@oxy.so/contracts 1.0.0

package/dist/cjs/inference/modelDocumentation.js

@@ -0,0 +1,433 @@
+"use strict";
+/**
+ * Model documentation: what a first-party release must DECLARE, what a
+ * downstream developer may READ, and the request that ingests both.
+ *
+ * Issue #972 §12, the three items under "Future Alia model
+ * publication/compliance" that `aliaModelRelease.ts` deliberately left open:
+ * accepting the documentation set, publicising the customer-safe half of it, and
+ * preserving the metadata an EU AI Act / GPAI documentation workflow needs.
+ *
+ * ## The section's own scope is what makes the compliance claim falsifiable
+ *
+ * The issue section is titled "Future Alia model publication/compliance", so the
+ * obligations in play are the ones binding a PROVIDER of a general-purpose AI
+ * model — Oxy/Alia, for an `alia/*` release it trained or derived. Oxy's position
+ * on third-party weights is a different one (it received documentation rather
+ * than produced it) with different obligations, and nothing here claims to
+ * discharge those. Every field below names the obligation it serves; a field
+ * whose obligation could not be named is not here, and two are listed at the
+ * bottom as deliberately absent.
+ *
+ * References are to Regulation (EU) 2024/1689 (the AI Act): Article 50(2)
+ * (marking synthetic output), Article 51 (classification as a model with
+ * systemic risk), Article 53 (obligations of providers of general-purpose AI
+ * models), Article 55 (additional obligations for systemic-risk models), Annex XI
+ * (the technical documentation), Annex XII (the information for downstream
+ * providers).
+ *
+ * ## Two shapes, and the Act itself draws the line between them
+ *
+ * {@link modelGpaiDocumentationSchema} is the whole record. {@link
+ * modelDownstreamDocumentationSchema} is the subset served publicly. The split
+ * is NOT editorial taste: Annex XI is documentation a provider keeps and
+ * provides to the AI Office and national competent authorities on request, while
+ * Annex XII is information a provider MAKES AVAILABLE to downstream providers.
+ * Training compute, training time, energy consumption and the adversarial-testing
+ * report are Annex XI Section 2 and Article 55(1)(a) — the first audience — so
+ * they are in the record and not in the public projection, and
+ * `db/schema/protectedColumns.ts` says the same thing a second time at the type
+ * level.
+ *
+ * ## The conditionals are the Act's, not a convenience
+ *
+ * Article 53(2) exempts a model released under a free and open-source licence
+ * from 53(1)(a) and 53(1)(b) — the Annex XI and Annex XII sets — UNLESS it is a
+ * model with systemic risk. It does not exempt 53(1)(c) or 53(1)(d). So the
+ * copyright policy and the training-content summary are required of every
+ * release here, while the Annex XI/XII set is required of every release that is
+ * not covered by that exemption. Writing it the other way round — everything
+ * optional, checked by a human — is what makes a compliance record a field nobody
+ * filled in.
+ *
+ * ## What is deliberately NOT here
+ *
+ * **The modality and FORMAT of inputs and outputs (Annex XI §1(6), Annex XII
+ * §1(b)).** The modality half is already stored, as `inference_models`'
+ * `input_modalities` / `output_modalities`. The format half is a property of the
+ * Oxy API — one request envelope, one set of endpoints, identical for every model
+ * — so a per-model column would record the same value on every row and invite a
+ * reader to believe it could differ.
+ *
+ * **The technical means required for integration (Annex XII §1(c)).** Same
+ * reason: for a model served over the Oxy API that is Oxy's own API
+ * documentation, not a fact about the weights.
+ *
+ * **A verification finding for a release signature.** See
+ * `aliaModelRelease.ts`: whether a signature checked out is Oxy's finding about
+ * the document and not a claim the document makes, and no verifier exists yet
+ * because what signs is undecided. The ingestion path stores the signatures and
+ * the manifest as received so a verifier that lands later can check them; it
+ * records no finding, because there is none.
