THE AMPLIFIED ECONOMY

Amplification Initiative — Minimum Viable Measurement Framework

Package 0.5 Discussion Draft | September 2026

Companion to:

The Amplified Economy — Conceptual & Experimental Architecture

Amplification Initiative — Newfoundland and Labrador Full Learning Pilot Portfolio Design

Amplification Initiative — Experiment Operating Protocol

Status

Provisional measurement architecture. The goal is not to create a universal productivity index or a single amplification score. It is to define the minimum evidence needed to compare experiments, make stage-gate decisions, detect exceptional outcomes, preserve uncertainty and learn across very different sectors without drowning participants in reporting.

1. Measurement objective

The measurement system should answer six questions for every intensive experiment:

1. Did something materially improve?

2. How confident are we that the improvement is real?

3. What changed in the human-AI-workflow configuration?

4. If a local gain occurred, how did it propagate through the wider system, and what constrained, redirected or absorbed it?

5. What became possible or scarce next?

6. Is the result worth stopping, redesigning, replicating, scaling or diffusing?

The framework should make it difficult to confuse:

• speed with productivity;

• output volume with quality-adjusted value;

• a prototype with an operational system;

• a local workflow gain with a system-level gain;

• self-reported usefulness with verified performance;

• correlation with causation;

• one exceptional worker with a transferable practice; or

• private return with public additionality.

2. Measurement principles

2.1 Minimum sufficient evidence

Collect the minimum evidence required to make the next resource-allocation decision.

Measurement is part of the experiment, not an administrative layer added afterward.

The program should prefer a smaller set of decision-useful measures over large dashboards that create reporting burden without changing decisions.

2.2 Effect size and evidence quality are separate

A 300% apparent gain measured poorly should not outrank a 20% gain measured credibly.

Every result should therefore carry two separate descriptions:

EFFECT

How large and valuable does the observed change appear to be?

CONFIDENCE

How strong is the evidence that the observed change is real, attributable and durable?

These dimensions should never be collapsed into a single score.

2.3 Quality before volume

More output is not automatically more value.

Output measures should be adjusted for:

• correctness;

• usefulness;

• completeness;

• safety;

• compliance;

• customer or user acceptance;

• expert quality; or

• another domain-relevant quality criterion.

2.4 Measure the whole result path

Time and cost accounting should include:

• generation;

• verification;

• rework;

• coordination;

• exception handling;

• integration;

• waiting;

• deployment; and

• other material downstream work.

A workflow that generates output ten times faster but creates nine times more review work may show little realized system gain in its current configuration. That does not mean the local amplification is unimportant. It may instead reveal that review has become the new binding constraint. The program should distinguish between a local capability gain, the system value currently realized, and the additional system value that could be unlocked if the new constraint were relieved economically, safely and without unacceptable human cost.

2.5 Pre-specify what matters

Before the intervention, each experiment should define:

PRIMARY OUTCOME

The main result that determines whether the experiment is useful.

MINIMUM WORTHWHILE EFFECT

The smallest improvement that would justify continued investment, given cost, risk and implementation burden.

FAILURE CONDITION

The result that would trigger STOP or REDESIGN.

This prevents post-hoc metric selection.

2.6 Domain-specific quality, common evidence structure

A software defect, a construction estimate, a policy analysis and a public-service interaction require different quality measures.

The program should not force them into one numerical quality scale.

Instead, every experiment uses domain-appropriate quality metrics inside a shared evidence structure.

2.7 No single amplification score

The program should not combine economic return, human return, strategic value and evidence confidence into one headline number.

Single scores invite false precision, hidden value judgments and Goodhart effects.

The decision record should preserve the separate dimensions.

3. Minimum Evidence Packet

Every intensive experiment should produce one compact core evidence packet.

The packet contains six domains:

A. PRIMARY RESULT

B. QUALITY AND VERIFICATION

C. TIME, COST AND CAPACITY

D. HUMAN RETURN

E. STRATEGIC / LEARNING RETURN

F. EVIDENCE CONFIDENCE

Not every metric is equally important in every experiment, but every domain must be considered and marked as measured, not material, or unresolved.

4. A — Primary Result

4.1 Primary outcome

Each experiment should select one primary outcome before the intervention.

Examples:

• quality-adjusted units completed per week;

• turnaround time for a defined service;

• cost per compliant output;

• error-free estimates completed per staff-hour;

• cases resolved within service standard;

• viable design alternatives evaluated;

• revenue-producing proposals generated;

• previously outsourced work completed internally;

• number of economically viable projects identified.

The primary outcome should directly represent the problem being solved.

4.2 Secondary outcomes

Secondary outcomes may capture:

• speed;

• cost;

• quality;

• capacity;

• worker experience;

• customer experience;

• error;

• resilience;

• new capability; or

• option creation.

