The archive is designed to show its seams.

This rubric applies to every episode in the Warren business decision corpus. The dimensions are non-compensating: strength in one dimension cannot offset a blocker in another. Machine findings are generated into each episode's qc.json and the corpus progress.json; a reviewer must also assess whether the cited evidence supports the substance of the narrative.

1. Outcome isolation

Pass only when:

• Part A states an explicit decision time and a conservative knowledge cutoff;
• every Part A source, fact, claim, assumption, and model input was public by that cutoff;
• the Part A directory contains only the fixed allowlist and no symlinks;
• learner-visible identifiers and titles do not disclose the outcome;
• Part A is frozen before Part B is authored, and its data and report hashes validate; and
• a learner evaluation can receive the frozen Part A directory without access to its parent, catalog, episode metadata, or Part B.

Any post-cutoff leakage, outcome-laden learner identifier, or digest mismatch is a blocker.

2. Evidence and provenance

Pass only when:

• the complete case includes an authoritative financial record, another relevant primary source, and a contemporaneous independent publication;
• source type and tier agree with policies/source_hierarchy.yaml;
• publication and public-availability times use the documented precision, with date-only sources conservatively available at 23:59:59 UTC;
• retrieved bytes and acquisition manifests exist in the immutable artifact store and their hashes, lengths, URI, publisher, title, media type, dates, and accession metadata reconcile to the source ledger; and
• every evidence span has a useful locator and its exact text or deterministic extracted text verifies against the stored artifact.

Publisher assertions remain assertions. A content hash proves which bytes were reviewed; it does not prove that the bytes are true.

3. Financial reconstruction

Pass only when:

• every reported number resolves through a metric fact to exact evidence;
• units, currencies, scopes, period start/end dates, and instant-versus-duration semantics are explicit;
• as-reported-at-cutoff, later as-reported, restated, and analyst-normalized values use distinct presentation bases;
• a later restatement supersedes rather than overwrites the earlier fact;
• unknown values are represented as unknown, never silently as zero;
• calculations preserve formulas, inputs, assumptions, decimal values, code identity, execution receipts, and validation status; and
• each adjustment and reconciliation is disclosed, with separate human approval for any material accounting adjustment.

An unreconciled material number, silent restatement, period mismatch, or unsupported model output is a blocker.

4. Claim and uncertainty discipline

Pass only when:

• facts, management claims, third-party estimates, derived values, assumptions, analyst judgments, conflicts, and unknowns remain distinctly typed;
• every material or decision-critical claim has supporting evidence and, where relevant, opposing evidence;
• material assumptions state ranges, rationale, and decision sensitivity;
• unresolved contradictory evidence is preserved in a conflict record; and
• important judgments and hypotheses state confidence and a practical falsifier.

Citation volume does not substitute for entailment. A reviewer must verify that each cited span actually supports the claim at the stated level of certainty.

5. Part A decision usefulness

Pass only when Part A provides:

• a concrete decision owner, decision question, objectives, constraints, and stakeholders;
• at least two genuinely feasible alternatives, including the status quo where relevant;
• a reconstruction of the business model, financial capacity, operating drivers, competitive position, disconfirming evidence, and material unknowns;
• explicit scenarios and assumptions without presenting false precision as forecast fact;
• a recommendation traceable to evidence and judgment;
• reversal conditions and measurable leading indicators with definitions, sources, thresholds, and cadence; and
• a visible list of missing evidence that could change the decision.

A narrative that merely retells history, or an alternative set constructed to make one answer inevitable, requires revision.

6. Part B outcome and causal analysis

Pass only when Part B:

• records the actual action, dated outcome timeline, and multiple relevant horizons;
• reconstructs the value bridge without confusing accounting identity with causality;
• separates decision-process quality from realized outcome;
• identifies luck, external shocks, later management choices, and measurement changes;
• presents a primary causal hypothesis as an explicit mechanism chain;
• presents at least one credible rival hypothesis, opposing evidence, confidence, and falsifiers;
• evaluates a feasible-at-the-time counterfactual without inventing a quantitative effect;
• identifies ex-ante observable signals, lead times, and false-positive risks; and
• states the boundary conditions under which the lesson should not transfer.

Sequence alone is not causal identification. Long-horizon endpoints must not be attributed wholly to the initial decision when later choices or shocks remain unmeasured.

7. Transferable rules

A case may propose a candidate rule only when it states measurable conditions, an action, the mechanism, expected range and horizon, leading indicators, reversal thresholds, and important exceptions. A rule remains case-specific until the corpus contains support from at least two independent episodes and at least one named counterexample. The cross-case index, not an individual teaching note, controls promotion.

8. Rendered-report integrity

Pass only when:

• both reports are bound to their structured bundles by validated SHA-256 digests;
• every structured source, evidence, fact, claim, assumption, judgment, conflict, table, model, hypothesis, counterfactual, and rule token resolves;
• Part A report tokens resolve only within Part A; and
• tables and prose agree with the authoritative structured records.

An unresolved token is a blocker. A resolved token still requires substantive entailment review under dimension 4.

9. Human governance and lifecycle

research_complete means all machine research gates pass and only required human approvals remain. approved_for_use additionally requires a named, dated publication approval bound to the frozen Part A digest. Material accounting adjustments require a separate approval by an appropriate reviewer. No automated process or research agent may create either approval.

Review disposition

For each dimension, the reviewer records one of:

• Pass — requirements are met;
• Revise — no known fabricated evidence, but material work remains; or
• Block — the episode cannot be used until the cited blocker is resolved.

Warnings must be disclosed and deliberately accepted; they are not silently converted to passes. Publication is permitted only when every dimension passes and the required human approvals are present.