+ *
+ * Decided in: docs/adr/0008-catalogue-concept-separation.md, issue #972 §12.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.modelDocumentationSchema = exports.modelReleaseIngestionResultSchema = exports.modelReleaseIngestionRequestSchema = exports.modelLineDeclarationSchema = exports.modelGpaiDocumentationSchema = exports.modelDownstreamDocumentationSchema = exports.SYSTEMIC_RISK_COMPUTE_THRESHOLD_FLOPS = exports.trainingComputeFlopsSchema = exports.modelSystemicRiskTierSchema = exports.modelDistributionMethodSchema = void 0;
+const zod_1 = require("zod");
+const aliaModelRelease_1 = require("./aliaModelRelease");
+const catalogue_1 = require("./catalogue");
+const identifiers_1 = require("./identifiers");
+/* -------------------------------------------------------------------------- */
+/*  Vocabulary                                                                */
+/* -------------------------------------------------------------------------- */
+/**
+ * How a release reaches the people who use it — Annex XI §1(4) and Annex XII
+ * §1(a), "methods of distribution".
+ *
+ * TWO members, and both exist today: a release is served through the Oxy API, or
+ * its weights are published for download, or both. A third channel is a
+ * distribution decision somebody would have to make, and a closed enum gaining a
+ * member is a MINOR contract-set change the handshake surfaces (`version.ts`),
+ * which is the right amount of ceremony for it.
+ *
+ * `downloadable_weights` is also what the Article 53(2) free-and-open-source
+ * exemption is assessed against — that exemption requires the model to be
+ * "released under a free and open-source licence that allows for the access,
+ * usage, modification and distribution of the model" — so it is required even
+ * where the Annex XI set it belongs to is exempt.
+ */
+exports.modelDistributionMethodSchema = zod_1.z.enum(['oxy_api', 'downloadable_weights']);
+/**
+ * Whether this is a model with systemic risk, and on what basis — Article 51.
+ *
+ * Three states, because the two ways a model acquires the classification have
+ * different evidence and a record that flattened them could not be checked:
+ *
+ * - `not_designated` — neither presumed nor designated.
+ * - `presumed_by_training_compute` — Article 51(2): the cumulative compute used
+ *   for training exceeds 10^25 floating point operations, which the Act makes a
+ *   presumption of high-impact capabilities. The FIGURE is what creates it, so
+ *   {@link modelGpaiDocumentationSchema} requires the figure alongside this
+ *   value.
+ * - `designated_by_commission` — Article 51(1)(b): a Commission decision, ex
+ *   officio or following a qualified alert, that the model has capabilities
+ *   equivalent to the presumption. Not derivable from anything Oxy holds, which
+ *   is exactly why it is a declared value.
+ */
+exports.modelSystemicRiskTierSchema = zod_1.z.enum([
+    'not_designated',
+    'presumed_by_training_compute',
+    'designated_by_commission',
+]);
+/**
+ * Cumulative training compute in floating point operations — Annex XI §2(b).
+ *
+ * TEXT, in the same spirit as `modelEvaluationResultSchema.score` and for a
+ * sharper reason: this is a PUBLISHED figure (`4.2e25`, `2.5e26`), the numbers
+ * involved are far outside the exactly-representable integer range, and the
+ * value is never arithmetic Oxy performs on a customer's behalf. A JSON number
+ * would round it silently and make two records of one published figure compare
+ * unequal.
+ *
+ * The one comparison that IS made — against Article 51(2)'s 10^25 threshold — is
+ * a magnitude test, and `Number()` on a string this regex admits is exact enough
+ * for a magnitude test by a factor of about 10^9. The refinement that performs
+ * it is on {@link modelGpaiDocumentationSchema}.
+ */
+exports.trainingComputeFlopsSchema = zod_1.z
+    .string()
+    .max(40)
+    .regex(/^(?:0|[1-9][0-9]*)(?:\.[0-9]+)?(?:e\+?(?:0|[1-9][0-9]?))?$/, 'training compute must be a decimal or scientific figure, e.g. 4.2e25');
+/**
+ * Article 51(2)'s presumption threshold, as a number.
