a stated success probability. They do not measure how long an agent ran or whether it can replace a job.
APx DESIGN RULEEvery APx score names a task family and reference cohort.
08 MAY 2026METR
The clocks get harder with concurrency
A followup productivity study identified selection bias and unreliable task timing when developers ran concurrent agents.
APx DESIGN RULEKeep wall time, agent runtime, and human labor distinct.
24 FEB 2026OpenAI · GDPval
Grade professional deliverables
GDPval evaluates work across 44 occupations using expert judging. Reported inference speed comparisons omit oversight and integration.
APx DESIGN RULEInclude review, repair, and integration before crediting output.
25 SEP 2025Mercor · APEX Agents
Work spans tools and applications
APEX Agents provides 480 professional tasks with prompts, files, rubrics, and gold outputs.
APx DESIGN RULEAPx is a distinct measurement proposal, not Mercor’s APEX benchmark.
23 FEB 2026O*NET Resource Center
Use a task taxonomy, not job stereotypes
The occupational content model distinguishes tasks, work activities, context, skills, and abilities.
APx DESIGN RULEKeep job, function, capability, and environment as separate tags.
ACCESSED 07 SEP 2026METR
Tests can miss real acceptance
A small study found that passing functional tests did not imply a change was ready to merge. Results were specific to its older model and tasks.
APx DESIGN RULERequire an independent acceptance gate beyond machine tests.
13 AUG 2025NIST
Protect evaluation integrity
The voluntary AI risk framework supports contextual evaluation and trustworthiness across a system’s lifecycle.
APx DESIGN RULESafety failures block publishable scores. APx is not NIST certified.
AI RMF 1.0 · 2023THE FIRST SCORE SHOULD EARN YOUR TRUST.
Less counting code.
More counting outcomes.
rUv’s measured APxNot measured yet.
Explore the specification The earlier 20 workday estimate and 100 million line count are not calibration data. A real score needs timed human reference trials and independently accepted agent outputs.
The next step: a matched pilot on bounded software maintenance tasks.