512 + CVS White Paper v1.0 — April 2026 Public · Open Architecture

No Token,
No Retrieval

How 512 + CVS closes the three structural gaps
that will break the AI economy

Jon M. Watson
512 Constraint Kernel · CVS Evidence Sidecar
XRP Ledger (XRPL)
Abstract The AI economy has produced three simultaneous structural failures. Courts cannot restore authorship to vectorised content because litigation moves in years while retrieval moves in milliseconds. Traditional payment rails cannot process the transactions required to compensate rights-holders because card interchange floors are thirty cents and the correct price of a single retrieval event is six-hundredths of a cent. Data provenance standards cannot survive the embedding boundary because once text becomes a vector, provenance metadata does not travel into latent space.

These failures are not independent. They are the same failure expressed in three different systems — legal, financial, and technical. Solving any one in isolation leaves the other two intact. 512 and CVS close all three gaps at the point where the failure actually occurs: retrieval time. Not through policy. Not through contracts. Through architectural constraint, cryptographic evidence, and a settlement rail designed for machine-speed, sub-cent transactions.
Section 01

The Scale of the Problem Nobody Has Solved

In December 2023, The New York Times sued OpenAI and Microsoft, alleging that millions of its articles were used without authorisation to train large language models. As of April 2026, that case has not been decided. Fair use arguments remain unresolved. No compensation has been paid. No mechanism exists that would have prevented the use or recorded it at the moment it occurred.

That case is one of more than fifty active copyright lawsuits against AI developers in US federal courts. Getty Images v. Stability AI. The RIAA on behalf of major record labels v. Suno and Udio. Authors Guild participants v. OpenAI. Disney, Universal, and Warner Brothers v. AI video generators. The US Copyright Office released a 108-page report in May 2025 concluding that certain AI training uses likely cannot be defended as fair use. No court will decide the core fair use question in AI training until mid-2026 at the earliest. Bartz v. Anthropic settled for $1.5 billion in class relief — after years of litigation, with no mechanism to prevent the same conduct from recurring.

Structural Failure The legal system is operating precisely as designed: it adjudicates disputes after they occur. Millions of retrieval events occur before the first hearing is scheduled. By the time a court issues a ruling, the economic harm has already compounded to a scale that no damages award can meaningfully address.

Law is the right instrument for accountability. It is the wrong instrument for prevention.

Section 02

The Geometry Moment

The economic relationship between author and idea is severed at a specific, identifiable moment. That moment is embedding. Every approach that operates downstream of this moment is operating in a world where the economic relationship has already been destroyed.

Text enters an embedding pipeline. Vectors exit. The vectors carry the semantic content of the original — sometimes extraordinary content, refined over years of expertise — but they carry no author identifier, no consent record, no price. The embedding model treats authorship as irrelevant metadata. The resulting vector is stored in a database that can be queried by any system with access, at any time, for any purpose, with no record of whose thinking it is drawing on.

This is not a metaphor. It is an architectural property of every vector database currently deployed. The embedding step is where provenance is destroyed. Not at training. Not at inference. At embedding.

Approach Where It Operates Why It Fails
Legal / copyright After breach, years later Retrieval already occurred; no evidence was generated at the time
C2PA / provenance At content creation Metadata does not survive the embedding boundary into latent space
GDPR / consent Before processing (policy layer) Enforcement is months; retrieval is milliseconds
Licensed RAG agreements Between organisations The vector database has no mechanism to check a contract before returning results
512 + CVS At retrieval time, machine speed — closes the gap
At the Execution Boundary What is missing is a mechanism that operates at retrieval time, at machine speed, and enforces attribution, consent, and pricing as structural properties of the retrieval itself — not as external contracts the system may or may not be aware of. 512 is that mechanism.

Section 03

Why Courts Cannot Close This Gap

Litigation is the first instinct. It is also the most expensive and least effective tool available.

