intent-eval-core 0.7.0__tar.gz → 0.9.0__tar.gz

{intent_eval_core-0.7.0 → intent_eval_core-0.9.0}/.gitignore

@@ -55,3 +55,9 @@ python/.venv/
 *.py[cod]
 *.egg-info/
 .pytest_cache/
+
+# lychee request cache (never committed; offline checks create none)
+.lycheecache
+
+# git-worktree isolation dirs (build-workflow artifact)
+.worktrees/

{intent_eval_core-0.7.0 → intent_eval_core-0.9.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: intent-eval-core
-Version: 0.7.0
+Version: 0.9.0
 Summary: Python (Pydantic v2) distribution of @intentsolutions/core — the canonical contracts kernel for the Intent Eval Platform. The same entity + predicate contracts the TypeScript package exposes as Zod validators + JSON Schemas, as Pydantic v2 models.
 Project-URL: Homepage, https://github.com/jeremylongshore/intent-eval-core
 Project-URL: Repository, https://github.com/jeremylongshore/intent-eval-core
@@ -80,7 +80,7 @@ print(gate.gate_decision)                 # GateDecision.pass_
 
 ## What's exported
 
-13 canonical entities + 3 predicate bodies (Blueprint B § 2 / § 7):
+15 canonical entities + 3 predicate bodies (Blueprint B § 2 / § 7 + SkillVersion DR-028 T1 + UsageEvent DR-103 D1):
 
 | Entity models | Predicate-body models |
 | --- | --- |

{intent_eval_core-0.7.0 → intent_eval_core-0.9.0}/README.md

@@ -47,7 +47,7 @@ print(gate.gate_decision)                 # GateDecision.pass_
 
 ## What's exported
 
-13 canonical entities + 3 predicate bodies (Blueprint B § 2 / § 7):
+15 canonical entities + 3 predicate bodies (Blueprint B § 2 / § 7 + SkillVersion DR-028 T1 + UsageEvent DR-103 D1):
 
 | Entity models | Predicate-body models |
 | --- | --- |

{intent_eval_core-0.7.0 → intent_eval_core-0.9.0}/pyproject.toml

@@ -8,7 +8,7 @@ name = "intent-eval-core"
 # