04Chapters 4.0 – 6.0 · The Three Layers

Commons, agents and a sovereign mesh

Layer I turns contribution into licensable assets. Layer II generates the demand. Layer III supplies compliant capacity at a price that demand can bear.

Layer I · 4.1 · The Commons

Proof of Contribution

Contributions arrive in five classes, and the class decides how the work is reviewed. Treating them uniformly would over-constrain the verifiable classes or under-constrain the subjective ones.

C1

Linguistic material

Speech, transcripts and dialect variants with a metadata envelope. Verified largely mechanically: signal checks, agreement between two independent transcribers, duplicate detection.

C2

Domain annotation

Labelled entities, classification and structured extraction. Verified by sampled re-annotation and inter-annotator agreement against a domain-specific threshold.

C3

Evaluation judgement

Pairwise ranking, preference elicitation and error identification over model output. Reviewed by panel agreement statistics — and the source of the flywheel signal.

C4

Translation pairs

Parallel text verified through round-trip checks and consistency scoring against a reference subset.

C5

Structured local knowledge

Procedural, regulatory, agricultural, customary and administrative knowledge as Q&A or decision trees. Verified by subject-matter review — the costliest method and the highest tariff.

Consensus path

Four states from submission to settlement

Submitted

Assigned to the peer review pool. Class is contestable during the challenge window.

Provisionally credited

Required independent reviews received. The tariff is escrowed, not transferable.

Validated

An Adjudicator confirms or escalates. On confirmation the challenge window opens.

Settled

The window closes without a sustained challenge. Credit becomes withdrawable.

Repairable

Returned with the deficiency named

  • Insufficient metadata
  • Contaminated audio
  • Missing consent fields
Non-repairable

Carries a reputation penalty

  • Fabricated provenance
  • Duplicate submission under altered metadata
  • Consent scope violation

Most rejected submissions are defective rather than fraudulent. Treating both identically would train contributors to stop after their first mistake.

Fig. 4.1Proof of Contribution — end-to-end consensus flow across five swimlanes.
4.2 – 4.3 · Adjudicators and licensing

Competence is attested per language

01

Attestation

Earned through scored adjudication against a held-out set, and degraded when a validator diverges from panel consensus. Competence in one low-resource language grants nothing in another.

02

Bond sizing

Scales with adjudication volume and inversely with recent accuracy — so the most reliable validators are also the cheapest to operate.

03

Three-way slashing

False validation returns the bond to the harmed party. Fabricated adjudication is burned. Procedural breach routes to treasury. A single penalty pool would turn slashing into a cost of doing business.

Stated plainly

Low-resource languages are low-resource partly because qualified validators are scarce. Tiered validation, cross-language panels for closely related varieties and a treasury-funded validator programme change the slope of this constraint. None of them removes it.

4.3 · Tariff classes

Three licences, one published band

Research

Non-commercial use. The lowest tariff and the narrowest redistribution rights.

Commercial training

Derived weights permitted; redistribution of the corpus itself prohibited.

Commercial redistribution

Onward sale permitted, with the highest tariff and a downstream reporting obligation.

Sensitive categories

Compute-to-data, not transfer

For health, financial behaviour or identifiable attributes, the licensee submits a job into a controlled environment inside the residency perimeter and receives aggregate results — never the corpus.

Revocation

Governs the future, honestly

Withdrawal removes a corpus from future licensing, deletes it from active storage and blocks new derived assets. It cannot recall weights already trained — so consent scopes define training use at the point of contribution.

4.4 · The revenue waterfall

Six stages before any residual

Splits shift by tariff class against a published band. Licensees negotiate against the band rather than a bespoke deal — which shortens contracting and removes the most common source of dispute.

01
Gateway and settlement costsPayable to third parties regardless of fee size.
02
Protocol treasuryOperations, audits, the validator programme and legal defence.
03
Validator poolSized to adjudication volume, not licence value.
04
Contributor poolThe largest single allocation, by contribution class and quality tier.
05
BurnApplied at the TLVY conversion step — not on the fiat fee.
06
ResidualHeld in treasury under the Chapter 10 governance path.
Layer II · 5.0 · The Agent Economy

Built for the business below the software threshold

A business sits below the threshold when acquiring, configuring and maintaining enterprise software costs more than the workflow it would automate. Many run real inventory, receivables and counterparties — they lack the margin per transaction to absorb implementation cost.