Secondary measures should not be substituted for the primary outcome after results are known.

4.3 Minimum Worthwhile Effect

Every experiment should define a Minimum Worthwhile Effect, or MWE.

The MWE is:

The smallest measured improvement that would justify continued investment after considering implementation burden, full cost, risk and organizational disruption.

The MWE should be established before the main experiment. Passing it establishes that an improvement may justify continued investment; it does not by itself establish amplification. If amplification is claimed, pre-specify a separate, context-appropriate magnitude criterion and resource basis for the material and disproportionate quality-adjusted gain.

Examples:

• at least 20% reduction in total cycle time while maintaining quality;

• at least 10% improvement in accepted output per labour hour;

• positive full-lifecycle net benefit within a defined payback period;

• ability to perform previously outsourced work at comparable quality and lower total cost;

• reduction of a service backlog large enough to materially affect waiting time.

The framework should not impose one MWE across sectors.

Where a precise MWE cannot initially be established, the experiment should state the decision logic qualitatively and tighten it before scale.

MWE forms

Use the form that matches the mechanism rather than forcing every experiment into a percentage-improvement target.

PERFORMANCE MWE

A minimum improvement in speed, cost, quality or capacity.

VIABILITY MWE

The point at which a previously uneconomic or impractical activity becomes worth doing after full lifecycle cost and risk.

CAPABILITY MWE

A defined new capability becomes possible at acceptable quality, cost and risk even when the baseline output was zero or absent.

Capability counterfactual

Creation and Frontier experiments should record both the operational baseline and, where relevant, the counterfactual route to obtaining the new capability: conventional development, outsourcing, specialist hiring, indefinite deferral, or no project. Do not express a transition from “not available” to “available” as an arbitrary percentage improvement.

5. B — Quality and Verification

5.1 Quality measure

Every experiment must define what “good” means independently of speed.

Preferred quality evidence includes:

• blinded expert review;

• objective correctness;

• defect rate;

• rework rate;

• acceptance rate;

• downstream success;

• compliance pass rate;

• customer outcome;

• benchmark performance; or

• a pre-defined rubric scored by qualified reviewers.

Self-rated quality alone should normally be insufficient for SCALE.

5.2 Quality-adjusted output

Quality-adjusted output is a concept, not necessarily one universal formula.

The preferred approach is:

Report output quantity and quality separately, then state the combined operational implication.

Example:

Baseline:

20 estimates/week

95% accepted without material correction

Experiment:

32 estimates/week

94% accepted without material correction

Interpretation:

60% higher volume with essentially stable quality.

If a domain has a credible established weighting system, a combined quality-adjusted measure may be used.

If the weighting would be arbitrary, do not manufacture a synthetic score.

Where downstream capacity constrains additional volume, quality per completed unit may become a primary expression of amplification. In those cases, record whether more time, testing, alternatives, iteration or verification are being invested per unit and whether independently measured quality actually improves.

5.3 Verification burden

Verification burden should be measured when AI output requires checking.

Minimum measures:

• verifier minutes per unit;

• verification share of total labour time;

• proportion of outputs requiring correction;

• severity of detected errors; and

• residual error after verification.

Useful derived indicator:

Net time gain

= baseline total time

− experiment total time including verification and rework.

The program should report the amplified-stage gain and current realized system gain separately where they differ materially. Where a downstream constraint is newly binding, it should also estimate amplification potential: the system value that could plausibly be unlocked after a costed, feasible response to that constraint.

5.4 Rework and defect burden

Track:

• first-pass success;

• correction cycles;

• defect severity;

• downstream failure; and

• cost of correction.

A workflow should not be described as having higher realized system productivity if increased rework erases the apparent speed gain. However, a validated local capability gain should still be recorded as an amplification signal when rework or review has become a newly binding constraint. The next question is whether that constraint can be redesigned, automated, stratified, parallelized or resourced at acceptable cost and risk.

6. C — Time, Cost and Capacity

6.1 Time-to-result

Time-to-result should include the entire relevant workflow.

Possible components:

• active human time;

• AI processing time where operationally material;

• waiting time;

• handoff time;

• review time;

• rework;

• approval delay; and

• exception handling.

The experiment should distinguish:

ACTIVE WORK TIME

Time people or systems are actively performing the task.

ELAPSED CYCLE TIME

Total time from start to usable result.

Both may matter.

6.2 Cost-to-result

Minimum direct cost should include:

• labour;

• AI / model usage;

• software;

• specialist support;

• data-processing costs;

• incremental infrastructure; and

• material review or quality-assurance cost.

For early experiments, report:

PILOT MARGINAL COST

What the test itself cost.

