Contextual Scores and Decision Safeguards

Draft 0.1 — 2 August 2026
This document sets product safeguards. It is not a complete legal compliance program for employment, credit, housing, insurance, education, investment, or other regulated decisions.

Objective

ANUKA may need computed summaries to match people with work, prioritize companies for review, identify trustworthy evidence, or communicate product readiness.

A useful score can compress complexity. A careless score can create false certainty, discrimination, manipulation, and irreversible reputation damage.

The governing rule is:

No score exists without a named purpose, defined evidence set, limited audience, expiration, explanation, and contest path.

Universal scores are prohibited

ANUKA must not calculate or publish one score intended to summarize a person’s or company’s total social value.

The following patterns are prohibited:

Human score: 84;
Founder quality: 92 without a defined task and method;
one permanent rank across unrelated capabilities;
a score that silently follows a person into every context;
a score whose inputs or consequences cannot be inspected.

Allowed score classes

Descriptive summary

Compresses observed history without predicting future behavior.

Example:

18 accepted contributions in SaaS onboarding, 12 independently reviewed, during the past 24 months.

Readiness score

Evaluates conformance to a declared checklist for one action.

Example:

Release readiness: 7 of 9 required controls passed; security review pending.

Match score

Estimates similarity between an opportunity and a participant’s relevant evidence.

Example:

Opportunity match: high, based on four comparable activation experiments and current availability.

Confidence score

Summarizes confidence in a particular claim or evidence package.

Example:

Metric-source confidence: moderate; data is source-connected but not independently reviewed.

Risk signal

Flags a defined operational or integrity concern.

Example:

Review conflict risk: related-party relationship detected.

A risk signal is not proof of wrongdoing.

Quality conformance score

Shows how many requirements of a named ANUKA quality profile are currently satisfied.

It must identify the profile version, failed checks, exceptions, and assessment date.

Score contract

Every score implementation must publish a Score Contract containing:

score name;
purpose;
subject type;
audience;
permitted uses;
prohibited uses;
input fields and sources;
feature definitions;
weighting or model method;
missing-data behavior;
time range and expiry;
confidence and uncertainty;
material limitations;
fairness and abuse tests;
human-review rules;
appeal and correction process;
version and change history;
responsible owner.

No score may ship with an undocumented private interpretation known only to the development team.

Naming

Score names should describe the actual decision context.

Prefer:

SaaS onboarding opportunity match;
Verified metric evidence confidence;
Investor data-room completeness;
Release-readiness conformance;
Reviewer independence signal.

Avoid:

Trust score;
Talent score;
Founder score;
Social score;
Success score.

Evidence eligibility

A score may use only inputs allowed by its Score Contract.

Each input should declare:

source class;
verification state;
relevance;
freshness;
visibility;
dispute state;
whether it is self-asserted, inferred, or independently reviewed;
legal and policy restrictions.

Disputed or stale evidence must not be treated as ordinary active evidence without explicit handling.

Protected and sensitive information

High-impact decisions must not use protected traits unlawfully.

The system should identify and control:

direct protected-class fields;
obvious proxies;
location or language features that may create unlawful disparate treatment;
disability-related information;
medical or genetic data;
age data;
immigration or national-origin signals;
family or caregiving information;
union or collective activity;
private communications.

Removing a direct field does not prove that the model is free from proxies or disparate impact.

Missing data

A missing record can mean:

no experience;
private experience;
unintegrated source;
source outage;
old data;
participant choice not to disclose;
data-quality failure.

The score must not automatically interpret missing data as negative evidence.

Sample size and confidence

Scores should not display false precision.

Requirements:

show sample size;
identify the number of comparable contexts;
expose whether evidence is concentrated in one project;
downgrade confidence for sparse or stale data;
distinguish observed performance from prediction;
avoid percentage outputs when the denominator is too small to be meaningful.

Context similarity

Matching should explain similarity dimensions such as:

problem type;
product stage;
business model;
customer segment;
technology;
risk level;
metric;
role;
evidence age.

A participant successful in one context is not automatically predicted to succeed in every context.

Explainability

Every displayed score should answer:

What is being measured?
Why is it being shown?
Which evidence influenced it most?
Which evidence was excluded?
How recent is it?
What uncertainty exists?
What can the subject do to correct inaccurate data?
Who made the final decision?

Explanations must be understandable without reverse-engineering a model.

Score output shape

{
  "scoreId": "anuka:score:01J...",
  "scoreType": "opportunity-match",
  "scoreVersion": "0.1.0",
  "subject": "anuka:person:01J...",
  "opportunity": "anuka:opportunity:01J...",
  "result": {
    "band": "high",
    "numericValue": 0.78,
    "confidence": "moderate"
  },
  "topFactors": [
    {
      "factor": "comparable-reviewed-outcomes",
      "direction": "positive",
      "evidenceCount": 4
    },
    {
      "factor": "evidence-freshness",
      "direction": "negative",
      "explanation": "No comparable evidence in the last six months"
    }
  ],
  "limitations": [
    "Not an employment background report",
    "No prediction of long-term job performance"
  ],
  "generatedAt": "2026-08-02T00:00:00Z",
  "expiresAt": "2026-08-09T00:00:00Z"
}

If a numeric value is not necessary, the API should return an evidence summary or band instead.

