Value Attribution and Outcome Measurement

Draft 0.1 — 2 August 2026
Measurement supports decisions and reward allocation. It does not guarantee scientific certainty, future performance, or an independently audited financial result.

Objective

ANUKA should connect contribution to outcome more rigorously than ordinary feedback tools and more honestly than marketing case studies.

The system must answer:

what changed;
who contributed;
what evidence existed before the change;
which users or systems were affected;
what happened afterward;
which alternative explanations remain;
what reward rule applies;
what can be claimed publicly.

Separate the claims

The protocol distinguishes:

Activity claim

A participant performed an action.

Example: Contributor submitted onboarding copy.

Acceptance claim

The product accepted the contribution.

Release claim

The contribution reached a declared environment or audience.

Association claim

A metric changed after or alongside the release.

Causal claim

The contribution caused some or all of the observed change under a declared method.

Financial-value claim

The result created revenue, savings, avoided loss, or another monetized effect.

A release plus a rising metric does not automatically justify a causal or financial claim.

Measurement plan

Every outcome-dependent opportunity MUST define a measurement plan before release.

Required fields:

hypothesis;
intervention;
eligible population;
assignment method;
baseline;
primary metric;
guardrail metrics;
measurement window;
minimum sample or evidence threshold;
data sources;
known confounders;
analysis method;
stopping rules;
rollback conditions;
attribution rule;
reward formula;
public-claim policy.

Evidence strength levels

E0 — Narrative

A reasoned account without source-connected quantitative evidence.

E1 — Before/after observation

The metric changed across a time boundary, but alternative explanations are not controlled.

E2 — Matched or segmented comparison

The result is compared against a relevant cohort, historical pattern, matched segment, or synthetic baseline.

E3 — Controlled experiment

A/B or other controlled assignment supports a stronger causal inference.

E4 — Replicated or independently reviewed result

The effect was reproduced, externally reviewed, or observed across multiple relevant contexts.

The evidence level must be displayed with the result.

Evidence indicator

Evidence-strength ladder

Higher levels strengthen the inference; they do not erase uncertainty, confounders or the need for clear public disclosure.

E0
Narrative
A reasoned account without source-connected quantitative evidence.
E1
Before / after
A change across time, with alternative explanations still open.
E2
Matched comparison
A relevant cohort, segment, history or synthetic baseline informs the result.
E3
Controlled experiment
Controlled assignment supports a stronger causal inference.
E4
Replicated or reviewed
The effect is reproduced, independently reviewed or observed in multiple relevant contexts.

Experiment types

Randomized A/B test

Preferred where technically, ethically, and commercially appropriate.

Switchback test

Useful for time-dependent systems where treatment alternates across periods.

Phased rollout

Compares cohorts introduced at different times.

Holdout

Maintains an untreated group for evaluation.

Interrupted time series

Evaluates change against pre-intervention trend.

Matched cohort

Compares similar users or accounts when randomization is unavailable.

Qualitative evaluation

Uses structured interviews, usability tests, or review panels for outcomes not captured adequately by one metric.

Operational verification

Confirms completion, performance, reliability, security, or compliance criteria without claiming commercial causality.

Primary and guardrail metrics

Each experiment has one declared primary metric unless a justified multi-objective design exists.

Guardrails detect harm such as:

fraud;
refunds;
latency;
accessibility regression;
support load;
churn;
complaint rate;
security incidents;
margin erosion;
concentration risk.

A primary win with a material guardrail failure is not a successful loop.

Metric contracts

Every metric SHOULD have a machine-readable contract:

identifier;
human definition;
owner;
formula;
source system;
event schema;
inclusion and exclusion rules;
time zone;
currency;
update frequency;
allowed transformations;
quality checks;
privacy class;
version;
deprecation policy.

Changing a metric definition during an experiment requires a new version and disclosure.

Baselines

Possible baselines include:

prior-period value;
same-period seasonal comparison;
control cohort;
expected trend;
contract or service-level threshold;
industry reference;
model prediction.

The baseline source and limitations must be visible.

Attribution models

Direct completion

Reward is based on accepted delivery, not downstream business outcome.

Binary outcome

A reward is earned when a defined threshold is met.

A reward is calculated from estimated incremental value attributable to the intervention.

Contribution points

A declared rule allocates credit among idea, implementation, review, rollout, and measurement roles.

Panel attribution

A reviewer panel assigns bounded credit using evidence and disclosed criteria.

Hybrid

A fixed delivery reward is combined with a capped outcome bonus.

The MVP SHOULD prefer fixed or capped models over uncapped revenue participation.

Attribution comparison

Attribution model comparison

Choose a model whose evidence burden matches the claim and reward risk. The MVP defaults toward fixed or capped approaches.

