BiasFromJoin alpha

03 / Integrity

Every conclusion retains the path that produced it.

Who decided, under which policy, from what evidence, with which model assistance, how comparable cases were treated, and what happened next—all versioned, permissioned, attributable, and time-aware.

Join the private alpha

The decision graph

A record that can answer back.

Each object keeps provenance, policy, permissions, time, and ownership instead of flattening consequential judgment into a final score.

Policy
The lawful purpose, governing rules, and prohibited uses.
Decision case
The complete, isolated record for one consequential decision.
Subject
The person affected, with explicit rights and data permissions.
Criterion
A relevant, observable standard defined before the outcome.
Evidence
A sourced artifact, observation, outcome, or telemetry event.
Context
A consented condition that can explain opportunity or constraints.
Reviewer
A human with role, training, conflicts, and calibration history.
Assessment
A criterion-level judgment with evidence and uncertainty.
Model run
The versioned AI input, output, prompt, provider, and trace.
Fairness test
A documented metric, cohort, threshold, and interpretation.
Decision
The accountable outcome, rationale, owner, and review date.
Appeal
A correction or challenge with status, remedy, and final response.

The AI review board

Specialist agents challenge the case. A human owns the outcome.

No single model receives unlimited authority. Agents have narrow roles, explicit inputs, hard boundaries, budgets, traces, and escalation rules. Disagreement sends a case to another accountable person.

Intake guardianBEFORE review begins
  1. VERIFY lawful purpose, consent, access, and retention
  2. CHECK that criteria existed before the outcome
  3. BLOCK prohibited or incomplete uses
Evidence mapperWHEN evidence enters a case
  1. LINK each item to a relevant criterion
  2. SEPARATE observation, hearsay, inference, and fact
  3. FLAG missing, stale, or contradictory support
Reviewer challengerWHEN a human assessment is drafted
  1. TEST for vague, personality-coded, or unequal language
  2. COMPARE standards with similar reviewed cases
  3. REQUEST evidence or a second human review
Disparity analystWHEN enough governed cases exist
  1. MEASURE outcomes and error rates across cohorts
  2. CONTROL only for legitimate, documented criteria
  3. ESCALATE patterns without exposing individual identities
Policy sentinelDURING every model-assisted action
  1. ENFORCE prohibited-feature and autonomy boundaries
  2. LOG model, prompt, inputs, outputs, and confidence
  3. PAUSE on drift, incidents, or missing oversight
Appeal explainerAFTER a decision is proposed
  1. GENERATE a plain-language evidence map
  2. SHOW omissions, uncertainty, and available correction paths
  3. ROUTE the appeal to an independent accountable human

The integrity loop

Fairness is a monitored process—not a one-time audit.

Policies, reviewers, models, and populations change. Immutable historical records make it possible to learn without quietly rewriting what happened.

  1. 01

    Define

    Set the purpose, criteria, evidence, rights, and prohibited factors.

  2. 02

    Observe

    Collect relevant facts and consented context without covert inference.

  3. 03

    Review

    Require criterion-level human judgment supported by inspectable evidence.

  4. 04

    Challenge

    Run independent AI checks, cohort tests, and reviewer calibration.

  5. 05

    Decide

    Keep a named human accountable and disclose how AI informed the case.

  6. 06

    Learn

    Resolve appeals, monitor outcomes, and improve policy without rewriting history.

The measurement system

Measure fairness from more than one angle.

Every dashboard states the population, lawful purpose, metric, uncertainty, sample limits, and action threshold behind what it shows.

Evidence coverage
How much of a decision is supported by relevant, current evidence.
Reviewer variance
How outcomes change across reviewers, locations, and time.
Conditional disparity
Whether comparable cases receive different treatment across cohorts.
Error parity
False-positive and false-negative differences where ground truth becomes available.
Appeal quality
Appeal access, correction rate, overturn reasons, and time to remedy.
Outcome validity
Whether the decision predicts the legitimate outcome it was meant to support.
Process dignity
Whether affected people understand the result and feel able to correct the record.
Operational value
Review time, rework, legal exposure, inconsistency, and avoidable repeat cost.

The operating model

Govern decisions like critical infrastructure.

Align automation to mission and policy, trace every action, join system evidence with human context, and improve systems rather than surveil individuals.

Governed agent operations

Mission → policy → case → review → decision.

Inspired by Paperclip: each agent has a role, budget, approval boundary, and trace. Humans can pause, override, reassign, or terminate work.

Human context + system signals

Evidence explains what. People help explain why.

Inspired by Swarmia: operational data meets consented qualitative context, focused on cohort improvement rather than ranking.

People source of truth

Role → goals → feedback → growth decision.

Informed by HiBob: synchronize structure and outcomes, then add decision provenance, challenge, and appeal rights.

Non-negotiable boundaries

Responsible AI is part of the product—not a policy page.

Employment and education AI can affect fundamental rights. BiasFrom is designed as high-risk infrastructure from day one.

Never
  • Infer protected traits, stress, personality, or emotion from faces, names, voices, accents, or behavior.
  • Use emotion recognition in workplaces or schools.
  • Automatically hire, fire, promote, grade, license, or deny an appeal.
  • Apply a secret race, nationality, disability, or migration-status multiplier.
Always
  • Tell people when and how AI influenced review.
  • Expose criteria, relevant evidence, uncertainty, model versions, and the accountable owner.
  • Provide correction, accessibility, independent review, and meaningful appeal.
  • Test subgroup performance before launch and continuously after deployment.

The European Commission classifies AI used in employment and exam scoring as high-risk and prohibits uses including workplace or education emotion recognition and certain biometric categorization. BiasFrom should exceed those controls. See the EU AI Act overview.

The category strategy

Start with one painful review. Become the integrity layer for all of them.

Prove one repeatable workflow, earn trust with measurable outcomes, then expand the same graph and assurance system into adjacent markets.

01 · Wedge

Performance review integrity

Open role architecture, goal lineage, evidence, voluntary well-being pulses, feedback quality, calibration, decision records, and employee appeals.

02 · Expand

Global assessment integrity network

Candidate passports, marker calibration, offline evidence, local-language explanations, independent appeals, and regulator-ready assurance.

03 · Platform

Decision integrity network

Policy, evidence, agent, fairness, appeal, and audit APIs for organizations where humans and AI make consequential decisions together.

The build standard

The person affected by a decision is a participant—not a data point.

Evidence is easy to inspect, context safe to contribute, expectations consistent, AI visible, and appeals possible without specialist knowledge.

Private alpha

Make the next decision answerable.

Join for product previews, research invitations, and early-access openings. BiasFrom updates only.

One address. No unrelated campaigns.