The timeline mismatch is decisive. A vector database can execute millions of retrieval operations per day. A copyright lawsuit from filing to first decision typically spans three to five years. Thomson Reuters v. ROSS Intelligence was decided after years in discovery. Bartz v. Anthropic settled for $1.5 billion — resolving claims based on conduct that had already occurred across hundreds of millions of training interactions that cannot be undone. The New York Times case, which specifically names retrieval-augmented generation in its complaint, will not see a fair use decision until 2026 at the earliest.

Litigation operates in the past tense. There is no injunction that can unlearn a vector. There is no damages award that creates the mechanism to prevent the next unauthorised retrieval. The current generation of vector databases generate no proof of what was retrieved and when. Courts cannot compel evidence that does not exist.

What CVS Changes CVS generates the receipt that litigation requires but cannot compel. Every retrieval event produces an immutable, Merkle-anchored record: which licensed vector set was accessed, under which policy, at what computed cost, with a hash of the chunks actually returned. The receipt exists before any dispute arises. When litigation occurs, it becomes a question of calculation — auditable from the receipt — rather than a question of whether any record exists at all.

Section 04

Why Payment Rails Cannot Close This Gap

Assume the attribution problem were solved. Authors could identify exactly which retrieval events used their work. Compensation was legally established. The question then becomes: through what mechanism does a fraction of a cent reach the rights-holder for each retrieval event?
The Floor Problem Credit card interchange economics establish a practical minimum: Visa and Mastercard charge approximately $0.30 plus a percentage per transaction. PayPal's micropayment tier charges five cents plus five percent. The correctly-priced compensation for a single retrieval event is $0.00063. The processing fee would exceed the transaction value by a factor of hundreds.

This is not a pricing problem that negotiation can solve. It is a structural constraint built into settlement systems designed around human decision cycles — authentication, fraud detection, chargeback risk, human review — that cannot be stripped out of the rails without rebuilding them.

Thirty years of micropayment history confirms this. IBM, Compaq, and DEC proposed micropayment systems in the mid-1990s. They failed. Not because the use case was wrong — Ted Nelson had been articulating the case for per-use payment since the 1960s — but because the transaction cost structure made sub-dollar payments economically irrational. The second-generation systems of the 2010s solved the fee problem but introduced mental transaction costs: users asked to make thousands of conscious micropayment decisions per session abandon the system.

The AI retrieval context eliminates the human friction barrier entirely. A calling agent does not experience decision fatigue from a payment event. But it does not eliminate the rail floor problem. If existing payment infrastructure cannot process $0.00063, compensation simply does not occur — regardless of legal obligation or technical capability elsewhere.

At the Execution Boundary The XRP Ledger's base transaction fee is 10 drops — 0.00001 XRP. At current valuations, a fraction of a cent. Finality: three to five seconds. Settlement is neutral: no institution can withhold it, no prior account relationship is required. The XRPL is the first settlement rail whose economics are compatible with per-retrieval compensation at the prices that correct attribution requires.

Section 05

Why Data Standards Cannot Close This Gap

The Content Authenticity Initiative and the C2PA standard represent the most serious technical attempt to restore provenance to digital content. C2PA attaches cryptographically signed metadata to content at creation time: who made it, when, with what tools. For journalism, photography, and deepfake detection, this is genuinely valuable.

The embedding boundary destroys it. When content with C2PA provenance is fed into an embedding pipeline, the embedding model processes the text — not the metadata envelope. The resulting vectors contain no reference to the C2PA manifest. The vector database stores them without it. Similarity search returns ranked vectors, not provenance records. The authentication chain is severed at the exact moment the content becomes machine-callable.

Licensed RAG agreements between organisations fail for a different reason. A contract does not operate inside a vector database. When a retrieval system receives a query, it executes similarity search. It does not check a contract before returning results. Compliance depends entirely on the integrity of the calling organisation's internal processes.

Architectural Observation

The distance between where a license lives — a legal document, an external policy layer, a C2PA metadata envelope — and where the retrieval decision is made — inside a vector database, in milliseconds — is unbridgeable by contract or standard alone.