01

Heavy repetition, low value

An order confirmation or payment reminder saves a fraction of a currency unit. Per-seat software cannot earn its cost at that granularity.

02

Discontinuous attention

The owner runs most functions personally. A sentence spoken between two other tasks beats a login and a dashboard.

03

Mobile-primary devices

Any assumption of desktop access, stable bandwidth or peripherals excludes most of the category.

Can the task be done in fewer owner-seconds than the manual alternative?

Falsification test

A voice-primary interface should lift trial-to-retention more than any gain in task accuracy. Measured separately for voice-primary and text-primary cohorts within the same operator.

5.2 · Execution pipeline

Seven stages — the seventh exists because the sixth is not enough

Voice is the primary input surface. Parsing tolerates ambiguity instead of guessing, and every stage that can commit the user to an obligation passes a confidence gate set by consequence class.

Intent capture

The raw utterance is kept alongside its normalisation — it is contribution material.

Intent parsing

Action type and slots extracted; every unfilled slot is flagged.

Task planning

A sequence is drafted against a capability template.

Confidence evaluation

Scored on slot completeness and the template's history for this user.

Confirmation gate

Only ambiguous or consequential fields are shown for confirmation.

Execution

External systems bound to the template are called.

Reconciliation

Returned state is compared with intent; divergence is logged as a failure record, never silently corrected.

Fig. 5.2Agent execution pipeline with consequence-weighted confidence gates.
5.3 · The capability ladder

Five levels, gated per domain

A domain advances only after measured accuracy over a defined task volume. An operator can serve one vertical at level four and another at level two — an aggregate score would let a strong domain carry a weak one into a failure class it has not earned.

L1

Assistive output

Drafts, catalogue translation, thread summaries. Nothing is committed.

L2

Record-keeping

Transactions, inventory state and receivables. Records become the basis for later decisions.

L3

Outbound communication

Follow-ups, confirmations and dispute responses. A confirmation gate is mandatory regardless of confidence.

L4

Transaction commitment

Supplier orders, delivery schedules and settlement terms. Gated on operator bonding.

L5

Cross-system fulfilment

Coordination across systems never designed to interoperate. Presented as a future option.

5.4 · Unit economics

One variable decides it: tasks resolved without a human

Inference scales with volume and is bounded by mesh pricing. Orchestration rises in steps with template count. Settlement fees fall as ticket size rises. Support is the residual — and the component that decides whether the model works.

Support clusters in the first 60 days per enterprise. Distribution runs through aggregators — distributors, trade associations, payment providers with merchant bases — and selling level-four capability before level two has cleared is prohibited at the operator level.

Fig. 5.4Contribution margin waterfall and break-even sensitivity.
Layer III · 6.0 · The Sovereign Compute Mesh

Residency is a commercial requirement

A material share of demand cannot lawfully route certain data across certain boundaries. A counterparty unable to use cross-border inference will pay a premium for compliant inference — bounded by the cost of the alternative, which is often building nothing at all.

Three workload classes

W1

Regulated inference

Usually absolute residency rules — non-compliant providers are excluded, not penalised.

W2

Commons training

Bound by consent scopes, enforced at licence issuance rather than execution.

W3

Aggregate workloads

Least constrained, most valuable per unit — run wherever cost is lowest.

Three node classes

N1

Metropolitan clusters

High-throughput baseline capacity near the densest enterprise concentration.

N2

Regional clusters

Satisfy residency boundaries and distant contributor populations at higher unit cost.

N3

Edge aggregation

Pre-processing, normalisation and residency attestation — no frontier workloads.

Fig. 6.2Mesh topology — the residency perimeter drawn as a physical wall.
Fig. 6.3Centralised baseline against mesh-aggregated cost; the gap funds Layer II.
6.3 · Cost structure

Cheaper inside a band — not everywhere

Aggregated regional capacity reaches a lower cost per unit at moderate utilisation, because operators have already borne the hardware capital cost. At high utilisation, scheduling overhead and reliability variance invert the advantage.

Above the band, unconstrained workloads route to external capacity while the mesh is reserved for residency-constrained work — preserving the Layer II subsidy at the cost of ceding the commodity segment.