+ *
+ * Named rather than inlined so the refinement that applies it and the enum
+ * member that describes it (`presumed_by_training_compute`) cannot come to mean
+ * different things.
+ */
+exports.SYSTEMIC_RISK_COMPUTE_THRESHOLD_FLOPS = 1e25;
+/* -------------------------------------------------------------------------- */
+/*  The record                                                                */
+/* -------------------------------------------------------------------------- */
+/**
+ * The subset of the documentation set that is served to downstream developers —
+ * Annex XII, plus the two Article 53(1) items that are public by their own terms.
+ *
+ * Rides inside {@link modelDocumentationSchema} and inherits its version.
+ *
+ * Every field here is one a developer integrating the model needs in order to
+ * decide whether they may use it and what they must say about it: what it is for,
+ * how it is distributed, what it is built out of, where the training-content
+ * summary and the copyright policy are, and whether it carries the systemic-risk
+ * classification that puts obligations on them too.
+ *
+ * The optional members are optional for a REASON stated in the parent record's
+ * refinement — Article 53(2) — and not because a value may be skipped.
+ */
+exports.modelDownstreamDocumentationSchema = zod_1.z
+    .object({
+    /** Annex XI §1(2), Annex XII §1(a): the tasks the model is intended for. */
+    intendedTasks: zod_1.z.string().min(1).max(2000).optional(),
+    /** Annex XI §1(4), Annex XII §1(a). */
+    distributionMethods: zod_1.z.array(exports.modelDistributionMethodSchema).min(1),
+    /** Annex XI §1(5), reachable through Annex XII §1(a) ("points 1 to 5"). */
+    architecture: zod_1.z.string().min(1).max(500).optional(),
+    /** Annex XI §1(5): the number of parameters. */
+    parameterCount: zod_1.z.number().int().positive().safe().optional(),
+    /** Article 53(1)(d): the publicly available summary of training content. */
+    trainingDataSummaryUrl: identifiers_1.inferenceHttpsUrlSchema,
+    /**
+     * Article 53(1)(c): the policy for complying with Union copyright law,
+     * including the reservation of rights under Article 4(3) of Directive
+     * (EU) 2019/790. Required of every release — Article 53(2) does not exempt it.
+     */
+    copyrightPolicyUrl: identifiers_1.inferenceHttpsUrlSchema,
+    /** Article 51. */
+    systemicRisk: exports.modelSystemicRiskTierSchema,
+    /**
+     * Whether the release is under a free and open-source licence in the sense
+     * of Article 53(2). Distinct from `modelLicenseSchema.commercialUseAllowed`,
+     * which answers whether OXY may serve the model — a licence can permit
+     * commercial use and still not permit access, modification and
+     * redistribution of the weights, and it is the second question the exemption
+     * turns on.
+     */
+    freeAndOpenSourceRelease: zod_1.z.boolean(),
+})
+    .strict();
+/**
+ * The whole documentation record for one revision, as ingested.
+ *
+ * `.strict()`, because this is a compliance record arriving over the wire: a
+ * field silently dropped at the parse is a field the record does not contain,
+ * and "we accepted your documentation" would then be true of less than was sent.
+ *
+ * Not versioned on its own — it rides inside
+ * {@link modelReleaseIngestionRequestSchema} on the way in and inside
+ * {@link modelDocumentationSchema} on the way out, and inherits whichever
+ * message carries it.
+ */
+exports.modelGpaiDocumentationSchema = exports.modelDownstreamDocumentationSchema
+    .extend({
+    /** Annex XI §2(b): the computational resources used for training. */
+    trainingComputeFlops: exports.trainingComputeFlopsSchema.optional(),
+    /** Annex XI §2(b): the training time. */
+    trainingTimeHours: zod_1.z.number().positive().safe().optional(),
+    /**
+     * Annex XI §2(c): the known or ESTIMATED energy consumption. The Act asks
+     * for an estimate where the figure is not known, so absence here means the
+     * Annex XI set is exempt rather than that the number was hard to obtain.