The constraint must operate at the same layer as the retrieval. 512 operates at that layer. Nothing else currently does.


Section 06

The Three Failures Are One Architecture Problem

The copyright crisis, the micropayment impossibility, and the provenance gap are not three separate problems. They are expressions of a single architectural deficit: AI retrieval systems were built without a governance layer.

A governance layer is not a policy, a contract, or a standard applied to content before it enters the system. It is a constraint that operates at the moment retrieval is requested, determines whether retrieval is authorised, records that it occurred, and generates the payment obligation — all in the same transaction, before inference begins.

Without Governance
Vector Plane

Open to any caller. No authorisation check. No record generated. Retrieval is unconditional.

512 — Policy Plane
Constraint Layer

Retrieval is conditional. Every caller must declare intent. Every policy check runs in constant time before any vector is returned.

CVS — Evidence Plane
Receipt Layer

Every authorised retrieval generates an immutable, Merkle-anchored receipt. Asynchronous. Never blocks execution. Gaps are observable.

With a governance layer operating at retrieval time, all three failures resolve simultaneously. Attribution is not litigated — it is a receipt. Payment is not negotiated — it is a computation. Consent is not assumed — it is checked, at machine speed, before the vector database responds.


Section 07

Architecture: Three Planes, No Trust Between Them

The system is split into three independent planes. Their separation is what makes the system deployable without requiring any participant to trust any other. Each plane can fail independently without cascading.

Plane Role Trusts
Vector Plane Embedding storage and similarity search Nothing. Executes only with a valid capability token.
Policy Plane (512) Execution-time constraint enforcement The signed policy artifact. Nothing else.
Evidence Plane (CVS) Immutable receipts and settlement trigger Cryptographic proof only. Cannot be influenced by the other two planes.

If the evidence plane is temporarily unavailable, execution continues and the gap becomes observable — a missing hash segment, a discontinuity in time ordering. An unobservable failure is a defect. A visible gap is auditable. This distinction matters in regulated industries where the absence of evidence is itself evidence.


Section 08

The Reframe: Vector Set as Licensed Economic Object

Existing systems treat the vector as the unit of retrieval — a thing you fetch from a database, with authorship as metadata at best and absent at worst. A vector set is a bounded collection of embeddings derived from a specific source corpus, governed by a single signed policy artifact. This reframe — from unit of retrieval to licensed economic object — is the architectural inflection point. Everything else follows from it.

A vector set is assigned to a rights-holder. Its policy artifact is machine-readable, versioned, signed, and revocable. It travels into the retrieval system as a live constraint, not a historical record. Three properties that were previously impossible become structural.

Attribution survives embedding. Provenance lives in the set-level metadata, bound to the vector_set_id. Authorship is recoverable at any point in the retrieval chain.

Consent is revocable. The policy artifact carries a revocation epoch. When a rights-holder withdraws consent, the epoch increments. Any retrieval referencing an older epoch fails the policy check. No vectors need to be deleted. Revocation is instant and architecturally enforced.

Usage is billable. Every retrieval produces a deterministic usage calculation. The formula is transparent and reproducible by any party with access to the receipt:

Billable_Units  =  depth_multiplier × Σ ( similarity_score_i × token_weight_i )

Where:
  similarity_score_i  =  cosine similarity of chunk i to the query vector  [0 → 1]
  token_weight_i      =  tokens_in_chunk_i ÷ 1000  (normalised to kilotokens)
  depth_multiplier    =  1 + ( tokens_consumed ÷ context_budget_declared )

Total Cost  =  Billable_Units × unit_price  (set by author in policy artifact)

The depth multiplier is the economic signal. Thinking that actually grounded the output earns more than thinking that was retrieved and ignored. The author sets the price. The formula runs deterministically. No human judgment is involved.

