LLMs at the Commitment Boundary
Classifying every meaningful LLM variant — and why the harness, not the model, sets the trust
Large language models are the most practically consequential technology to apply this series’ trustworthiness framework to. Part IV established how trustworthy a situation is — six facets in The Facets Composed (including processor fidelity / disposition in Disposition and Processor Fidelity). Part V — quality (Strengthening the Commitment Boundary), legitimacy (When Social Stability Matters More Than Quality), meritocracy (Meritocracy — The Sweet Spot). This article (Part VIII) applies that stack where LLMs are deployed at scale — in exactly the contexts where those facets matter most, and where the most common deployment patterns sit in the most dangerous cells of the classification grid.
They are also the industrial-scale version of The Confident Deck: fluent, confident output with no native signal of how much is known and how much is invented.
The central insight is this: the trustworthiness classification of an LLM system is determined entirely by its epistemic harness — not by the model itself. The same underlying model can sit at L3-U-high (naked, no citations) or approach L1-U-med (RAG with verified citations, deterministic extraction with downstream verification). The investment in harness is the investment in trustworthiness — see the Artifact Classification map for how each harness variant lands in the grid.
The baseline: naked LLM
A general-purpose LLM answering questions from parametric memory alone sits at L2-U-high — or L3-U-high when its output is formatted for human consumption.
The training data is a compression of internet content of unknown provenance, unknown quality distribution, and unknown vintage mix — plus every editorial, labelling, and filtering choice applied before and during training. The model cannot distinguish what it knows from what it has confabulated. Its output is a L2 probabilistic projection (multi-step transformation from heterogeneous sources) with high uncertainty (uncharacterised, uncalibrated, no citations). That is frozen collective subjectivity: collective human judgment frozen into weights, applied as if it were objective, usually without the model knowing it carries subjectivity at all — high probabilistic computation loss of the systematic, directional kind.
For negligible-severity tasks — brainstorming, drafting, creative exploration, personal productivity — this is acceptable. The LLM’s fluency and breadth are genuine assets. Correctness is a preference rather than a requirement.
For anything above negligible severity, the naked LLM is in the prohibited cell for automated decisions and the “caveat required” cell for human consumption — per When the Commitment Boundary Needs Reinforcing and Artifact Classification grid. It belongs at L3-U-high: human-optimised output with uncharacterised uncertainty. A human reader who treats it as authoritative is making a decision based on an artifact whose uncertainty they cannot assess.
The contamination rule
The most urgent case those frameworks forbid: LLM output stored directly as a L0 fact without a commitment boundary crossing — the inference-as-fact failure mode, and exactly the failure that opened this series with The Diagnosis (a probabilistic inference hardened into committed fact).
When an LLM hallucinates a fact and that fact is recorded as a L0 event — a decision, an attestation, a conclusion commitment — the hallucination becomes permanent ground truth. The log is append-only. Correcting events can be appended, but the original contamination remains. Every derivation built on that L0 event inherits the contamination.
This is not hypothetical. It happens when:
- LLM-generated content is automatically ingested as structured data
- LLM summaries are stored as the record of what was reviewed
- LLM-produced metadata is written to databases without verification at the boundary
- Agentic LLMs write to records as part of automated workflows with no recorded delegation L0
Architectural rule: LLM output must never become L0 by write-through. It stays pre-bind evidence — typically L2 or L3; harness-dependent, sometimes approaching L1 — referenced by the commit event, not the commit itself. The model has no standing to commit; authority weight attaches to the named human, collective, or authorized process that crosses the commitment boundary, per severity in When the Commitment Boundary Needs Reinforcing.
That rule does not require a human to click approve on every instance. It requires that some authority — with weight, policy, and accountability — committed before or at the bind. Contamination is bypassing that structure, not automating per se.
When the LLM should commit — recorded delegation
The contamination rule is not a bias toward human judgment. It is a bias toward honest commitment structure. On a specific, bounded task, a harnessed LLM may outperform humans — more accurate, more consistent, better calibrated — and that superiority can be demonstrated from the record, not asserted from vendor slides.
The full pattern — evidence base (human L0 history, outcomes, counterfactual replay), delegation L0, per-decision commits with policy ref, drift monitoring, severity constraints, and refusal of rubber-stamp theatre — is recorded delegation.
Summary:
- Rational to delegate when calibration shows the LLM reliably beats the human baseline on the classified task — deferring to measurement is epistemically honest (Rule 3 — rubber-stamp approval is worse than honest automation).
- Delegation itself is L0 — direct policy commit or indirect thresholds in an authorized process; humans remain in the accountability chain via authorship, certification, and monitoring — not necessarily per-click approval.
- Catastrophic severity: Rule 2 still requires a human at each consequential bind; recorded delegation applies fully at significant and minor where the matrix permits automation with escalation and audit.
