Roadmap

Status: Gate-based public roadmap
Current release posture: v0.1 Public Draft; official numeric scoring disabled

This roadmap describes sequence and exit criteria, not delivery promises. A calendar target cannot waive a security, privacy, measurement, accessibility, governance, legal, or integrity gate.

1. Delivery principles

• Publish the construct and its limitations before presenting results.
• Validate demand and measurement assumptions before building a large proprietary model.
• Separate the public specification, private platform, and private evaluator.
• Use synthetic data until real-data controls are approved.
• Keep official scoring disabled until a later governed release.
• Prefer deterministic contracts and reproducible evidence lineage over interface theater.
• Publish what is live, limited, experimental, controlled, and planned.
• Publish methodology and index definitions before any comparative value renders.
• Use an original visual language; research-platform rigor does not authorize copying another product’s layout, CSS, components, typography, color system, or page architecture.
• Publish privacy-protected method failures and corrections as part of benchmark stewardship.

2. Foundation: public kernel

Target sequence: Initial 30 days
Purpose: Produce a coherent artifact that is safe to review publicly.

Deliverables:

• normative specification, benchmark card, scoring architecture, and limitations;
• charter, governance, conflicts, contribution, conduct, appeal, security, data, and trademark policies;
• independence and institutional firewall with external-majority Commission, commercial-role separation, reserved-claim, and structural-entity gates;
• ten-domain ontology with 55 capability families and stable identifiers;
• public schemas, invariant rules, and valid/invalid synthetic fixtures;
• experimental deterministic reference scoring over synthetic cases only;
• provider-neutral evaluator interfaces and deterministic mock behavior;
• premium public explanatory site and synthetic evidence narrative;
• methodology pages for the proposed Demonstrated Capability profile, Time to Production research candidate, Evidence Strength, Recency, confidence/coverage, and domain packs;
• a research/publications information architecture and a privacy-preserving data-explorer design, with no named-person ordering;
• version manifest, changelog, source ledger, and live-versus-planned register;
• automated schema, scorer, evaluator, build, accessibility, responsive, and claims checks.

Exit gates:

• all public documents and links agree on v0.1 status;
• official numeric publication is impossible in code and policy;
• no single global human number, named-person leaderboard, or cost-per-person axis is present;
• every sample is visibly synthetic;
• ontology and schema release tests pass;
• site source passes desktop/mobile, keyboard, reduced-motion, contrast, metadata, performance, and link checks;
• current-source, license, trademark, privacy, security, measurement, and governance review outcomes are recorded;
• no outside organization is represented as a partner, sponsor, signatory, validator, or endorser;
• operational and reporting-channel gaps remain visibly blocked;
• founding-steward publication approval is explicitly recorded after the other gates, without overriding any failed gate.

2A. Phase 2 source evolution: trust and aggregate register

Sequence: After the foundation source gate; before any real-person intake
Purpose: Make the future trust and register model inspectable without activating it.

Deliverables:

• a proposed T0–T4 trust-progression ladder with operational readiness and result lifecycle modeled separately from it, E0–E5 evidence maturity, capability, disclosure holds, and credential state;
• explicit v0.1 score suppression at every trust status;
• prestige-neutral verification and corroboration rules;
• behavioral-integrity reason codes, temporary routing/hold actions, and appeal requirements, with no country allow/block rules;
• a proposed multidimensional composition registry that keeps AI-native delivery separate and keeps evidence strength, coverage, and recency out of capability aggregation;
• a synthetic aggregate-register preview showing participation and evidence density, never geographic aptitude;
• minimum-cell, organization-diversity, dominance, complementary-suppression, coarse count/share band, denominator-recovery, exact prior-snapshot differencing, shared disclosure-budget, sampling-bias, and publication-review requirements;
• private verification, fraud-review, human-review, appeal, and disclosure-control architecture;
• premium, accessible trust, profile-anatomy, and register surfaces using only synthetic data.

Exit gates:

• every Phase 2 artifact is marked Proposed or Not authorized;
• no trust status can activate a score, credential, real-person profile, or public API response;
• synthetic/demo and production-data paths have bidirectional leakage tests;
• geography, employer prestige, school prestige, fame, and referral association cannot alter integrity risk, queue priority, verification standard, trust progression, or capability; any legally required service-eligibility or processor-route constraint must remain in a separate purpose-scoped jurisdiction service;
• an integrity signal can route, step up, rate-limit, or temporarily hold, but cannot silently reject or publish a fraud label;
• all restrictions carry a reason code, policy version, expiry, human-review route, and appeal state;
• the aggregate preview suppresses small cells and labels sampling bias in the visualization itself;
• desktop, mobile, keyboard, reduced-motion, accessibility, source-integrity, and production-build checks pass.