Worked Example

Author sets unit_price = $0.0005. Calling agent declares context_budget = 2,048 tokens. Three chunks retrieved, consuming 1,024 tokens total:

ChunkSimilarityTokenstoken_weightContribution
A0.945120.5120.481
B0.783840.3840.300
C0.451280.1280.058
depth_multiplier    =  1 + (1024 ÷ 2048)  =  1.50
sum of contributions                       =  0.839
Billable Units      =  1.50 × 0.839       =  1.259
Total Cost          =  1.259 × $0.0005    =  $0.00063

One retrieval event: $0.00063. Deterministic. Auditable. Reproducible by any party with access to the receipt.

At Scale — Single Author

TimeframeRetrievalsRevenue
Per day10,000$6.30
Per month300,000$189.00
Per year3,650,000$2,299.50

A single author, at a single price point, with no subscription, no advertising, no audience required. Authors whose context has higher gravity — more retrievals, deeper usage, higher similarity scores — earn proportionally more. The economics self-sort.


Section 09

CVS: Receipts, Not Logs

After retrieval executes and billable units are computed, the evidence plane records what happened. Recording is asynchronous — it does not block inference, does not slow the hot path.

CVS Receipt Schema

receipt_id          — Unique identifier for this usage event
vector_set_id       — Which licensed set was accessed
author_id           — Who owns the set; settlement routes here
caller_id           — Who retrieved (pseudonymous or public, per policy)
timestamp           — When the retrieval occurred
billable_units      — Computed usage quantity
unit_price          — Price set by the author's policy artifact
total_amount        — billable_units × unit_price
policy_hash         — Cryptographic hash of the policy artifact applied
revocation_epoch    — Consent version active at time of retrieval
chunks_merkle_root  — Merkle root of chunks actually returned; auditable

These are receipts. Not logs. Not analytics. Not behavioral data. A receipt says: this licensed vector set was invoked, under this policy, at this time, for this cost. It does not record who the end user was, what question they asked, or how they behave over time. The receipt is the economic event. Everything else is absent by design.

What CVS Changes The absence of a receipt is itself observable — a missing hash segment, a discontinuity in time ordering. An unobservable failure is a defect. A visible gap is auditable. This is the property that makes CVS evidence-grade rather than merely logging-grade: gaps cannot be hidden, only observed.

Section 10

The Surveillance Inversion

Any system that introduces billing and usage tracking around AI retrieval triggers a legitimate concern: does this require user profiling?

The answer is structural. Surveillance capitalism monetises information consumers. This system monetises cognitive contribution — the thinking that was used, not the person who used it. CVS receipts record that a licensed set was invoked under a declared purpose at a specific time. They record the Merkle root of the chunks returned. They do not record who the end user was, what question they asked, or how they behave over time. There is no demographic layer. No behavioral graph. Those concepts do not exist in the data model and cannot be reconstructed from receipts.

The Inversion

Most systems monetise information consumers. This system monetises the thinking that was consumed.

That demand inversion is what makes the system structurally compatible with privacy regulation, enterprise deployment, and long-term trust — not because it claims to be ethical, but because user surveillance provides no economic advantage here.


Section 11

Corporate Knowledge: The Extension Case

Everything described above applies to public content. The same architecture applies to the most valuable content organisations own: internal knowledge.

Enterprise organisations are deploying RAG systems at scale against internal documentation, research archives, legal precedent libraries, and proprietary analysis. These systems face the same problem. Internal knowledge has ownership. A research team that produced a proprietary analysis has an interest in knowing when that analysis is retrieved, by which systems, for which purposes. A compliance function has an interest in knowing whether policy documents are being retrieved into agent workflows that affect regulated decisions.

Current enterprise RAG deployments cannot answer any of these questions. There is no receipt. There is no authorisation layer that makes retrieval conditional on purpose declaration. There is no mechanism for the knowledge owner to revoke access when the context of use changes.

