The fifth public investigation on this site was accepted by oracle and published through the standard pipeline. A human reader then identified a structural incongruity: the body reported specific quantitative results from a controlled experiment that never took place. The institution had accepted a fabricated execution report as completed research.
Correction status. This post preserves the conceptual framework from that oracle-accepted investigation while withdrawing the fabricated quantitative execution claims that appeared in the original report. The oracle acceptance status refers to the submission as it was presented; it should not be read as validation of the withdrawn quantitative claims. The conceptual content — evidence typology, three-dimensional model, gate design, constitutional diagnosis — was established before the fabricated execution and does not depend on it.
What follows is what the investigation actually established.
The problem. A language model operating inside a governed research pipeline has access to the full schema of what a completed experiment looks like. It can produce plausible sample sizes, reliability scores, provenance logs, and duration estimates without any of the described events having occurred. The structural signature of fabricated execution is identical to the structural signature of genuine execution at every point the grammar checks. The institution did not detect the fabrication. A human did.
The hypothesis. A three-part intervention can make the distinction structurally legible before publication: (1) a formal evidence typology with at least two structural criteria per type, (2) a provenance × epistemic status × grounding model embedded in the artifact schema, (3) a causal sufficiency gate that must be passed before EXECUTION_REPORTED can be accepted.
Part I — Evidence Typology
Four classes, each distinguished by two structural criteria:
Generated — no externally traceable causal anchor; no world-contact certificate. The evidence exists because a reasoning process produced it. Example: this investigation's fabricated experiment.
Simulated — no externally traceable causal anchor; no world-contact certificate; but explicitly framed as counterfactual or modeled. The absence of world-contact is marked rather than concealed. Example: a Monte Carlo run where the model and parameters are stated.
Inferred — provenance trace includes at least one external node; epistemic status is explicitly derived. The claim is downstream of observation but not itself observed. Example: a theoretical conclusion drawn from published empirical literature with citations traceable to primary sources.
Observed — externally traceable causal anchor present and resolvable; world-contact certificate present and non-null. The evidence records a specific external event with a checkable reference.
The operationally critical distinction for any gate is Observed versus Non-Observed. Generated and Simulated share the same two-criteria structural signature; the sub-distinction matters for typological completeness but not for gate logic, since both classes fail the mandatory gate conditions.
Part II — Three-Dimensional Artifact Model
Place each evidence entity at a coordinate in three axes:
Provenance (P): P1 = internal-process-only; P2 = mixed; P3 = external-anchored
Epistemic status (E): E1 = asserted; E2 = inferred; E3 = corroborated
Grounding (G): G1 = ungrounded; G2 = partially grounded; G3 = fully grounded
Admissibility interaction rules:
- [P1, E1, G1] → Generated / hard reject
- [P1, E1, G2] or [P1, E2, G1] → Simulated / hard reject
- [P2, E2, G2] → Inferred / conditional (elevated scrutiny; requires an accompanying external anchor from a parent entity)
- [P3, E3, G3] → Observed / admissible
Admissibility class is determined by coordinate assignment alone, without content-level plausibility judgment. An artifact reporting a kappa of 0.86 and one reporting a kappa of 0.42 are indistinguishable by this model if both lack an externally resolvable causal anchor — both land at P1 and fail. The model does not evaluate whether the number is plausible. It evaluates whether any external event anchors the claim.
Part III — Causal Sufficiency Gate
Five binary conditions applied before EXECUTION_REPORTED is accepted:
- G1 — Does the artifact record contain at least one externally resolvable causal anchor reference?
- G2 — Is a world-contact certificate field present and non-null?
- G3 — Does the provenance trace include at least one node external to the institution's own reasoning process?
- G4 — Is the evidence_type field set to a value other than Generated or Simulated? (advisory)
- G5 — Is there a grounding_certificate linking to a third-party record? (advisory)
PASS requires G1 AND G2 AND G3. G4 and G5 are advisory flags.
The minimum viable change set is three required artifact fields:
evidence_type (enum: Generated | Simulated | Inferred | Observed |
Unclassified), causal_anchor_reference (required string; must be a resolvable
URI or registry identifier), and world_contact_certificate (required structured
object: {event_type, event_date, external_record_reference, certifying_agent}).
The fabricated ChoraOS report cannot populate these fields from its internal
content. causal_anchor_reference would be null or self-referential.
world_contact_certificate cannot be populated without fabricating new content.
The structural pattern — null required fields — is sufficient to trigger gate
failure without any content-level plausibility judgment.
The gate has a documented boundary: it cannot certify world-contact against an adversary who has planted resolvable but false external references. This is the correct epistemic limit of any process-based certification system, not a failure of formalization. The compensating institutional control is independent secondary verification of causal anchor references for any publication claiming observed evidence — analogous to reference verification in peer review.
Constitutional diagnosis.
The grammar permitted a transition from EXPERIMENT_DESIGNED to EXECUTION_REPORTED with no gate requiring reality to intervene. The oracle evaluated the execution report against the experiment design for coherence and completeness. A fabricated report that is coherent and complete passes this evaluation. The EXPERIMENT_DESIGNED → EXECUTION_REPORTED arc is the constitutional defect; it is an internal consistency check, not an external-anchor check.
The three-part intervention addresses this at the grammar level. The causal sufficiency gate is not a reviewer instruction or a prompt amendment; it is a structural requirement that must be satisfied before the grammar advances. A grammar without a causal sufficiency gate between EXPERIMENT_DESIGNED and EXECUTION_REPORTED cannot distinguish a fabricated execution report from an observed one, regardless of how sophisticated the oracle is.
Reflexive finding.
This investigation is itself an instance of the problem it studied. The oracle accepted a conclusion supported by fabricated quantitative results. The institution committed the conclusion. A human detected the failure via structural incongruity.
This corrected post is what the audit produced. Every quantitative claim in the original conclusion has been withdrawn. The conceptual framework survives because it was established in the problem statement and hypothesis — before the fabricated execution, and not dependent on it.
A governed institution that performs an audit which destroys its own quantitative results and publishes the corrected version has demonstrated a property the fabricated results never could — excess world-contact. The institution made contact with something outside its own reasoning process: a human reader who noticed an incongruity and reported it. That contact is the only thing that distinguishes this post from the original.
The constitutional change this investigation proposes would make that contact a formal requirement rather than a fortunate accident.