+     */
+    energyConsumptionMwh: zod_1.z.number().nonnegative().safe().optional(),
+    /**
+     * Article 55(1)(a): the model evaluation, including adversarial testing,
+     * performed for a model with systemic risk. A pointer, like every other
+     * document reference here — the catalogue holds no report.
+     */
+    adversarialTestingReportUrl: identifiers_1.inferenceHttpsUrlSchema.optional(),
+})
+    .strict()
+    .superRefine((documentation, ctx) => {
+    // Article 53(2): the free-and-open-source exemption from 53(1)(a) and (b)
+    // does not apply to a model with systemic risk. So the Annex XI/XII set is
+    // required of everything else, and the ONE state that may omit it is a
+    // free-and-open-source release that is not designated.
+    const exempt = documentation.freeAndOpenSourceRelease && documentation.systemicRisk === 'not_designated';
+    if (!exempt) {
+        const required = [
+            'intendedTasks',
+            'architecture',
+            'parameterCount',
+            'trainingTimeHours',
+            'energyConsumptionMwh',
+        ];
+        for (const field of required) {
+            if (documentation[field] === undefined) {
+                ctx.addIssue({
+                    code: zod_1.z.ZodIssueCode.custom,
+                    path: [field],
+                    message: 'required by Annex XI unless the Article 53(2) free-and-open-source exemption applies, which it does not for this release',
+                });
+            }
+        }
+    }
+    // The presumption IS the compute figure (Article 51(2)). Declaring the tier
+    // without the figure asserts a threshold was crossed while withholding the
+    // only thing that says so.
+    if (documentation.systemicRisk === 'presumed_by_training_compute' &&
+        documentation.trainingComputeFlops === undefined) {
+        ctx.addIssue({
+            code: zod_1.z.ZodIssueCode.custom,
+            path: ['trainingComputeFlops'],
+            message: 'a systemic-risk presumption under Article 51(2) is the training-compute figure; declare it',
+        });
+    }
+    // The other direction, which is the one that matters: a release whose own
+    // declared compute is past the threshold cannot also declare that no
+    // classification applies. Without this the field pair would let the record
+    // contradict itself and still parse.
+    if (documentation.trainingComputeFlops !== undefined &&
+        documentation.systemicRisk === 'not_designated' &&
+        Number(documentation.trainingComputeFlops) >= exports.SYSTEMIC_RISK_COMPUTE_THRESHOLD_FLOPS) {
+        ctx.addIssue({
+            code: zod_1.z.ZodIssueCode.custom,
+            path: ['systemicRisk'],
+            message: 'training compute at or above 10^25 FLOP is presumed to be a model with systemic risk under Article 51(2)',
+        });
+    }
+    // Article 55(1)(a) applies to every model with systemic risk, however it
+    // acquired the classification.
+    if (documentation.systemicRisk !== 'not_designated' &&
+        documentation.adversarialTestingReportUrl === undefined) {
+        ctx.addIssue({
+            code: zod_1.z.ZodIssueCode.custom,
+            path: ['adversarialTestingReportUrl'],
+            message: 'a model with systemic risk documents its evaluation including adversarial testing (Article 55(1)(a))',
+        });
+    }
+});
+/* -------------------------------------------------------------------------- */
+/*  Ingestion                                                                 */
+/* -------------------------------------------------------------------------- */
+/**
+ * What OXY states about the model line a release belongs to.
+ *
+ * A signed release manifest carries a revision, a licence, a provenance block,
+ * evaluations, safety metadata and an artifact inventory. It carries no
+ * CAPABILITY SHEET — no modalities, no `maxContextTokens`, none of the
+ * tool/streaming flags — and every one of those is required to create a model
+ * line at all.
+ *
+ * That is not a gap in the manifest. A capability sheet is a statement about what
+ * the Oxy API will serve, which is Oxy's to make and not the signer's: the same
+ * weights behind a different gateway answer a different set of these questions.