ESTIMATED OPERATIONAL COST

Expected cost if the configuration becomes routine.

FULL LIFECYCLE COST

Total cost including integration, maintenance, security, training, monitoring, exception handling, vendor dependence and replacement risk.

These should not be conflated.

6.3 Capacity

Capacity measures what the constrained resource can now accomplish.

Examples:

• cases per staff member;

• projects per engineer;

• estimates per estimator;

• inspections prepared per analyst;

• designs evaluated per week;

• clients served with a fixed team.

Capacity should be measured at the actual constrained resource.

If AI increases analyst throughput but managerial approval remains fixed, analyst capacity increased while realized system capacity may not have. The measurement task is then to estimate whether managerial approval is a legitimate, irreducible constraint or an economically addressable bottleneck whose redesign or added capacity could convert the local gain into realized system value.

6.4 System throughput

Where bottleneck migration is plausible, track three distinct results:

LOCAL THROUGHPUT

Output of the amplified stage or constrained resource.

REALIZED SYSTEM THROUGHPUT

Completed value delivered by the whole process under the current configuration.

AMPLIFICATION POTENTIAL

Additional system value plausibly unlockable after a feasible, costed response to the newly binding constraint.

This distinction prevents weak current end-to-end throughput from erasing a validated material and disproportionate local quality-adjusted productivity or capability gain. Amplification potential must not be reported as realized value.

6.5 Capacity reinvestment and quality under constrained throughput

A binding downstream constraint can change how upstream amplification is expressed. If additional quantity cannot be absorbed, an amplified worker or team may use some released capacity to improve each unit rather than produce more units. Possible uses include deeper testing, additional design alternatives, iterative refinement, stronger verification, robustness, completeness, safety work, experimentation, learning, or reduced cognitive pressure. This is a testable hypothesis, not an assumed benefit.

Record, where material:

• unconstrained local throughput potential;

• actual output volume;

• downstream capacity, queue or utilization;

• released time or effort per completed unit;

• how released capacity was used: more volume, quality, testing or verification, alternatives or design exploration, experimentation, learning or skill development, backlog reduction, idle or unallocated time, or other;

• objective or independent evidence of any quality change;

• whether quality gains produce economic, user or public value; and

• whether extra refinement creates gold-plating, overengineering, added complexity or downstream burden.

CAPACITY REINVESTMENT

Observable redeployment of AI-released capacity into valuable non-volume uses.

QUALITY REINVESTMENT

A subset of capacity reinvestment in which released time improves the quality, robustness, completeness, safety or validation of each unit.

The program should distinguish a direct AI quality effect from a quality improvement enabled by released human time. Flat output volume under a binding downstream constraint should therefore not be interpreted as absence of amplification without examining quality and capacity allocation.

st, risk, demand and human consequences.

7. D — Human Return

Human return should be measured only where workers or participants are materially affected, but it should not be treated as optional simply because the primary experiment is economic.

The minimum human-return set should combine observable indicators with short structured worker reporting.

7.1 Workload and effort

Measure where feasible:

• hours spent on the workflow;

• overtime;

• number of handoffs;

• attention switching;

• queue pressure;

• exception load; and

• perceived workload; and

• perceived time pressure and attention available per completed unit.

A reduction in routine work that creates constant exception handling may not improve the human experience.

7.2 Drudgery

Before and after, estimate the share of work experienced as:

• repetitive;

• frustrating;

• administratively burdensome;

• physically unpleasant; or

• low-value relative to the worker's expertise.

Use short repeated measures rather than one retrospective question at the end.

7.3 Agency and autonomy

Ask whether the configuration changes:

• control over how work is performed;

• ability to challenge AI output;

• discretion;

• decision authority;

• visibility into how the result was produced; and

• ability to opt out or escalate.

This is especially important in AI-led workflows.

7.4 Skill and understanding

Record whether the configuration:

• develops new skill;

• preserves important existing skill;

• shifts skill toward verification or orchestration;

• reduces practice of an essential skill; or

• creates dependence on a tool without adequate understanding.

Where skill retention matters for safety or resilience, use an observable task or scenario test where possible rather than survey response alone.

7.5 Intervention capacity

For systems where humans may need to intervene, measure whether the responsible person can:

• recognize failure;

• understand enough context to diagnose it;

• take control;

• correct the result; and

• recover the process.

A human technically “in the loop” but unable to intervene effectively should not count as meaningful oversight.

7.6 Workforce distribution

Where material, track:

• roles created;

• roles eliminated;

• roles redeployed;

• entry-level opportunities;

• training requirements;

• promotion pathways;

• compensation changes; and

• who captures the productivity gain.