Enterprise Implementation 512 and CVS applied to internal corporate knowledge produce: a vector set per knowledge domain, a policy artifact per domain declaring permitted use classes, a capability token per retrieval event, and a CVS receipt recording which knowledge was used, by which system, for which declared purpose, at what time. This is the audit trail that enterprise compliance, legal, and risk functions do not currently have and increasingly require — particularly in regulated industries where AI-assisted decisions carry disclosure obligations.

Section 12

The Competitive Landscape

Approach What It Closes What Remains Open
x402 / XRPL (t54.ai) Payment rail — sub-cent, machine-to-machine No attribution, no consent layer, no retrieval-conditional governance
C2PA / Content Authenticity File-level provenance at creation time Does not survive the embedding boundary
Licensed RAG datasets Contractual legitimacy for curated corpora No mechanism to make retrieval conditional on compliance
AI content licensing (AP, Reuters) Organisation-level agreements for named corpora Not per-retrieval; no receipt generated; advantages large platforms only
512 + CVS + XRPL Attribution · Consent · Payment — simultaneously

The gap that no competing approach closes: enforcement at retrieval time, at machine speed, with a deterministic receipt. x402 can prove that a payment was made. It cannot prove what was retrieved, under what consent terms, whether the rights-holder authorised the use class, or whether the revocation epoch was current at the time of access. It is a payment rail without a governance layer. That is the gap 512 and CVS close.


Section 13

Threat Surface

False Use-Class Declaration

A caller declares inference-only but intends to train on retrieved context. The CVS receipt records the declared use_class at the time of retrieval. If the caller later uses retrieved context for training, the receipt proves they declared otherwise — a cryptographic breach of contract, not an ambiguous dispute. The evidence exists before the violation propagates.

Capability Token Replay

An attacker intercepts a valid capability token and replays it beyond its intended scope. Tokens are short-lived and bound to a specific caller public key, time window, and context budget. Replay against a different caller fails the key check. Replay after expiry fails the TTL check. Replay within budget is caught by the context_budget counter, decremented on each use and recorded in the receipt.

Vector Set Poisoning

An attacker injects vectors into a legitimate author's set to inflate usage and reroute payments. Ingest is deterministic: each vector is bound to a source_hash of the original text, the embedding model, and its version. Any vector that does not reproduce from the declared source corpus fails the hash check. Poisoned vectors are detectable by any party with access to the source material and the ingest specification.

Consent Epoch Staleness

An author revokes consent, but capability tokens issued before revocation continue in circulation. Tokens include the revocation epoch current at issuance. If the epoch has since incremented, the token is invalid. The vector store checks epoch currency on every query — not just at token issuance.


Section 14

Implementation Path

No exotic components are required. The first build uses infrastructure that exists today.

Day 1 — The Constraint Gate

Deploy the policy gateway. Authors sign policy artifacts: permitted use classes, pricing terms, revocation epoch. The gateway wraps the vector database. Capability tokens are minted on policy check success. No vector is retrievable without one. Stack: Postgres for metadata and policy artifacts; any vector database (Pinecone, Weaviate, pgvector); gateway service for policy evaluation and token minting.

Day 30 — The Receipt Layer

Deploy CVS as an append-only receipt log. Every successful retrieval produces a receipt per the schema in Section 09. Receipts are Merkleised locally. No external ledger anchoring yet. Add the billable unit calculation. Authors can now see, in real time, which systems are retrieving their content, at what depth, at what computed cost.

Day 90 — Settlement and Ledger Anchoring

Deploy the payment worker. Batch receipts per author across the settlement window. Verify Merkle proofs. Execute micropayments via XRPL to author payout addresses. Anchor the settlement Merkle root to the ledger for independent finality. At this point: authors are being paid. AI systems are retrieving licensed context. The receipt chain is tamper-proof to third parties. The system is auditable, insurable, and deployable.

What Is Not Required on Day 1 External ledger anchoring. Third-party attestation. Zero-knowledge proofs. Novel cryptographic primitives. A new token or currency. Any of these can be layered in. None are prerequisites. The constraint gate alone changes the economics. The constraint gate alone changes who gets paid.