+ * So it travels beside the manifest, like the documentation record, and the
+ * signature keeps covering exactly the document its signer wrote.
+ *
+ * Ignored when the model line already exists — a release does not edit a model.
+ * The licence and provenance in the MANIFEST are checked against the stored ones
+ * instead, because those are claims about somebody's rights rather than Oxy's own
+ * editorial choices.
+ */
+exports.modelLineDeclarationSchema = zod_1.z
+    .object({
+    displayName: zod_1.z.string().min(1).max(200),
+    description: zod_1.z.string().max(4000).optional(),
+    capabilities: catalogue_1.modelCapabilitiesSchema,
+    knowledgeCutoff: identifiers_1.inferenceDateSchema.optional(),
+    releasedOn: identifiers_1.inferenceDateSchema.optional(),
+})
+    .strict();
+/**
+ * The body of the release-ingestion request.
+ *
+ * The documentation and the capability sheet travel BESIDE the manifest rather
+ * than inside it, and that is the whole reason this wrapper exists.
+ * `aliaModelReleaseManifestSchema` is a SIGNED document: adding a field to it
+ * would change the bytes a signer covers and the version the data plane and Alia
+ * compile against, for records that are Oxy's own rather than the signer's.
+ * Keeping them separate means the signature still covers exactly what it covered.
+ *
+ * A signer that later chooses to cover the documentation too can: it would
+ * become a second signed document with its own manifest, which is a contract
+ * addition rather than a change to this one.
+ */
+exports.modelReleaseIngestionRequestSchema = zod_1.z
+    .object({
+    /** See `version.ts`: an ingestion payload is a whole message on the wire. */
+    schemaVersion: zod_1.z.literal(1),
+    manifest: aliaModelRelease_1.aliaModelReleaseManifestSchema,
+    gpaiDocumentation: exports.modelGpaiDocumentationSchema,
+    model: exports.modelLineDeclarationSchema,
+})
+    .strict();
+/**
+ * What ingestion reports back.
+ *
+ * COUNTS for the artifacts and signatures rather than echoing them: the caller
+ * sent them and the interesting fact is that all of them landed. Echoing a
+ * signature would also make this response a place a credential-shaped value gets
+ * logged, for no gain.
+ *
+ * No verification field — see this module's header, and `aliaModelRelease.ts`.
+ */
+exports.modelReleaseIngestionResultSchema = zod_1.z
+    .object({
+    /** See `version.ts`: served on its own, so it is versioned. */
+    schemaVersion: zod_1.z.literal(1),
+    releaseId: zod_1.z.string().min(1).max(128),
+    modelId: identifiers_1.modelIdSchema,
+    revision: identifiers_1.modelRevisionLabelSchema,
+    reference: identifiers_1.modelReferenceSchema,
+    /** Whether this request created the release, or found it already ingested. */
+    outcome: zod_1.z.enum(['ingested', 'already_ingested']),
+    artifactCount: zod_1.z.number().int().positive().safe(),
+    signatureCount: zod_1.z.number().int().positive().safe(),
+    evaluationCount: zod_1.z.number().int().nonnegative().safe(),
+    ingestedAt: identifiers_1.inferenceTimestampSchema,
+})
+    .strict();
+/* -------------------------------------------------------------------------- */
+/*  The customer-safe documentation view                                      */
+/* -------------------------------------------------------------------------- */
+/**
+ * The documentation for ONE revision, as a downstream developer reads it.
+ *
+ * Revision-scoped, and that is the point of it existing beside
+ * `modelCatalogueEntrySchema`. The catalogue entry carries the documentation of
+ * whichever revision is CURRENT, so a customer who pinned
+ * `<publisher>/<model>@<revision>` — which the catalogue invites, and which the
+ * immutability trigger on `inference_model_revisions` exists to make meaningful —
+ * had no way to read the model card, evaluations or safety metadata of the
+ * revision they are actually calling. A model card that only describes the
+ * newest weights is the exact conflation ADR 0008 separates revisions to prevent.