The first pilot may not be long enough to establish durable labour-market effects, but it should record leading indicators.

8. E — Strategic and Learning Return

8.1 Threshold crossing

A threshold-crossing claim requires evidence that AI changed a decision from:

not economically or operationally viable

to

viable enough to proceed.

Required evidence:

• credible description of the previous constraint;

• conventional alternative or historical cost where available;

• AI-enabled full lifecycle estimate;

• quality comparability;

• implementation burden; and

• decision outcome.

Threshold status:

NOT TESTED

The experiment did not address economic viability.

POSSIBLE

Prototype economics suggest crossing, but lifecycle evidence is incomplete.

VALIDATED

The full-lifecycle case is credibly positive relative to the previous alternative.

REALIZED

The organization actually proceeds because the economics changed.

The strongest evidence is observed changed behaviour, not hypothetical ROI.

8.2 Option creation

An option is a newly viable opportunity created by the capability gain.

Examples:

• a new product;

• a new service;

• a previously infeasible client project;

• internalization of previously outsourced capability;

• entry into a new market;

• a follow-on investment;

• a reusable software or process asset;

• a new research or engineering path.

Classify options by maturity:

IDENTIFIED

Plausible new opportunity.

VALIDATED

Basic business or operational case established.

COMMITTED

Resources formally allocated.

REALIZED

The option produces operating, commercial or public value.

Do not count speculative brainstorming as realized option creation.

8.3 Bottleneck movement

Every successful experiment should answer:

What became the binding constraint after the intervention?

Record:

• previous bottleneck;

• new bottleneck;

• evidence that the constraint moved;

• whether the new constraint is legitimate or avoidable;

• estimated value of relieving it;

• resources and cost required to relieve it;

• expected realized system throughput or value after relief;

• risks, quality constraints and human consequences of the response; and

• proposed response.

Where no bottleneck movement is observed, record that as well.

Bottleneck categories should include not only technical and physical constraints but also adoption and behavioural friction: workflow compliance, switching cost, trust, training burden, integration habits and incentives can become the true constraint even when the technology works.

8.4 Transferability

Transferability should be treated as a ladder rather than a yes/no judgment.

T0 — ORIGINAL CASE ONLY

Works for the original worker/team/task.

T1 — SAME PERSON, NEW TASK

The original practitioner reproduces the result in another relevant case.

T2 — SECOND USER, SAME ORGANIZATION

Another qualified person reproduces the practice.

T3 — SECOND ORGANIZATION, SAME SECTOR

The method works under different local conditions.

T4 — CROSS-CONTEXT

The underlying method transfers beyond the originating sector or setting.

No experiment should be described as broadly transferable at T0. Report the transferability level separately from the assistance needed to achieve it. For each transfer attempt, record the recipient’s starting expertise, data, digital capability and implementation authority; adaptations made; who supplied coaching or technical support; and recipient and provider time and cost. Distinguish independent reuse from assisted transfer and report unsuccessful attempts. Include continuing support and maintenance in the net-value assessment: a T3 result achieved with intensive coaching is not evidence that low-cost dissemination will work.

8.5 Additionality

Additionality asks what public participation changed.

Evidence can include:

• participant cost-sharing decisions;

• documented prior deferral;

• counterfactual interview before funding;

• timing difference;

• increased experimental scope;

• added measurement rigor;

• cross-firm coordination;

• reusable learning obligations;

• access to expertise otherwise unavailable; or

• public-sector value.

Additionality status:

LOW

The activity likely would have occurred similarly without the program.

MODERATE

The program materially changed timing, scope, quality or measurement.

HIGH

The experiment or transferable learning was unlikely to occur without the intervention.

Additionality should be assessed before and after the experiment because participant expectations can be wrong.

8.6 Strategic capability

Where relevant, record whether the experiment creates:

• durable Canadian expertise;

• domestic technical or operational capability;

• reduced reliance on scarce external suppliers;

• reusable IP or know-how;

• export capability;

• institutional learning; or

• resilience to supply, labour or market disruption.

These are descriptive outcomes, not automatic reasons to scale.

9. F — Evidence Confidence

Evidence confidence should be assessed separately for each major claim.

Four dimensions matter most.

9.1 Baseline quality

STRONG

Reliable existing data or well-designed prospective baseline.

MODERATE

Useful baseline with some missing data or measurement noise.

LIMITED

Reconstructed baseline with material uncertainty.

9.2 Comparison strength

STRONG

Parallel control, randomized allocation, credible matched comparison, repeated within-subject crossover, or another design that materially limits alternative explanations.

MODERATE

Before/after comparison with stable conditions and reasonable controls for major confounders.

LIMITED

Uncontrolled before/after observation or retrospective comparison.

