Coverage Analysis · Triodian

Coverage analysis

How much can actually be governed at the point of action?

The deterministic-versus-semantic boundary isn't one line, it's three enforceable strata and a residue. This is our estimate of how much economic activity a deterministic governance layer can reach, where the reach stops, and why the boundary is the whole point.

The headline

Roughly half to two-thirds of activity is governable. The rest is meaning, and we cede it.

~20%

Zone 1 · rule-checkable

Correctness is a pass/fail against an explicit spec.

~15%

Zone 1.5 · symbolic

Not decidable from form, but a symbol once classified.

~28%

Zone 2 · hybrid drift

Semantic core, but populations bound against a baseline.

~38%

Zone 3 · semantic residue

Meaning, intent, contested values. Ceded to human judgement.

Zones 1 + 1.5 + 2 give a combined governable surface of ~60% (50–70%). For AI-mediated activity specifically it runs to the upper end (~65–70%), because AI outputs pass through hardware chokepoints where enforcement attaches cheaply. Figures are blended GDP and employment estimates, not survey results.

The strata, defined

Governance surface is set by the kind of correctness a task requires.

ZoneCorrectness conditionWhat enforces itExamples
Zone 1
Rule-checkable
Compliance with an explicit, machine-checkable specification. A validator returns pass/fail with no interpretation.Constrained decoding, schema validators, protocol conformance, functional-safety envelopes.Payments, tax, regulated reporting, safety-critical control, logistics execution.
Zone 1.5
Symbolic
Not decidable from surface form, but fully expressible as discrete symbols once a span is classified.A tagger (outside the verdict path) plus deterministic composition: typed-span gates, speech-act interlocks, bounds registers, a session ledger over prohibited combinations.Commitments and waivers, PII/PHI emission, dose and amount bounds, re-identification mosaics.
Zone 2
Hybrid drift
The individual output needs judgement, but populations of outputs can be bounded against an attested baseline.Aggregate-drift monitoring, structure and provenance attestation at hardware chokepoints.Large-scale inference, content moderation at scale, trading/AML surveillance, insurance portfolios.
Zone 3
Semantic residue
Correctness depends on meaning, intent, context or contested values that can't be reduced to a rule, symbol lattice, or statistical envelope.Nothing deterministic. Human judgement and evaluation.Legal interpretation, original creative work, negotiation, therapy, teaching, frontier research.

The Zone 1.5 / Zone 3 boundary is the tagger: wherever a property can be reliably classified into an enumerated taxonomy, with a calibrated confidence bound and a fail-closed disposition for the unclassifiable, it's enforceable as a symbol. Wherever classification itself is irreducibly contested, it stays in Zone 3.

The revision that matters

Much of what looks "irreducibly semantic" is really symbolic once classified.

A nine-digit string is or isn't a tax file number depending on reference. "I'll waive the fee" binds the principal only in context. These properties are more than syntax but less than full meaning, and decisively: once a span is tagged, comparators, finite-state machines and bounds registers over those tags are bit-identical deterministic hardware.

The one probabilistic element, the tagger, sits outside the verdict path, in the same proposer/enforcer separation the architecture already uses. The memoryful piece, a session ledger that fires on a prohibited combination of individually-permissible spans, closes the multi-turn escalation gap that memoryless per-decision checks structurally cannot catch.

Where it lands, by sector

The decomposition cuts across sectors at the task level.

SectorBinding zoneRationale
Finance & insuranceZone 1–2 + a 1.5 commitment / typed-disclosure layerTransaction processing, reporting and AML surveillance are rule/drift-governable; underwriting and advice have semantic cores; commitments and MNPI mosaics are symbolic.
Professional & business servicesZone 2–3 + a 1.5 commitment layerLegal, consulting and design are semantic; accounting, payroll and admin are rule-governable; contractual commitments and mandatory disclaimers are symbolic.
Healthcare, education, social assistanceZone 2–3 + a 1.5 dosage / typed-disclosure layerBilling, coding and clinical protocols are rule/drift-governable; diagnosis, teaching and care are semantic; dose bounds and PHI emission are symbolic.
InformationZone 2 + a 1.5 typed / mosaic layerContent platforms and AI services are drift/attestation-governable at the boundary; editorial and creative content is semantic; re-identification mosaics are symbolic.
Manufacturing, utilities, transportZone 1–2Process control, QC, routing and safety envelopes are rule/drift-governable; design and last-mile human judgement are semantic; parameter bounds are symbolic.
GovernmentMixedBenefits, tax and procurement rules are Zone 1; policy, adjudication and management are Zone 3.

Why AI-mediated work is more governable

The chokepoints are denser in the AI stack than in the economy at large.

Structural

Constrained decoding guarantees form and grammar by masking invalid tokens at the logit level. Guarantees nothing about meaning.

Symbolic

Typed-span, speech-act, bounds and combination policy over classified spans, including the memoryful guarantee no per-decision check can provide. Silent on whether the judgement was correct.

Aggregate & representation

Population-level drift bounds plus representation-directional gating on a policy-defined feature. A different axis than distance; still not a truth guarantee.

Prevent vs. detect

Only some modalities deterministically prevent a non-compliant emission; watermarking, signed traces and attestation only detect. Enforcement value concentrates in the prevent-grade set.

The hard ceiling

Enforcement guarantees structure, symbol and population, never meaning.

Constrained decoding provably guarantees a schema is satisfied. A session ledger provably blocks a prohibited combination of individually-permissible typed spans. Neither provably guarantees the content is true, appropriate, or intended. That asymmetry is not a limitation we're working around, it's the defensible boundary that defines what the layer is for.

The symbolic stratum enforces that a commitment was made with authority and in sequence, not whether making it was wise. Where classification is calibrated, it inherits the tagger's error and handles the unclassifiable fail-closed. The interpretive core, whether an argument is sound, a diagnosis right, a design good, stays with people.

How to read these numbers

  • The Zone 1.5 share is the newest and softest estimate; it depends on classifier reliability that is still improving. Treat ~15% (10–20%) as a reasoned framework figure, not a survey result.
  • Governable ≠ automatable ≠ should be automated. This measures where enforcement can attach, not whether AI can do the work or whether governing it this way is desirable.
  • Market and sector figures are order-of-magnitude, drawn from public national-accounts and labour statistics; the strata are a spectrum, not bins, and the boundaries are genuinely contested.

The three strata map onto the enforcement ladder: rules and symbols ship deterministically; drift bounds populations; meaning is ceded, on purpose. The enforcement ladder →

Want the stratum-by-stratum detail behind these figures?

Talk to us →