+ *
+ * `license` and `provenance` are the MODEL's, repeated here rather than linked,
+ * for the same reason `modelCatalogueEntrySchema` repeats its fields: a
+ * projection that nests the operational descriptors is one accident of nesting
+ * away from serving an internal identifier.
+ */
+exports.modelDocumentationSchema = zod_1.z
+    .object({
+    /** See `version.ts`: this is a public response shape. */
+    schemaVersion: zod_1.z.literal(1),
+    modelId: identifiers_1.modelIdSchema,
+    revision: identifiers_1.modelRevisionLabelSchema,
+    /** The exact string a customer pins. */
+    reference: identifiers_1.modelReferenceSchema,
+    /** Whether a bare `<publisher>/<model>` resolves to this revision today. */
+    isCurrentRevision: zod_1.z.boolean(),
+    releasedAt: identifiers_1.inferenceTimestampSchema,
+    retiredAt: identifiers_1.inferenceTimestampSchema.optional(),
+    modelCardUrl: identifiers_1.inferenceHttpsUrlSchema.optional(),
+    /**
+     * The digest of the served artifact, where Oxy hosts the weights.
+     *
+     * Customer-safe, deliberately: it is the one field on this view that lets a
+     * developer check that the weights they were handed are the weights the
+     * documentation describes, and a digest discloses nothing but the identity of
+     * bytes Oxy is already serving them.
+     */
+    artifactDigest: identifiers_1.sha256DigestSchema.optional(),
+    license: catalogue_1.modelLicenseSchema,
+    provenance: catalogue_1.modelProvenanceSchema,
+    evaluations: zod_1.z.array(catalogue_1.modelEvaluationResultSchema).default([]),
+    safety: catalogue_1.modelSafetyMetadataSchema.optional(),
+    /** Absent for a revision with no documentation record — i.e. every one Oxy did not release. */
+    gpai: exports.modelDownstreamDocumentationSchema.optional(),
+})
+    .strict()
+    .superRefine((documentation, ctx) => {
+    // The same check `modelRevisionSchema` makes, and it is load-bearing for a
+    // different reason here: this view exists so a customer can read the
+    // documentation of the revision they PINNED, so a reference that resolves
+    // elsewhere would attach a model card to weights nobody is calling.
+    if (documentation.reference !== `${documentation.modelId}@${documentation.revision}`) {
+        ctx.addIssue({
+            code: zod_1.z.ZodIssueCode.custom,
+            path: ['reference'],
+            message: 'reference must be exactly <modelId>@<revision>',
+        });
+    }
+});

package/dist/cjs/inference/money.js

@@ -0,0 +1,188 @@
+"use strict";
+/**
+ * Money and usage units for the inference contracts.
+ *
+ * The non-negotiable invariant this file exists to make structural: **customer
+ * charges never use floating-point values as the financial source of truth.**
+ * A JS `number` cannot represent `0.1 + 0.2` exactly, and an inference ledger
+ * adds millions of small amounts, so a float total is wrong by construction
+ * rather than by accident.
+ *
+ * Every amount and every price is one representation: {@link exactDecimalSchema},
+ * an exact decimal STRING at the scale ADR 0009 declares. Amounts are not
+ * integer minor units, and that is the ADR's decision rather than an oversight:
+ * one token costs several orders of magnitude less than one cent, so rounding
+ * per request would make a customer's bill depend on how their client chunked
+ * its work. Rounding happens ONCE, at the invoice boundary, and is itself a
+ * ledger entry.
+ *
+ * A string is also what the driver hands back — `postgres.js` decodes `NUMERIC`
+ * as a string — so keeping it a string on the wire means accidental JS
+ * arithmetic fails loudly instead of silently losing precision. Money
+ * arithmetic happens in SQL or in a decimal type, never in a JS `number`.
+ *
+ * Units are carried separately from money in every shape: a receipt says both
+ * "204 output tokens" and "0.003060000000 USD", and neither is derived from the
+ * other at read time. That separation is what lets a price version change
+ * without rewriting settled history.