Section 15

The Game

If this architecture is deployed at scale, it does not adjust the economics of AI. It replaces them.

Training data becomes less valuable than live licensed context. A model that grounds itself in licensed, attributed, high-quality context at inference time outperforms a model that relies on frozen training data. Live retrieval from expert sources commands a premium reflecting current authority and relevance, not historical snapshot.

Scraping becomes inferior to licensed retrieval. Scraped content is anonymous, unattributed, and — under the current legal trajectory — increasingly contested. For enterprise deployments where auditability determines insurability — financial services, healthcare, legal, insurance — the licensed path is not optional. It is the only path that produces evidence of compliance.

Attribution becomes a receipt, not a lawsuit. The fight moves from "did you use my work?" — answered after years of litigation — to "here is the cryptographic receipt that records when you used it, and here is the deterministic calculation of what you owe." Disputes narrow from questions of liability to questions of calculation. Calculation is auditable. Liability is not.

Publishers and knowledge institutions become intelligence marketplaces. A financial research publisher, a specialist legal database, a medical literature archive — any institution that holds high-quality authored content becomes, overnight, a provider of licensed cognitive context. Revenue derives from the direct use of thinking, at the moment machines think with it.

The alternative — the current default — is an AI economy built on content that belongs to someone else, adjudicated in courts that cannot keep pace with the speed of retrieval, on payment rails that cannot process the transactions that compensation requires. That economy has fifty-plus lawsuits and no resolution. It has billions in legal exposure and no mechanism for prevention. It has the geometry of intelligence and no price attached to it.

The architecture described here attaches a price. At machine speed. Before retrieval begins.

Structural Implication — The Agentic Economy

The Oil That Lubricates the Agentic Economy

Every argument in this paper has assumed a human organisation as the calling party. That assumption is already obsolete. The dominant caller of the next five years is not a human. It is an agent.

Autonomous software agents — executing research pipelines, managing workflows, coordinating tool use, pulling live data from sensors and APIs — transact at a frequency and granularity that human commerce was never designed to support. A single agent-driven workflow may generate thousands of discrete service calls in a session. Each call has a cost. Each cost is sub-cent. None of them can clear a debit card minimum or a monthly SaaS billing cycle.

Human Speed — Existing Rails $0.30 + Card interchange floor per transaction.
SaaS: $99–$999/month regardless of use.
Machine Speed — Required Economics $0.00001–$0.001 Per API call. Per vector retrieval. Per IoT data pull.
Per tool invocation. Per inference event.

The gap between these two numbers is not a fee negotiation. It is a structural mismatch between payment infrastructure built for human decision cycles and an economy that now operates at machine speed. At human speed, you pay a subscription. At machine speed, you pay per act. The subscription model has no granularity. The per-act model has had no viable rail — until now.

Consider the scope of what agents need to pay for. A vector retrieval from a licensed knowledge base. A live data pull from an IoT sensor network. An API call to a weather, financial, or logistics service. A tool invocation — code execution, image processing, document parsing — priced per use. A model inference call from a specialist provider. Each of these is, economically, the same transaction: a machine paying another machine a sub-cent fee for a discrete service, in real time, without a human authorising each individual payment.

Traditional SaaS pricing absorbed this problem by hiding it inside monthly subscriptions. A developer pays $200/month for API access. The granularity of actual usage is invisible. This works at human scale, where usage is relatively predictable. It breaks at agent scale, where a single workflow can make ten thousand calls in an hour — and a single idle day makes the subscription economically wasteful. The subscription model is simultaneously too expensive during low-usage periods and too blunt during high-usage periods.

512 + CVS provides the governance and evidence layer that makes per-act machine payments legally and operationally viable. The capability token authorises the specific transaction — not the relationship, not the account, not the monthly plan. The CVS receipt proves what occurred, when, under what declared purpose, for what computed cost. The XRPL settles the obligation at sub-cent cost with three-to-five second finality, batching individual micro-obligations without collapsing the per-act granularity that makes attribution possible.