9.3 Outcome objectivity

STRONG

Objective operational, financial, quality, service or independently scored result.

MODERATE

Structured expert or user judgment with defined rubric.

LIMITED

Self-report, anecdote or unstructured impression.

9.4 Confounding / attribution risk

LOW

Few credible alternative explanations.

MODERATE

Some material alternative explanations remain.

HIGH

Novelty, selection effects, simultaneous process change, demand shifts, staffing changes or other factors make attribution uncertain.

9.5 Overall confidence label

HIGH

Strong or moderate baseline, credible comparison, reasonably objective outcomes and low-to-moderate confounding.

MEDIUM

Useful evidence but at least one important limitation remains.

LOW

Promising signal whose apparent effect cannot yet be separated from major measurement or attribution uncertainty.

The evaluator should record the reason for the label in one sentence.

No arithmetic confidence score is required.

10. Experiment Result Profile

Every stage-gate review should use the same compact result profile.

EXPERIMENT

Name / participant / cohort.

PRIMARY OUTCOME

What was the pre-specified main measure?

MWE

What minimum effect was required?

OBSERVED EFFECT

Magnitude and direction.

QUALITY

What happened to quality?

LOCAL AMPLIFICATION

What changed at the measured stage or constrained resource? State the outcome type, baseline and comparator, workflow boundary, observation period, quality standard, absolute gain and any meaningful quality-adjusted productivity multiple; show the human, technical and organizational resources used and whether the separate amplification magnitude criterion was met.

REALIZED SYSTEM EFFECT

What whole-system value is captured under the current configuration after verification, rework and downstream constraints?

AMPLIFICATION POTENTIAL

If a new bottleneck is binding, what additional system value could plausibly be unlocked, what supporting resources would be required, and at what estimated cost and risk?

COST

Pilot marginal / estimated operational / lifecycle.

CONFIGURATION

Human / hybrid / AI allocation by major function.

HUMAN RETURN

Material positive, negative or unresolved effects.

THRESHOLD CROSSING

Not tested / possible / validated / realized.

OPTION CREATION

Identified / validated / committed / realized.

RELEASED-CAPACITY USE / QUALITY REINVESTMENT

If throughput was constrained, how was released time or effort actually used? Did quantity, quality, testing, experimentation, learning or idle time change, and what evidence shows that the change created value rather than gold-plating?

BOTTLENECK

What became scarce next?

TRANSFERABILITY

T0 / T1 / T2 / T3 / T4; independent or assisted transfer; recipient readiness; adaptation, coaching and continuing-support time and cost; outcome after those costs.

ADDITIONALITY

Low / Moderate / High.

EVIDENCE CONFIDENCE

High / Medium / Low, with reason.

OPERATIONAL DECISION

STOP / REDESIGN / REPLICATE / SCALE-HANDOFF.

KNOWLEDGE DISPOSITION

DIFFUSE / DO NOT DIFFUSE / DIFFUSE WITH RESTRICTIONS.

STOP REASON CODE

Mechanism / Performance-Quality / Lifecycle Economics / Additionality / Risk-Governance / Adequate Existing Solution / Insufficient-Infeasible Evidence / Other.

NEXT EVIDENCE REQUIRED

What must be learned before the next commitment?

11. Decision thresholds

The framework should avoid universal numerical thresholds because sectors differ.

Instead, every experiment uses three threshold types.

11.1 Pre-specified MWE threshold

The primary outcome must meet or exceed the Minimum Worthwhile Effect unless there is a documented reason to redesign rather than stop.

11.2 Non-negotiable guardrails

Regardless of economic effect, do not advance when:

• unacceptable safety risk remains;

• material privacy or security risk is uncontrolled;

• quality falls below the minimum acceptable standard;

• legal or regulatory requirements cannot be met;

• human oversight is required but ineffective; or

• the result depends on concealed or unverifiable assumptions; or

• a consequence-specific quality guardrail is violated, such as required source provenance, safety-check completion, statutory compliance or another binary control that should not be averaged into a quality score.

11.3 Evidence threshold by decision

STOP

May occur at any confidence level when the mechanism, quality, economics or safety is clearly inadequate.

REDESIGN

Appropriate when evidence identifies a fixable mechanism or bottleneck.

REPLICATE

Preferred when effect exceeds MWE but evidence confidence is Medium/Low, transferability is T0/T1, or the result is unusually large.

SCALE-HANDOFF

Normally requires:

• effect meets MWE;

• acceptable quality;

• either a positive realized system gain or a credible, costed constraint-relief case showing that a validated local capability gain can be converted into positive system value;

• positive expected lifecycle economics;

• controlled risk;

• High or sufficiently strong Medium evidence for the consequence level;

• identified and economically credible bottleneck response where current system absorption is limiting value;

• organizational capacity, or a feasible plan to add the supporting capacity, required to absorb the change.