Phase 2 source completion does not authorize the founding cohort, a live register, numeric scoring, identity proofing, device fingerprinting, public profiles, or employment use.

3. Controlled-pilot preparation

Target sequence: Days 31–60, only after foundation gates
Purpose: Build the minimum private operating boundary for authorized feasibility work.

Platform work:

• create separate private platform and evaluator repositories;
• implement identity, organization, tenant, consent, and authorization foundations;
• implement evidence inventory, claim/link lineage, encrypted artifact storage, and minimized audit events;
• build case-scoped reviewer workflow, correction, conflict routing, and adjudication;
• build deterministic calculation and replay behind a nonofficial pilot flag;
• implement private trust/readiness projections and result-lifecycle references without public profiles, results, or credentials;
• configure security, conduct, privacy, and appeal reporting channels.

Governance work:

• seat the external-majority FDE Benchmark Commission at independence stage I1 before any real-person pilot authorization;
• publish its roster, terms, independence determinations, conflicts, funding, appointment/removal rules, first valid meeting record, and signed pilot decision;
• implement and test the training, employment, staffing/recruiting, sales, evidence, rating, calculation, issuance, appeal, and publication role/data firewall.

Security and data work:

• approve the threat model, data flow, classification, retention, deletion, incident, provider, and legal-role analyses;
• implement tenant isolation, key management, least privilege, access review, and break-glass controls;
• implement malware, archive, secret, file-type, and prompt-injection scanning;
• prove sandboxed, network-denied, quota-limited artifact analysis;
• test provider retention/training behavior and local or customer-controlled alternatives.

Measurement work:

• complete job analysis and external content review;
• write detailed rubrics for the highest-priority capability families;
• define critical-gate hypotheses and alternate demonstrations;
• design the Shared Work Challenge Battery around a common scenario substrate and separately declared human-only, model-only, and human-with-AI conditions, without presuming score comparability;
• pre-register feasibility-cohort questions, sampling, thresholds, data rights, and analysis;
• train reviewers and test conflict, accommodation, correction, and appeal flows;
• implement grader invariants for known-good, known-insufficient, and known-invalid synthetic responses;
• run learnability and answerability review before any scenario becomes controlled evaluation material;
• define adjudication-log sampling, grader-defect suspension, impact analysis, and participant-remediation procedures;
• obtain qualified counsel review of applicable worker/candidate, automated-decision, discrimination, privacy, notice, audit, accessibility, and human-review requirements for each intended jurisdiction and use.

Exit gates:

• independent security/privacy review appropriate to the pilot risk;
• tested participant notice, consent, withdrawal, deletion, correction, accommodation, and appeal;
• signed schemas and reproducible run/calculation manifests;
• named trained reviewers and non-conflicted appeal coverage;
• approved pilot protocol and residual-risk acceptance.

4. Feasibility calibration

Target sequence: Days 61–90, only after controlled-pilot approval
Purpose: Learn whether the construct, evidence model, workflows, and reviewer process are usable—not prove broad validity.

Planned work:

• recruit a permissioned cohort spanning non-FDE, developing, experienced, and senior/principal work profiles;
• obtain independent expert judgments and blinded artifact review where practical;
• measure evidence availability, reviewer disagreement, workflow burden, accommodations, and major failure modes;
• run evaluator stability, prompt-injection, irrelevant-attribute, and provider-conformance tests;
• revise ontology, rubrics, evidence rules, schemas, and reviewer guidance;
• publish an aggregate findings and limitations report only after snapshot-specific consent, disclosure, differencing, and safe-output approval;
• publish invalid scenarios, broken grader behavior, held findings, failed hypotheses, and corrective changes at a non-identifying level;
• study whether Time to Production can be defined without rewarding unsafe speed or penalizing organizational delay; retain it as a research candidate if not.

A cohort of roughly 20–30 may support feasibility and content revision. It cannot support percentiles, population norms, broad subgroup fairness, or predictive employment claims.

Exit gates:

• documented feasibility findings and protocol deviations;
• reviewer-agreement and adjudication analysis against predeclared thresholds;
• accessibility and irrelevant-attribute findings with remediation;
• no unresolved high-severity privacy, security, or integrity defect;
• public change rationale and updated benchmark card;
• governance decision on whether a larger external pilot is justified.