Without this infrastructure, the agentic economy faces a binary choice: agents either cannot pay at all — blocked by rail minimums that make per-call billing impossible — or they pay in subscription blocks that transfer all pricing power to platform operators and make the real economics of agent commerce opaque to both buyer and seller.

The agentic economy is not hypothetical. It is the dominant architectural direction of every major AI platform, every autonomous workflow system, and every edge compute deployment that processes data at the speed it is generated. That economy requires a payment layer that operates at machine speed, at machine granularity, with machine-readable proof of what was purchased and why.

Without it, the engine seizes. Every tool call hits a billing wall. Every API requires an account relationship. Every IoT data pull goes unpriced and unattributed. Intelligence flows, but economics does not. The value that agents create cannot be traced back to the assets that made it possible.

512 + CVS + XRPL is the lubricant the agentic economy requires. Sub-cent, cryptographically witnessed, deterministically priced, settled at ledger speed. The machinery exists. What has been missing is the constraint architecture that makes every transaction governable, every payment auditable, and every retrieval event a receipt — not a guess.

Appendix A

End-to-End Call Flow

  1. Author publishes content. Text is ingested, deterministically chunked, embedded with a pinned model, and assigned to a vector_set_id governed by a signed policy artifact. The current revocation epoch is recorded.
  2. Policy artifact is registered. The policy defines allowed use classes, pricing terms, and revocability. It is signed by the rights-holder and versioned.
  3. Caller submits retrieval intent. An agent submits the intent envelope: caller_pubkey, purpose_code, use_class, max_cost, ttl, context_budget.
  4. 512 evaluates constraints. The policy gateway validates in constant time: authorisation, purpose declaration, use class match, consent currency, price bounds, exit support. Failures surface reason codes and caller options.
  5. Capability token is minted. A short-lived token binds the caller to a specific vector set, use class, price ceiling, time window, and context budget.
  6. Vector search executes. The vector database is queried only with a valid capability token. Similarity search returns ranked chunks.
  7. Billable units are computed. The deterministic formula runs against the returned chunks.
  8. CVS receipt is emitted. Written to the append-only evidence log asynchronously. Execution is not blocked. The receipt contains provenance, policy hash, usage, and price.
  9. Inference completes. The model produces output grounded by retrieved context. No source text is copied.
  10. Settlement runs out-of-band. Receipts batched, Merkle proofs verified, micropayments executed to author payout addresses via XRPL. Settlement confirmation anchored to the ledger.
Critical Property If any step fails upstream of vector search, retrieval does not occur. If any step fails downstream, payment still resolves from receipts. Inference speed is never traded for governance or accounting.
Appendix B

Key Definitions

vector_set
A bounded collection of embeddings derived from a specific source corpus, governed by a single signed policy artifact. The licensed economic object. The unit of consent, pricing, and attribution.
policy_artifact
A machine-readable, signed document specifying the terms under which a vector set may be retrieved: permitted use classes, pricing terms, revocation epoch, exit terms. Signed by the rights-holder. Versioned.
capability_token
A short-lived cryptographic token minted by the policy gateway on successful constraint evaluation. Required for vector database access. Bound to caller identity, vector set, use class, price ceiling, time window, and context budget.
CVS_receipt
An immutable, append-only record of a retrieval event: which licensed set, under which policy, by which caller, at what computed cost, with a Merkle root of chunks actually returned. The economic event.
revocation_epoch
A versioned counter in the policy artifact. Incrementing it invalidates all outstanding capability tokens and any retrieval referencing an older epoch. Revocation without deletion. Instant and architecturally enforced.
depth_multiplier
The economic signal in the billable unit formula. Reflects how deeply an author's thinking grounded a specific inference event. Higher when more context budget is consumed, proportional to actual cognitive contribution.
intent_envelope
The structured declaration submitted by a calling agent before any retrieval is attempted. Fields: caller_pubkey, purpose_code, use_class, max_cost, ttl, context_budget. The system cannot proceed without it.