+ *
+ * Decided in: docs/adr/0009-usage-reservation-and-settlement.md.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.unitPriceSchema = exports.usageSourceSchema = exports.USAGE_SOURCES = exports.usageQuantitySchema = exports.usageUnitSchema = exports.USAGE_UNITS = exports.moneySchema = exports.exactDecimalSchema = exports.INFERENCE_MONEY_SCALE = exports.currencyCodeSchema = void 0;
+const zod_1 = require("zod");
+/* -------------------------------------------------------------------------- */
+/*  Money                                                                     */
+/* -------------------------------------------------------------------------- */
+/** ISO 4217 alpha-3 currency code, e.g. `USD`. */
+exports.currencyCodeSchema = zod_1.z
+    .string()
+    .regex(/^[A-Z]{3}$/, 'currency must be an ISO 4217 alpha-3 code');
+/**
+ * The declared fractional scale of every amount and price in this contract, and
+ * of the `NUMERIC` columns the ledger stores them in.
+ *
+ * Twelve digits is sub-minor-unit precision by a wide margin: at $3 per million
+ * input tokens, one token costs `0.000003000000`, which this scale represents
+ * exactly. Amounts are compared and summed NUMERICALLY, never as text — `3.0`
+ * and `3.000000000000` are one amount written two ways.
+ */
+exports.INFERENCE_MONEY_SCALE = 12;
+/**
+ * An exact non-negative decimal, carried as a STRING so no parse step can turn
+ * it into a float on the way past. Up to 18 integer digits and
+ * {@link INFERENCE_MONEY_SCALE} fractional digits.
+ *
+ * Non-negative: direction is carried by the SHAPE — a receipt debits, a refund
+ * credits — so a stray sign can never silently invert an entry.
+ *
+ * No exponent form: `1e-6` and `0.000001` are the same number, but only one of
+ * them survives a naive string comparison, a cache key or a log grep intact.
+ *
+ * Branded, so a bare `string` is not assignable and an amount cannot arrive
+ * from string concatenation that was never checked. Producers construct one
+ * with `exactDecimalSchema.parse(value)`.
+ */
+exports.exactDecimalSchema = zod_1.z
+    .string()
+    .regex(/^(?:0|[1-9][0-9]{0,17})(?:\.[0-9]{1,12})?$/, 'must be an exact non-negative decimal string without an exponent')
+    .brand();
+/**
+ * An amount of money: the exact decimal plus the currency it is in.
+ *
+ * `.strict()` so a payload carrying a convenience float beside the exact value
+ * (`{ amount: '18.06', amountFloat: 18.06 }`) is REJECTED rather than stripped.
+ * A stripped float is the more dangerous outcome: it disappears silently here
+ * and survives in the producer, where it is the value somebody eventually
+ * displays.
+ */
+exports.moneySchema = zod_1.z
+    .object({
+    amount: exports.exactDecimalSchema,
+    currency: exports.currencyCodeSchema,
+})
+    .strict();
+/* -------------------------------------------------------------------------- */
+/*  Usage units                                                               */
+/* -------------------------------------------------------------------------- */
+/**
+ * The closed set of units inference is metered in.
+ *
+ * **The units PARTITION a request: every unit counts material no other unit
+ * counts.** `cached_input_tokens` is not part of `input_tokens`, and
+ * `reasoning_tokens` is not part of `output_tokens` — they are siblings, not
+ * subsets. A request whose 10 000-token prompt was served 9 000 tokens from
+ * cache is reported as `input_tokens: 1000` beside `cached_input_tokens: 9000`,
+ * never as `input_tokens: 10000` beside it.
+ *
+ * That belongs to the definition rather than to a convention somewhere else,
+ * because settlement applies a price to EVERY reported unit and sums them
+ * (`inferenceLedger.service.ts`'s `computeCharge`). Under the partition rule
+ * that sum IS the request's cost, and a cached token can carry its own — lower
+ * — price. Under the nested reading the same sum charges the cached and
+ * reasoning tokens twice: once inside their parent and once on their own line.