5. Verified-profile beta

Earliest sequence: Months 4–6; not tied to a promised date
Purpose: Test a controlled verification service without claiming certification or general validity.

Potential scope:

• a larger, more diverse external practitioner cohort;
• public and private repository evidence under approved isolation;
• adaptive live evaluation and controlled cases;
• reference and deployment corroboration;
• capability-profile history with version-separated display;
• private-by-default profiles and opt-in disclosure;
• credential interoperability prototype without accredited-certification claims;
• three-layer research sequence: factual Verified Production Record, multidimensional Capability Profile, then a credential-interoperability candidate with issuance still disabled until its independent gates pass;
• first domain-pack public draft;
• enterprise cohort research with human-governed, non-adverse use terms;
• method-linked research publications and a privacy-protected aggregate data-explorer beta;

Required before beta designation:

• operational non-conflicted verification and appeals;
• documented calibration and reliability evidence appropriate to each displayed claim;
• external measurement, fairness, accessibility, privacy, and security review;
• controlled-case integrity, contamination monitoring, revocation, and incident response;
• stable cost, reviewer capacity, and service controls;
• public limitations and aggregate operations reporting.

6. v1.0 consideration

Earliest sequence: Months 7–12; release only when gates are met

Version 1.0 requires:

• a defined target population and intended-use boundary;
• externally reviewed construct, ontology, rubrics, and critical gates;
• at least two appropriately designed calibration cohorts;
• predefined and met reliability, calibration, and reproducibility thresholds;
• tested fairness, accessibility, accommodations, and irrelevant-attribute controls;
• operating correction, appeal, suspension, and revocation;
• secure controlled cases and contamination response;
• operational privacy, consent, retention, deletion, provider, and incident controls;
• independent review for high-conflict results;
• public benchmark card, methodology, limitations, version policy, and release vote;
• evidence that employment, training, platform, payment, sponsor, prestige, and fame fields cannot alter results;
• jurisdiction- and use-specific counsel review without representing research issue spotting as legal compliance.
• independence stage I3, an external-majority Commission vote for the exact release, an operating role/data/funding firewall, and a published structural-entity decision.

Predictive employment claims require separate longitudinal validation for the specific use and population. They are not implied by a v1.0 label.

7. Research and publication program

The public platform should become a citable research record through method depth, not human ranking. Planned publication types are:

• benchmark and ontology release notes with exact version scope;
• method pages for every displayed dimension and domain pack;
• aggregate cohort and evidence-availability studies;
• reliability, accessibility, fairness, contamination, and evaluator reports;
• system-failure and correction reports covering held findings, broken graders, invalid scenarios, and failed hypotheses;
• a “State of FDE Capability” report only when sample design and privacy support its claims;
• a versioned data explorer for disclosure-controlled aggregate distributions, coverage, evidence strength, recency, and research findings.

Every figure must link to method, population, period, provenance, confidence, coverage, limitations, and corrections. Small-cell controls and disclosure review must prevent reconstruction of individual records. Sorting and filters must not create a named-person leaderboard.

After the FDE construct passes its validation and governance gates, the research program may consider other AI delivery roles. Each role requires a separate job analysis, construct, ontology mapping, evidence and validation plan, benchmark card, and governance release. FDE evidence cannot be assumed to generalize.

8. Parallel workstreams

9. Explicitly deferred

• A new general-purpose FDE foundation model.
• Public leaderboards and universal rankings.
• Any single global human capability number in v0.1, and any later scalar that has not passed a new interpretation-and-use, measurement, consequence, independence, and release gate.
• Cost-per-person, compensation-efficiency, or “value for money” indices.
• Public ordered lists of named people, including views that reconstruct one through filtering or sorting.
• Copying another research platform’s trade dress, components, styles, layout, or page architecture.
• Automated employment recommendations or adverse decisions.
• A staffing, recruiting, placement, or commercial recommendation funnel inside the neutral benchmark authority surface.
• Formal personnel certification or accreditation.
• Broad paid individual tiers.
• Unreviewed on-premises or VPC deployment.
• A neutral foundation created only for appearance.
• Direct coupling to a training system, a particular evidence platform, or a single model provider.
• Expansion to other AI roles before the FDE construct has met its validation and governance gates.

10. Roadmap changes

Roadmap changes follow GOVERNANCE.md. A material scope or claim change requires a public rationale, conflict review, changelog entry, and benchmark-card update where applicable.

Current availability is controlled by LIVE_VS_PLANNED.md, not this roadmap.