The Facets Composed

We have, across this part, taken trustworthiness apart into six facets — tier, commitment status, authority weight, atomicity, uncertainty, and processor disposition — and insisted each time that they are independent. Now we put them back together. Because no real decision ever asks about a facet in isolation. It asks a single, practical question:

How much may I trust this, for what I am about to do with it?

The facets are the inputs. Trust — and its inverse, risk — is the output. This article is about how the inputs compose into the output: what gates what, what caps what, what compounds along a chain, and why the answer always depends not only on the artifact but on the stakes of the use.

        Epistemic tier        "How strong is the claim?"        (tiers 2–13)
        Commitment status     "Did an authority commit?"        (committed / not)
        Authority weight      "How much does that authority count here?"
        Atomicity             "How far from the substrate?"     (L0–L3)
        Uncertainty           "How fit is the artifact for use now?"  (U-low / med / high)
        Disposition           "How faithfully is the processor reading?"  ([processor fidelity](/terms/processor-fidelity/))
                                          │
                                          ▼
                          Trust / risk for THIS decision

What is not a facet of epistemic trustworthiness — and a stress test

The six facets are deliberately narrow: they answer how much may I trust this for an epistemic use? — objective strength, accountability, standing, structure, artifact fitness-for-use, and processor fidelity. Many properties of statements matter to organisations and to public life but do not compose into that judgment. They belong elsewhere — in an ontology of relationships between claims, in normative Values, or in perception and reaction layers we have not developed here.

Contested and disputed claims are the clearest example. When two sides in a conflict assert contradictory propositions — one tier-3 and established, the other tier-13 or tier-9; or both tier-8 with only oral assertion behind each — the honest response is separate tier on each claim plus relationships in the graph: contradicts, refutes, supersedes, open question on. How widely a claim is accepted, believed, or disputed in a population is subjective acceptance — politically and socially vital, but not a sixth input to epistemic trust. A tier-3 finding remains tier-3 whether forty or forty million people reject it; the rejection is data about people and communities, not about the claim’s evidential warrant. We considered naming contestation or dispute status as a facet; we dropped it because it would conflate popularity with strength and duplicate what an ontology should express as edges and community-scoped metadata, not as trust composition.

That boundary opens room for other statement relationships the knowledge layer may need later — without promoting them to facets: perceived credibility, affective reaction, adoption by a group, allegiance to a narrative, exposure in media. Useful for strategy, governance, and psychology; orthogonal to the trust function this part defines.

Stress test — candidates considered and where they land:

For epistemic trust and risk — the question this part exists to answer — the list is exhaustive at this level of abstraction. New axes that appear in practice (model card fields, bias scores, community labels) should map onto these six, onto tier and provenance, or onto relationships and perception outside the trust function — not multiply facets until the composition rules break. If a proposed seventh axis does not change how much may I rely on this for this stakes class after tier, commitment, weight, atomicity, uncertainty, and disposition are known, it is not a facet here.

The classification space

Atomicity and uncertainty together define a grid (L–U shorthand), and the grid tells you at a glance how usable an artifact is:

                    U-low          U-med          U-high
L0 (atomic)         Ideal          Acceptable     Degraded
L1 (inv. agg.)      Authoritative  Standard       Avoid
L2 (compound)       Usable         Operational    Exploratory only
L3 (rendered)       Reporting      Standard UI    Prohibited

Three cells make the point sharply. L0-U-high (Degraded) is the uncalibrated sensor: structurally atomic, perfectly committed, and still untrustworthy because its fidelity is poor — the atom is correctly identified, its accuracy is not. L1-U-low (Authoritative) is the real-time, single-writer balance: one step from the atoms, but lossless, current, and ordered — about as trustworthy as a derived artifact gets. L3-U-high (Prohibited) is the dashboard that mixes data of different vintages without disclosure — it looks authoritative and is inconsistent in ways the reader cannot detect. That cell should not exist in a well-designed system.

                 COMMITTED                         UNCOMMITTED
                 Atomic   Aggregate                Aggregate   Rendered
                 (L0)     (L1/L2)                  (L1/L2)     (L3)

Class A        ✓        ✓ + reservation + U≤med     ✗           ✗
(irreversible)          + chain to atomic

Class B        ✓        ✓ + chain, U≤med            ✗           ✗
(audit)

Class C        ✓        ✓ + staleness SLA           ✓ + SLA     ✗
(operational)

Class D        ✓        ✓                           ✓           ✓ (if U disclosed)

How the facets compose

The matrix already encodes most of the composition rules. Stated plainly, they are four:

Commitment gates. For accountable uses — Classes A and B — commitment status is a gate, not a dial. An uncommitted derivation, however fresh and accurate, has no authority behind it; when it is wrong, there is nothing to hold accountable. No amount of low uncertainty buys an uncommitted artifact into Class A. This is the binary facet doing binary work.

Uncertainty and atomicity cap. Within the committed region, uncertainty and atomicity set ceilings. Class A is capped at L0/L1 and at U-med-with-reservation; it may never consume U-high regardless of commitment, and never L2/L3 regardless of freshness. The deeper the aggregation and the higher the uncertainty, the lower the class of use that may safely depend on it.

Disposition discounts. Low processor fidelity — fatigue, promotion inflation, prevention deflation, unsafe room — discounts the effective trust available from otherwise acceptable artifacts. A Class-C-eligible L2-U-med input consumed by a rubber-stamping approver does not deliver Class C judgment quality regardless of artifact coordinates. Record disposition signals where known (Disposition and Processor Fidelity); design bind mechanisms from Strengthening the Commitment Boundary to raise fidelity when severity warrants it.

Tier travels by weakest link. Along a chain of reasoning, the tier of the conclusion is the tier of its weakest premise. You cannot reason a tier-10 hunch into a tier-3 conclusion. Tier composes pessimistically, and honestly.

Uncertainty compounds multiplicatively. Along a chain of processing, information efficiency multiplies: three lossy steps at η = 0.9 leave about 0.73 of the original information. Distance from ground truth accumulates with every probabilistic or cross-stream step. This is not a heuristic; the next article shows it is a theorem.

The output: a risk judgment scaled to the stakes

Put together, the facets yield not a label but a judgment: given a claim of this tier, committed (or not) by an authority of this weight, sitting at this atomicity and this uncertainty, read by a processor of this fidelity — how much may a use of these stakes rely on it?

The decisive word is stakes. The thresholds in the matrix are not universal constants; they tighten as the cost of being wrong rises. An L2-U-med artifact is fine for a reversible operational nudge and unthinkable for an irreversible, high-severity commitment. The same six facets, the same artifact — different verdicts, because the question included the consequences.

Throughout this part, claims about uncertainty — that processing can only lose information, that losses compound multiplicatively, that quality is ceilinged at the point of observation — have been presented as well-motivated intuitions. They are not merely intuitions. They are consequences of information theory (What Information Theory Says), and the next article proves them.

Part IV ends when that proof is complete. Part V runs 34 → 19 → 20 → 35 → 36: false vs genuine uncertainty, when to reinforce, quality, legitimacy, then meritocracy as overlap (Meritocracy — The Sweet Spot).