Direct completion

Basis: Accepted delivery
Strength: Simple and auditable
Caution: Does not establish downstream impact

Binary outcome

Basis: Defined threshold
Strength: Clear trigger
Caution: Threshold design can create cliff incentives

Incremental value share

Basis: Estimated incremental value
Strength: Connects work to outcomes
Caution: Needs strong causal evidence and caps

Contribution points

Basis: Declared multi-role allocation
Strength: Makes team roles visible
Caution: Points do not prove company-level causation

Panel attribution

Basis: Evidence-based review
Strength: Handles complex context
Caution: Requires conflict controls and disclosed criteria

Hybrid

Basis: Fixed reward + capped bonus
Strength: Balances delivery and outcome
Caution: Keep bonus formula and cap explicit

Multi-contributor attribution

A result may depend on:

original signal;
hypothesis author;
product owner;
designer;
engineer or AI agent;
reviewer;
tester;
rollout steward;
analyst;
existing product and brand assets.

Credit allocation must not imply that one participant created the entire company-level effect.

Recommended fields:

role;
contribution record;
declared weight or formula;
reviewer adjustments;
maximum share;
conflict disclosure;
evidence links.

Financial value

Possible value categories:

Incremental revenue

Additional revenue reasonably attributable to the intervention.

Gross-margin contribution

Revenue adjusted for direct variable cost.

Cost savings

Reduction in spend or labor under a documented baseline.

Avoided loss

Estimated prevention of churn, fraud, downtime, penalties, or rework.

Capital efficiency

Improvement in conversion, payback, utilization, or runway.

Option value

Learning that enables or rejects a future investment.

Financial claims should state whether values are:

observed;
estimated;
modeled;
annualized;
gross or net;
cash or accounting;
independently reviewed.

Reward formulas

Reward model reference

How common reward models behave

Fixed plus bonus

Delivery fee + verified outcome bonusRewards reliable delivery and a separately evidenced result.

Capped incremental share

Lower of the cap or value × ratePrevents an outcome share from exceeding the agreed maximum.

Multi-role pool

Pool × role weight × evidence adjustmentAllocates one pool according to contribution role and evidence quality.

Threshold ladder

Below threshold 1Bonus 0
Threshold 1 to threshold 2Amount 1
Threshold 2 and aboveAmount 2

Set currency, timing, caps, rounding, reversals, taxes, and dispute rules alongside every model.

Every formula must specify currency, timing, caps, rounding, reversals, taxes, and dispute rules.

Negative and neutral results

ANUKA must preserve failed experiments.

Possible outcomes:

positive;
neutral;
negative;
inconclusive;
invalid measurement;
stopped for safety;
not enough data;
external event overwhelmed signal.

A contributor may still earn delivery and learning rewards when the hypothesis fails honestly.

Punishing all negative findings creates pressure to manipulate results.

External events

The plan must address:

pricing changes;
marketing campaigns;
outages;
seasonality;
competitor actions;
major customer wins or losses;
policy changes;
data pipeline changes;
simultaneous experiments.

Material events are added to the evidence package.

Data quality

Before attribution, the system should check:

event completeness;
duplicate events;
schema version;
source outages;
bot and fraud traffic;
cohort leakage;
assignment integrity;
missing values;
currency conversion;
late-arriving data;
metric-definition drift.

AI analysis

AI may:

draft hypotheses;
detect anomalies;
summarize evidence;
propose segments;
identify confounders;
generate reproducible analysis code;
explain results.

AI MUST NOT silently:

choose a favorable metric after results are known;
remove inconvenient cohorts;
claim causality from correlation;
fabricate missing data;
publish claims without approval;
calculate payments using an undisclosed model.

Material AI-generated analysis should include model, tools, inputs, and human review state.

Public result cards

A public improvement story SHOULD include:

product;
change;
contributors;
measurement window;
affected cohort;
metric definition;
baseline;
observed result;
evidence strength;
guardrail status;
attribution limitation;
verifier;
correction or revocation status.

Avoid headlines such as ANUKA increased revenue by 42% when the evidence supports only a before/after association.

Investor reporting

Traction reports may include verified source data and experiment results, but must distinguish:

company-reported;
source-connected;
reviewed;
attested;
audited;
estimated;
forecast.

ANUKA verification is not an audit unless a qualified independent auditor performed the audit.

Conflict of interest

Reviewers and analysts disclose:

financial interest;
reward eligibility;
product ownership;
related-party relationship;
prior involvement in the change;
model or vendor incentives.

High-value outcome bonuses SHOULD receive independent or multi-party review.

Reproducibility

An evidence package SHOULD preserve:

query or analysis version;
source snapshot references;
metric contracts;
experiment configuration;
relevant code hash;
reviewer decision;
publication version.

Sensitive raw data remains access-controlled.

MVP acceptance criteria

measurement plan before release;
primary and guardrail metrics;
evidence strength label;
metric contracts;
result states including negative and inconclusive;
capped reward formulas;
contributor role allocation;
external-event log;
public claim template;
dispute and correction;
no automatic causal language.

Sources

Verification record

Sources opened and checked: 2 August 2026
Protocol status: Founding draft
Statistical-method review required for production: Yes
Financial-claim review required: Yes