Epistemic Tier Collapse
Interpreted or probabilistic artifacts persisted and consumed as ground truth
L2 and L3 artifacts must not masquerade as L0 — epistemic tier must survive persistence and read paths.
Epistemic tier collapse is the anti-pattern of letting L2 or L3 artifacts look and behave like L0 in schemas, dashboards, and organisational habit — without a commitment boundary, without uncertainty disclosure, and without lineage to a named authority.
A closely related, ubiquitous form is CRUD as domain language — persistence verbs (Created, Updated, Deleted) standing in for domain vocabulary at tier one. That is tier collapse at the naming layer: capture mode and business meaning vanish into storage shape before inference or rendering even enter the picture.
Another is the confident deck — L3 polish consumed as proof: presentation erases the gap between how strong a claim looks and how strong it is.
It is not the same as frozen collective subjectivity, which names subjectivity frozen into trained model weights at corpus time. Tier collapse is what you choose in software and process when inference is stored, read, or acted on as fact. The two often appear together with LLMs; either can exist without the other.
In software systems
Model scores in operational tables. LLM summaries as the audited record. Sentiment and risk fields in master data with no inference flag. Materialised views that become the de facto read path while the event log keeps honest lineage elsewhere.
The schema says fact. The UI says fact. Downstream invariant enforcers consume “fact.” Epistemic uncertainty is zero in the contract because nobody published it.
In human organisations
“The committee concluded” with no vote. Meeting minutes that assign decisions never made. Expert opinion pasted into policy as if it were measurement. Judgment recorded as fact because the document layer has no tier distinction.
In socio-technical systems
Clinical codes from model suggestions without physician commitment. Compliance metadata from generated summaries. Agentic workflows that write structured fields from LLM output. Inference-as-fact is the failure mode when tier collapse hits the immutable log. The same collapse appears at the commitment boundary when tier-one captures of decisions — signatures, keystrokes, second-hand reports — are treated as the genuine inner commitment they only approximate: coerced commitment, transcription error, forged commitment, insincere commitment, capture-chain distortion, and second-hand commitment.
When the model also carries frozen collective subjectivity, tier collapse compounds the damage: biased interpretation, then mislabelled as ground truth.
Why it persists
Short-termism. Direct persistence is the fastest path to shipping “AI features.” See provenance deferred and short-termism.
Cognitive bias. If it is in the database, it must be true.
Missing knowledge. Engineers model JSON, not epistemic tiers.
Local optimisation. Each team publishes its scores for downstream convenience. See local optimization.
Misaligned incentives. Throughput and headline metrics outrank rerunnable chains — publication over provenance, metric substitution. See misaligned incentives.
Principled alternatives
Commitment boundary
Derivation and commitment stay separate. L2/L3 remain labelled and uncertainty-characterised until an authority crosses the boundary with evidence, policy, and moment — then L0. Tier collapse becomes visible in audit and schema rather than invisible in a shared table.