KNOWLEDGE DISPOSITION — DIFFUSE

May accompany any operational disposition. Normally requires:

• replicated result;

• transferability at least T2, and preferably T3 for broad sector claims;

• documented prerequisites;

• reproducible implementation;

• known failure modes; and

• public or sector value from sharing.

12. Frontier Result Measurement

Frontier Worker or Frontier Capability claims require a stricter evidence pattern than ordinary improvement claims.

A frontier signal should show:

• unusually large quality-adjusted improvement relative to the person's or team's prior baseline, disproportionate to added human, technical and organizational resources; efficiency or capacity gains can qualify without a capability or creation effect;

• evidence that the result is not simply lower quality at higher speed;

• repeatability across more than one task or period;

• identifiable role of human judgment, AI capability and workflow design;

• comparison with peers, prior practice or another credible reference where feasible;

• evidence confidence sufficient to justify additional resources;

• specialist/support intensity, including expert hours, privileged access, infrastructure and program assistance consumed; and

• evidence that the outlier is not simply a hidden team result attributed to one person;

• the gap between local amplification and currently realized system value; and

• a costed hypothesis for what downstream support or constraint relief would be required to convert the local gain into larger system value.

The program should distinguish:

OUTLIER RESULT

An unusual performance observation.

VALIDATED FRONTIER CASE

The unusual result survives quality and replication scrutiny.

FRONTIER PRACTICE

A method or configuration that can be described and tested.

TRANSFERABLE FRONTIER CAPABILITY

A practice or capability reproduced beyond the original case.

The program should not infer a stable worker trait from a single outlier result.

13. Portfolio-Level Measurement

Individual experiment metrics should feed a smaller set of portfolio measures.

13.1 Discovery efficiency

• opportunities screened;

• cost per opportunity screened;

• time from signal to disposition;

• percentage advanced to Opportunity Clinic;

• Frontier Signals generated;

• proportion of signals later validated.

13.2 Experimental discipline

• experiments with credible baselines;

• experiments with pre-specified MWE;

• experiments stopped early;

• experiments redesigned;

• average time to first decision gate;

• measurement burden per participant;

• specialist-review hours per experiment.

A high STOP rate is not inherently bad.

13.3 Portfolio value

• total public and participant investment;

• measured economic / public-service value;

• distribution of returns;

• share of measurable value generated by top 10% of experiments;

• validated threshold crossings;

• realized options;

• commercialization or operational scale outcomes;

• value created after full lifecycle cost where measurable.

Portfolio value should be reported with attribution caveats.

13.4 Replication and diffusion

• results entering replication;

• replication success rate;

• T2 / T3 / T4 transferability outcomes;

• reusable playbooks or modules;

• organizations adopting validated practices;

• evidence of actual reuse, with recipient readiness, assistance required and net value after transfer costs, not only publication of a playbook;

• time from discovery to replication;

• time from replication to routine use.

13.5 Human-return portfolio measures

• drudgery reduced;

• worker capability developed;

• material negative workload effects;

• skill-retention concerns;

• roles redeployed / displaced where observable;

• compensation or gain-sharing where material;

• human-return outcomes by cohort and sector.

13.6 Capacity reinvestment and quality

• experiments with material released capacity;

• how released capacity was allocated among volume, quality, testing, experimentation, learning, backlog reduction and idle time;

• independently measured quality changes where throughput remained constrained;

• cases in which reinvestment created valuable robustness, design improvement or learning; and

• cases in which reinvestment produced gold-plating, complexity or new downstream burden;

• the relationship among local amplification intensity, downstream absorption constraints and measured quality change; and

• whether larger absorption gaps are associated with more quality, testing or experimentation reinvestment. Treat this portfolio relationship as exploratory unless the experiment design supports causal inference.

13.7 Strategic learning

• bottlenecks identified;

• bottlenecks successfully relieved;

• validated local amplification cases whose system value is currently constrained;

• costed amplification-potential cases;

• proportion of bottleneck-response experiments that convert local gains into realized system gains;

• durable domestic capabilities created;

• public-program additionality distribution;

• option-creation rate;

• realized options;

• recurring sector patterns;

• experiments whose learning changed future portfolio selection.

14. Anti-Goodhart safeguards

Measurement can distort behaviour.

The framework should explicitly protect against participants optimizing for visible metrics rather than real value.

Safeguards:

• use multiple evidence dimensions rather than a single score;

• pre-specify primary outcomes;

• retain qualitative diagnostic review;

• rotate or refine metrics when gaming becomes visible;

• use independent quality assessment for high-value claims;

• audit a sample of reported results;

• distinguish participant-reported from evaluator-verified evidence;

• do not reward experiment count or AI usage volume;

• do not reward “success rate”;

• allow negative findings to count as valuable learning;

• separate funding decisions from automated scoring.