Decision tiers

Tier 0 — Informational

Examples:

sorting public content;
suggesting documentation;
showing a participant their own progress.

Safeguards: explanation, correction, privacy, and monitoring.

Tier 1 — Reversible opportunity discovery

Examples:

recommending bounties;
prioritizing optional outreach;
suggesting reviewers.

Safeguards: multiple discovery paths, no permanent exclusion, participant controls, bias monitoring.

Tier 2 — Material but reviewable

Examples:

shortlist for a paid project;
access to a permissioned data room;
higher bounty limit;
reviewer eligibility.

Safeguards: human review, documented criteria, appeal, identity assurance, audit log, time-limited decision.

Tier 3 — High impact or regulated

Examples:

employment hiring, firing, promotion, retention, or reassignment;
credit, insurance, housing, education admission, or regulated financing;
securities or investment recommendations;
denial of essential network rights;
fraud enforcement with severe consequences.

Tier 3 requires a dedicated legal basis, compliance program, independent review, adverse-action procedures where applicable, and human accountability. The general ANUKA scoring service must not enable Tier 3 by default.

Human review

Human involvement must be meaningful.

A reviewer should be able to:

inspect evidence;
understand the score method;
disagree with the model;
record rationale;
identify conflicts;
request additional information;
avoid relying on prohibited fields;
communicate decisions and appeal rights.

A rubber-stamp click is not meaningful human review.

Employment boundary

The CFPB states that third-party dossiers and algorithmic scores about workers used for hiring, promotion, reassignment, or retention are often governed by the Fair Credit Reporting Act.

FTC guidance explains that companies providing employment background reports may be consumer reporting agencies even if they use another label. Requirements may include permissible purpose, authorization, accuracy procedures, disclosures, and adverse-action processes.

EEOC guidance requires employers and employment agencies to avoid discriminatory use of background information and algorithmic tools.

Therefore ANUKA must not offer a generic hireability score or sell participant dossiers for employment decisions without a purpose-built, counsel-reviewed compliance program.

Company and investment scores

Company readiness summaries can also mislead.

An investor-readiness score must not imply:

guaranteed growth;
investment endorsement;
valuation;
creditworthiness;
securities suitability;
independent audit;
absence of fraud;
completeness beyond the reviewed evidence.

Prefer a checklist and evidence matrix:

corporate records present;
metrics source-connected;
cohort definitions documented;
contracts uploaded;
cap table reviewed;
security review current;
open risks listed.

Change management

A score version change must document:

reason;
changed inputs;
changed weights or model;
expected impact;
validation results;
fairness review;
migration date;
whether historical scores are recomputed;
rollback plan.

Material changes must not silently alter participant opportunity access.

Monitoring

Score owners should monitor:

output distribution;
missing-data patterns;
selection and rejection rates;
subgroup disparities where lawful and appropriately protected;
false positives and false negatives;
appeal and overturn rates;
source outages;
model drift;
manipulation attempts;
concentration effects where high-score participants receive all future evidence opportunities.

Feedback loops

Reputation systems can create self-fulfilling inequality: early winners receive more opportunities and therefore more evidence.

Mitigations include:

entry-level opportunities;
exploration quotas;
random or rotating exposure;
context-specific matching;
visible alternative qualification routes;
support for simulated and supervised practice;
separating lack of history from negative history;
monitoring opportunity concentration.

Participant rights

For material scores, a subject should be able to:

know that a score exists;
see its purpose and audience;
inspect material inputs;
correct inaccurate data;
dispute attribution or interpretation;
request human review;
know the expiration date;
see material changes in the method;
export relevant evidence;
restrict optional public display.

These product rights do not replace stronger rights under applicable law.

Agent-generated scoring

AI may help summarize evidence, but:

the model and prompt version must be logged;
the output must be labeled as generated or assisted;
source evidence must remain inspectable;
hallucinated facts must not enter the graph as verified evidence;
model confidence is not factual confidence;
sensitive decisions require deterministic checks and human accountability;
generated explanations must be verified against actual inputs.

Prohibited scoring patterns

ANUKA must not:

create a universal person or company score;
conceal material inputs;
use undisclosed browsing history;
infer protected traits;
penalize missing private data automatically;
make permanent decisions from short-lived behavior;
use an AI-generated explanation not grounded in the score inputs;
allow companies to purchase a higher trust score;
advertise correlation as causation;
enable automated adverse employment decisions through the general API;
display more precision than evidence supports.

MVP acceptance criteria

The first scoring release is ready when:

every score has a public Score Contract;
only descriptive, readiness, match, confidence, or risk-signal classes are supported;
no universal score exists;
sample size, freshness, evidence source, confidence, and limitations are shown;
subjects can dispute factual inputs;
missing data is not automatically negative;
high-impact use is blocked or routed to a separate reviewed policy;
every score expires;
method and version are logged;
the product can replace a numeric score with an evidence summary when precision is not justified.

Sources

Verification record

Sources opened and checked: 2 August 2026
FCRA text page observed revised March 2026
Decision-safeguard status: Draft
Employment, consumer-reporting, discrimination, investment, and state-law review required before high-impact use: Yes