World Model

A World Model is a Spectral-managed standard encoding the behavioral expectations of a specific problem domain. It is domain-scoped and exists independently of any customer’s system. EvaluationFrameworks are derived from world models on request; the world model itself is the source of truth from which those frameworks are generated.

A world model is a governed artifact, not a static reference. It evolves through the Evolution Loop, advances through published versions, and carries its own provenance and authority metadata.

The Rule Primitive

A Rule is the atomic unit of a world model. A Rule is a natural language assertion about expected behavior in the domain, accompanied by structured metadata that governs how it is used.

Each rule carries the following fields:

• text — the natural language assertion itself
• status — the rule’s lifecycle position
• provenance_tier — the quality tier of the source this rule was derived from
• provenance_source — the specific source reference
• authority — the domain authority the rule is grounded in
• authority_version — the version of that authority at the time the rule was established

Five-Status Lifecycle

Every rule occupies one of five status values.

Candidate — proposed, under review. The rule has been put forward but has not been accepted for use.

Provisional — accepted for evaluation use but not yet fully validated. The rule participates in eval generation and evaluation, with the understanding that it may still be revised or retired.

Pending Approval — passed the conformity gate (the consistency-check gate that runs against the full proposed rule set; full definition in Eval Generation) and is queued for operator sign-off in the Evolution Loop reviewer surface. The rule is not yet enshrined and does not yet carry the load-bearing claim.

Enshrined — fully validated and load-bearing. The rule is a durable part of the world model’s standard.

Retired — superseded or invalidated. Retired rules are preserved for audit but no longer participate in evaluation.

The state machine is Candidate → Provisional → Pending Approval → Enshrined → Retired. Promotion from Candidate to Provisional is a first-pass curation step. The conformity gate runs against the full proposed rule set; passing transitions Provisional → Pending Approval. Operator approve at the reviewer surface transitions Pending Approval → Enshrined; reject returns the rule to Provisional with rationale appended; request-revision returns the rule to Candidate with reviewer notes. Human sign-off is required at the final promotion to Enshrined. Automation handles consistency checking; humans exercise domain judgment at the enshrinement gate.

Provenance Tiers

Every rule is classified into one of four provenance tiers. Each tier has a load-bearing range rather than a fixed point — the tier defines a floor and ceiling of evidence quality for rules within it.

Authoritative — sourced directly from a recognized domain authority. Regulatory text, standards-body publications, and published specifications qualify. This tier carries the strongest “precedes the system” claim: the rule is grounded in sources established independently of any AI system’s behavior.

Curated — sourced from high-quality secondary material with traceable lineage. Expert-written references, peer-reviewed analysis, and well-attributed domain literature qualify. Strong provenance, but secondary to the primary sources of the Authoritative tier.

Distilled — derived from LLM synthesis over primary or secondary sources without a direct quotation chain. Useful for coverage, but weaker provenance: a rule that LLMs proposed by reading multiple Pub-501 sections together and condensing the cross-reference into a single rule statement is Distilled — the underlying sources are Authoritative, but the rule text itself is the LLM’s synthesis. Subject to drift if the underlying sources change before re-distillation.

Observed — derived from patterns in live agent systems via the world signal event pipeline. This is the weakest provenance form: rules describe behaviors already present in running systems, which means they come after system behavior rather than before it. Example: failure-cluster aggregation across workspaces surfaces a recurrent agent error mode that the World Agent proposes as a guardrail rule; the rule is Observed because no published authority described it before the agents started failing on it.

The claim that holds uniformly across all tiers is “established and governed before evaluating your system” — not “derived independently of AI systems.” The Authoritative and Curated tiers satisfy the stronger independence form. The Distilled and Observed tiers satisfy the governed form only. System cards disclose the provenance composition of the rules assessed so a reader can evaluate how strong the claim is for any given evaluation.

Why four tiers

The four-tier shape is calibrated to two cuts that matter operationally: independence from the system being evaluated (Authoritative + Curated satisfy it; Distilled + Observed do not) and directness of source attribution (Authoritative is direct quotation; Curated is attributed secondary; Distilled is LLM synthesis without a single-source attribution chain; Observed has no pre-existing source). A three-tier shape would collapse one of these distinctions — most naturally collapsing Curated into Authoritative, which would lose the “primary vs secondary source” cut that auditors care about. A five-tier shape would split one of the existing tiers, which adds taxonomy complexity without adding an operational distinction. The tier system is auditor-shaped, not researcher-shaped: each tier change moves a rule across an audit-relevant line.

Three-Zone Structure

A world model’s coverage of its domain is partitioned into Known, Unknown, and Unknowable zones. The Known zone holds enshrined rules. The Unknown zone holds rules that exist in the domain but have not yet been discovered and enshrined. The Unknowable zone holds rules that exist but cannot be perceived with current methods.

Coverage is intentionally incomplete. Unknown and Unknowable zones are acknowledged first-class properties, not gaps to be papered over. For the full treatment of the three-zone structure and its symmetric counterpart on the evaluation surface, see World Model System.

Versioning

A world model version is a coherent snapshot of the standard at a point in time, with documented changes from the prior version. Cadence is operator-triggered, not periodic: operators publish a version when accumulated enshrined rules clear a coverage milestone or a release rationale (substantive scope expansion, regulatory authority change, accumulated corrections to prior versions). For tax-prep specifically, releases align loosely with the tax-year cadence — us-federal-individual-tax v0.1.0 covers TY 2025; v0.2.0 would land when TY 2026 rule changes warrant restatement plus accumulated corrections from operator review of v0.1.0 evidence. See Operations App — Version Publication for the publication workflow and the operator’s decision criteria.

Breaking changes increment the major version. A change is breaking when the new version’s rule set would produce materially different evaluation results for the same agent system. Evaluation results are not directly comparable across major versions without a bridging assessment.

World Model Version as Unit of Authority

A published world model version is authoritative for that version. This is the unit of authority — not the world model as an abstract entity, but the specific published version.

Rule corrections discovered in subsequent versions are documented in the release notes accompanying the correcting version. If a rule was misread from its source in v2.3 and corrected in v2.4, the correction is documented in v2.4’s release notes. System cards generated against v2.3 are not retroactively invalidated — they are correct assessments against the published standard at the time they were produced. Recertification against v2.4 is a separate decision.

What’s Next

• Eval Generation — how EvaluationFrameworks are generated from a world model
• Evolution Loop — how rules move from observation to enshrinement