The program should monitor whether measurement itself changes participant behaviour in harmful ways.

15. Missing-data discipline

Small organizations will not always produce clean evidence.

The program should never fill missing data with false precision.

When evidence is unavailable:

• mark it unresolved;

• state why;

• reduce confidence accordingly;

• identify the cheapest next measurement that would change the decision.

“Unknown” is an acceptable result.

16. Measurement burden budget

The pilot should track measurement cost as a first-class variable.

AI-assisted measurement should itself be treated as an experimental configuration. Intake structuring, evidence extraction, check-in summarization, contradiction/provenance checks and draft evidence packets may reduce evaluator burden, but the human-execution baseline remains the reference case for staffing and costing. Time or cost savings should be credited only after comparison shows that evidence quality, provenance, escalation sensitivity and accountable human judgment are preserved.

For each experiment record:

• participant reporting hours;

• evaluator hours;

• specialist review hours;

• data-instrumentation cost; and

• time added to the workflow by measurement.

A useful portfolio metric is:

Measurement cost per decision-relevant finding.

The program should simplify or remove measures that rarely change stage-gate decisions.

17. Measurement by cohort

17.1 Fast-Feedback Digital

Likely emphasis:

• time-to-result;

• accepted output;

• defect / rework;

• verification burden;

• configuration iteration;

• cross-skill capability;

• replication speed.

17.2 White-Space SME

Likely emphasis:

• baseline labour / contractor cost;

• threshold crossing;

• full lifecycle economics;

• maintainability;

• security / integration burden;

• domain-expert capability;

• reusable sector pattern;

• existing-solution / adoption-friction check;

• support required to integrate and maintain the solution;

• system bottleneck.

17.3 Public / Institutional

Likely emphasis:

• service time;

• backlog;

• quality / compliance;

• public or internal user outcome;

• approval latency;

• staff workload;

• accountability;

• intervention capacity;

• procurement / governance bottlenecks.

The core framework remains common; the detailed metrics differ.

18. Measurement governance

18.1 Participant

Provides operational evidence and context but does not self-certify major success claims.

18.2 Measurement Lead

Owns:

• metric definition;

• baseline quality;

• evidence confidence;

• result interpretation;

• measurement burden.

18.3 Portfolio Lead

Owns the decision, informed by but separate from measurement.

18.4 Independent review

Use independent expert or evaluator review when:

• the claimed effect is unusually large;

• public funding escalation is material;

• the outcome is difficult to score objectively;

• the result may become a public flagship case;

• safety or high consequence is involved; or

• Frontier status is being considered.

19. What the MVMF deliberately does not do

This framework does not:

• create one national amplification score;

• assign monetary value to every human outcome;

• claim that all experiments are causally comparable;

• force all sectors into identical outcome metrics;

• require randomized trials for every project;

• treat self-report as worthless;

• treat self-report as sufficient for strong causal claims;

• require every experiment to show economic return;

• assume strategic value overrides weak economics;

• assume human-return benefits excuse poor quality;

• assume every local gain is transferable.

Its purpose is disciplined decision-making under heterogeneous evidence.

20. Initial red-team test

The strongest challenge is that the framework may still be too elaborate for small organizations and may create a professional evaluation bureaucracy that consumes the value it is meant to measure.

The framework should therefore be tested against five questions:

1. Can a five-person SME complete the required core measurement without specialist staff?

2. Can an evaluator understand the experiment from the Result Profile in under ten minutes?

3. Does each required metric change a plausible decision?

4. Can the framework distinguish a genuine 2x capability gain from a superficially impressive demo?

5. When downstream throughput is capped, can it distinguish valuable quality or experimentation reinvestment from idle time, gold-plating or hidden downstream burden?

If the answer to the first three is no, simplify.

If the answer to the fourth is no, strengthen verification.

21. Immediate next work

The MVMF is sufficiently defined to support the next two tasks.

Next:

1. COMPLETED — Stage 0 Signal Form, Opportunity Note, Experiment Charter and one-page Experiment Result Profile created in NL Pilot Forms & Templates v0.1: https://docs.google.com/document/d/1FfINsd2I_nMBZIZLGr0m6ejTu_o8nOS-mTEk5jnRu1U/edit

2. COMPLETED — Operating protocol and MVMF stress-tested across four hypothetical cases in Sector Stress Tests v0.1: https://docs.google.com/document/d/1MfsQiCWI2IRi5ICcil9ZYs1EGvQW2TQ3IGa_vpMKlP0/edit

3. COMPLETED PROVISIONALLY — The operational-capacity model now converts the stress-test burden estimates into peak weekly staffing demand across Low / Base / High cases. Under Base assumptions, 36 active experiments are supportable with three rolling waves and a practical core team of approximately 5 FTE plus an on-demand specialist / independent-review pool; the High-burden case requires more capacity or less concurrency.