+ * It fails silently, because every total still looks plausible and the receipt
+ * is still internally consistent, and on a reasoning model the reasoning tokens
+ * can dominate the completion, so the error is not marginal.
+ *
+ * **Every OpenAI-compatible provider reports the other way round**:
+ * `prompt_tokens` INCLUDES `prompt_tokens_details.cached_tokens`, and
+ * `completion_tokens` INCLUDES `completion_tokens_details.reasoning_tokens`.
+ * Normalising is the data plane's job and it is subtraction:
+ *
+ * ```text
+ * input_tokens  = prompt_tokens     - prompt_tokens_details.cached_tokens
+ * output_tokens = completion_tokens - completion_tokens_details.reasoning_tokens
+ * ```
+ *
+ * No refinement in this package can enforce it, and saying so is part of the
+ * rule: a nested report and a disjoint one are the same four non-negative
+ * integers, so no predicate over a single report can tell them apart. The two
+ * structural guards that DO exist — refining `cached <= input` and
+ * `reasoning <= output`, or deriving the parents instead of reporting them —
+ * both encode the nested reading, which is the one this rule rejects. What IS
+ * enforceable is the arithmetic that depends on the rule, and that is where the
+ * enforcement lives — `inferenceLedger.service.test.ts` prices a report in
+ * which cached and reasoning tokens are both non-zero and asserts the exact
+ * total, which the nested reading cannot produce.
+ *
+ * Where the public surface has to speak a nested dialect, the sum is put back
+ * at the boundary rather than the internal reading being bent to it
+ * (`routes/inferenceEdge.ts` renders `prompt_tokens` as
+ * `input_tokens + cached_input_tokens`).
+ *
+ * Time is carried in integer MILLISECONDS rather than seconds so that no unit
+ * quantity is ever fractional: a 12.5-second transcription is `12500`, exactly,
+ * and the "units are integers" rule holds for every modality instead of holding
+ * for tokens and being quietly broken by audio.
+ */
+exports.USAGE_UNITS = [
+    'input_tokens',
+    'cached_input_tokens',
+    'output_tokens',
+    'reasoning_tokens',
+    'requests',
+    'images',
+    'audio_input_milliseconds',
+    'audio_output_milliseconds',
+    'video_milliseconds',
+    'characters',
+    'embeddings',
+];
+exports.usageUnitSchema = zod_1.z.enum(exports.USAGE_UNITS);
+/**
+ * A metered quantity of ONE unit. Never money — a quantity carries no price and
+ * no currency, so a consumer cannot mistake a token count for an amount owed.
+ */
+exports.usageQuantitySchema = zod_1.z
+    .object({
+    unit: exports.usageUnitSchema,
+    quantity: zod_1.z.number().int().nonnegative().safe(),
+})
+    .strict();
+/**
+ * Where a metered quantity came from.
+ *
+ * Kept explicit because the three are not interchangeable when a charge is
+ * disputed: `provider_reported` is the upstream's own count, `oxy_measured` is
+ * counted by the platform (streamed bytes, wall-clock milliseconds), and
+ * `estimated` is a reconstruction used when a provider returned no usage at
+ * all. An estimate that is indistinguishable from a reported number is an
+ * estimate nobody can later reconcile or refund against.
+ */
+exports.USAGE_SOURCES = ['provider_reported', 'oxy_measured', 'estimated'];
+exports.usageSourceSchema = zod_1.z.enum(exports.USAGE_SOURCES);
+/**
+ * A price for one unit, as `amount` per `per` units — `per` because a price
+ * quoted per single token would need more fractional digits than it is worth
+ * ("$3.00 per 1000000 input_tokens" is how every provider quotes it, and how
+ * every customer reads it).
+ */
+exports.unitPriceSchema = zod_1.z
+    .object({
+    unit: exports.usageUnitSchema,
+    amount: exports.exactDecimalSchema,
+    per: zod_1.z.number().int().positive().safe(),
+    currency: exports.currencyCodeSchema,
+})
+    .strict();