Inference as Fact

A model score enters the database with the same schema weight as a sensor reading. Downstream systems treat it as input ground truth. The distinction between “the system inferred” and “a named authority decided” collapses at the persistence layer — and every derivative built on top inherits the confusion.

The fix is not to discard the inference. It is to record it as a conjecture — the act of inferring captured at full strength, its content marked at an honest tier — rather than collapsing act and content into a single “fact” the floor cannot tell apart from a measurement or a decision. Automated bind at scale uses recorded delegation: L0 policy, calibration study, per-decision commit with policy ref — not write-through.

In software systems

The failure manifests when L2 or L3 outputs are persisted without commitment metadata and consumed as if they were L0 events or L1 observations. A fraud score of 0.87 is stored in customer_risk_profile.current_score and overwritten nightly. A demand forecast becomes the inventory replenishment trigger with no record of model version, calibration date, or confidence interval. An LLM-generated summary is saved to case_notes.clinical_summary and displayed to clinicians without attribution to inference layer.

Materialised views and cache tables accelerate the error: the ephemeral computation becomes the system’s working truth because rebuild cost is high and lineage is untracked. Event logs may contain the original inference event, but operational systems read from tables that treat inference as fact. When the model is retrained and scores shift, history is silently replaced rather than superseded. The append-only log exists in theory; the operational read path mutates inference in place.

In human organisations

Organisations conflate expert judgment with expert opinion recorded informally. A senior analyst’s forecast is emailed to the planning team and enters the budget spreadsheet without a commitment event — no timestamped decision, no named authority, no evidence snapshot. It becomes organisational fact through circulation, not through commitment.

Meeting notes capture “the team believes churn will reach 12%” as if consensus were measurement. Consultants’ slide decks enter strategy documents as assertions. The failure is epistemic category collapse: probabilistic inference, deterministic derivation, and principled decision are stored in the same mental and documentary layer. When outcomes diverge from prediction, nobody can reconstruct who committed to what, with what evidence, at what moment — because nobody committed. The inference was never elevated to decision; it was simply repeated until accepted.

In socio-technical systems

Clinical pathways illustrate the compounded damage. A diagnostic model produces a probability; the EHR writes it to the problem list as a coded diagnosis. Billing, referral, and subsequent clinicians treat the code as documented fact. The physician who glanced at the suggestion never performed a commitment boundary crossing — but the record shows a diagnosis as if they had.

Regulatory reporting aggregates these conflated entries. Quality metrics measure model-assisted documentation as physician judgment. Patients receive letters citing conditions that exist only as uncommitted inference in the reasoning layer but committed fact in the billing layer. The socio-technical stack has aligned software persistence with organisational habit: both treat “what the system computed” and “what an authority decided” as interchangeable for operational convenience.

Structural causes

Epistemic tier collapse

Inference-as-fact is what epistemic tier collapse looks like in production: interpreted or probabilistic artifacts stored and read with the epistemic status of ground truth — no uncertainty characterisation, no commitment boundary, no lineage to the authority that would have to stand behind a Class A decision.

Model scores, forecasts, and LLM outputs are L2/L3 by nature. Many also inherit frozen collective subjectivity from training; tier collapse makes that inheritance invisible. When they are written to operational tables without promotion through a commitment boundary, the failure mode is inevitable: downstream systems calibrate action to “fact” because the schema says fact. The damage is not that inference occurred; it is that the structure denied anyone the ability to know it was inference.