- Refuse: agentic bind with no delegation L0, post-hoc thresholds, or decorative human approvers when a published automated policy with calibration evidence is the honest alternative.
The classification of meaningful LLM variants
Each variant below maps to a cell in the facet space from The Facets Composed. Explore cells interactively for examples, permitted consumer classes, and per-cell LLM variant detail.
Open Artifact Classification map →
RAG over a single verified authoritative source, exact citations, human verifies citations
Approaches L0-U-med. The LLM is performing a rendering step over a verified source. Citations are provided and verified. Medium uncertainty because the LLM may still paraphrase imperfectly and retrieval may miss a more relevant passage. Suitable for Class B audit consumption when human verifies.
Enterprise RAG over a curated knowledge base, citations with retrieval timestamps, staleness disclosed
L1-U-med. The canonical best-practice enterprise RAG case. The knowledge base is curated and regularly updated. Citations include document names, versions, and retrieval timestamps. Staleness window is documented and disclosed. Suitable for Class C operational support.
RAG without citations shown to the consumer
L2-U-high from the consumer’s perspective — regardless of the quality of the retrieval. The consumer cannot verify any claim. The retrieval may be excellent; the consumer has no way to know. From the consumer’s epistemic position, this is indistinguishable from naked LLM output. The citation is what makes uncertainty characterisable. Without it, uncertainty is not characterised.
LLM with web search, URLs shown, retrieval date disclosed
L2-U-med. Web sources are themselves of variable quality (L0 through L3 in the source hierarchy). The LLM synthesises across sources, introducing probabilistic uncertainty. The citation makes the uncertainty characterisable — the consumer can follow links and judge source quality. Medium uncertainty, not low, because the synthesis step is probabilistic and the sources are heterogeneous.
Domain-specific LLM trained on verified corpus, citations from training domain
L2-U-med within domain, L2-U-high at domain boundaries. Better calibrated than general internet models within the training domain. Training corpus version and cutoff date are mandatory metadata — without them the consumer cannot assess the temporal uncertainty. At domain boundaries or for events after the training cutoff, uncertainty slides to High without the consumer being able to detect it.
Multiple independent models, outputs shown separately
L2-U-med where models agree; explicit epistemic uncertainty signal where they disagree. The disagreement is information, not noise. Pre-aggregated ensemble destroys this signal. The harness must show individual model outputs and flag disagreements for human attention.
Agentic LLM taking consequential actions without recorded delegation
L2-U-high masquerading as L0 when no prior L0 delegation policy exists. The failure is uncommitted automation — see pattern entry for accountability chain requirements.
Agentic LLM operating under recorded delegation (L0 policy + drift monitoring + contestability)
Permitted where the calibration matrix allows automated bind — typically significant or minor severity with certified process, confidence escalation, and audit. Each commit references the delegation L0 and calibration study.
LLMs in the enterprise decision map
When the Commitment Boundary Needs Reinforcing established the three dimensions. Where do LLMs fit across the severity and nature dimensions? The Enterprise Decision Map plots hundreds of enterprise decisions and artifacts by domain, nature, and severity — filterable on the map or in a searchable list.
Negligible severity: full flexibility, including naked LLM. This is the LLM’s natural habitat.
Minor severity: LLM with logging and periodic audit. RAG preferred. Recorded delegation may replace per-instance human review when calibration L0 exists.
Significant severity, ad-hoc tasks: LLM with RAG and citations — human commits before bind unless delegation policy already covers the task class.
Significant severity, high-volume variable: Full harness + escalation — or automated bind under recorded delegation when track record evidence supports it.
Significant severity, high-volume uniform: Primary home for recorded delegation after counterfactual calibration — VM + confidence escalation per the matrix.
Catastrophic severity: LLM as research and context-gathering tool; mandatory human at each commitment boundary per Rule 2 — even when the model outperforms on historical metrics. Superior calibration informs harness design and review quality; it does not remove the human bind. Citations must be verified before relying on them.
Deterministic tasks at any severity: LLM not appropriate as primary mechanism. A correct algorithm replaces a probabilistic approximation with a correct answer.
What the harness provides
The difference between L3-U-high and L1-U-med for an LLM system is entirely the harness:
Each element of the harness moves the artifact’s position in the facet space. The model is the engine. The harness is the accountability infrastructure. Both matter. Only one is typically invested in — and the gap between them is where later work on AI and processes that help organisations approach truth under uncertainty will live.
That work extends harness beyond LLMs: an epistemic harness is structure around any capability — human teams, models, hybrids — that supplies tier, provenance, the commitment boundary, and feedback. Retrieval and citation are one layer; domain authority, experiment architecture, and collective quality mechanisms are others. The Enabling Intelligence and Synthesizing Knowledge series develop the full socio-technical design.
Part VIII continues with another application of immutable truth capture — science as event sourcing — in Event Sourced Science.