4. NEXT — Convert the capacity envelope into pilot cost and funding architecture, including core staffing, specialist review, participant support and staged escalation resources. Preserve a Wave-1 operating-capacity gate so later-wave concurrency can be adjusted using observed burden rather than the planning assumptions.

5. Define funding stages and co-contribution only after the evidence burden is understood.

6. Specify privacy / security / risk tiers.

7. Update the Red-Team Register with measurement-specific failure modes that emerge from the sector tests.

Working measurement principle

Measure enough to know whether the next dollar, hour and unit of institutional attention should be committed—but no more than the decision requires.

22. Canonical outcomes and finding efficiency

Every experiment should classify its primary quality-adjusted outcome as Type I Efficiency, Type II Capacity, Type III Capability or Type IV Creation. Configuration is recorded separately on the human-only ↔ hybrid ↔ AI-only continuum. Amplification is an evidence-dependent magnitude designation, not a fifth outcome type: a material and disproportionate shift in the quality-adjusted production frontier relative to added human, technical and organizational resources. It can apply to any of the four outcome types. Multiple gains in quality-adjusted productivity can qualify through efficiency or capacity alone; capability and creation effects are not prerequisites, and novelty alone is insufficient. Report combined effects without counting the same benefit twice.

For each amplification claim, state the baseline and comparator, workflow boundary, observation period, quality standard, absolute change and, where meaningful, the quality-adjusted productivity multiple. Specify the resource basis and include setup, implementation, oversight, verification, rework, compute, integration, maintenance and other material lifecycle costs. Pre-specify a context-appropriate criterion for a material and disproportionate gain; report uncertainty, repeatability within the original setting and replication by other workers or organizations separately. Where a new capability has no meaningful numeric baseline, use explicit feasibility, value and resource comparisons rather than an arbitrary multiplier.

Identify whether the demonstrated gain belongs to an individual, team or workflow, or organization or service. Local amplification, realized system value and costed amplification potential remain separate claims. A verified local gain can qualify at its measured boundary even when a downstream constraint limits organizational capture; work shifted to others must be counted, and a gain erased across the relevant workflow is not realized amplification at that wider boundary.

A credible, decision-relevant finding is a documented conclusion capable of changing or confirming an operational or policy decision and supported by a defensible baseline or counterfactual, quality adjustment, explicit limitations and review proportionate to the claim. Count distinct decision claims, not projects, reports, outputs or positive results. A replication normally strengthens an existing finding unless it answers a separate pre-specified question. STOP and REDESIGN conclusions count when they meet the same standard.

Report gross program cost divided by the number of accepted findings, with the numerator, denominator and excluded claims visible. Illustratively, $3.68 million divided by 18, 24 or 30 findings equals approximately $204,000, $153,000 or $123,000 per finding; $2.25 million divided by 8, 12 or 16 equals approximately $281,000, $188,000 or $141,000. These are sensitivity examples, not promised yields, and must be interpreted with claim importance, risk, replication strength and participant burden.

Document control and disclosures

Package 0.5 Discussion Draft. Last reviewed 20 September 2026. Next review: before external circulation and after any material host, funding, governance, scope, budget or capacity decision.

Authorship: Terry Hudson is the originator and principal author. This document was developed in a personal capacity. The author is employed by the Government of Newfoundland and Labrador; that employment is disclosed for transparency. The document was not commissioned, approved or endorsed by the Government of Newfoundland and Labrador, and the views expressed are the author’s own. No institutional partner or co-signatory is represented by this draft. Contact: terry@hed1.ca.

AI-use disclosure: This document was developed with substantial AI assistance under the author’s direction. The project’s originating thesis, problem definition and core hypotheses were developed by the author from professional experience, observation and reasoning. AI supported research and evidence discovery, structured challenge, elaboration and organization of ideas, comparison of alternatives, synthesis, drafting and revision. The author directed the inquiry, supplied the institutional context, judged which outputs to retain or reject, and made the substantive policy and design decisions. The author remains responsible for source verification, accuracy, judgment and final content. The development process informed hypotheses about human–AI work, but it is not evidence that those hypotheses are true.

Conflict of interest: [AUTHOR DECISION: insert the author’s and any affiliate’s relevant financial, commercial or institutional interests before external circulation.] No tool, vendor or proprietary product is endorsed, preselected or given preferential access.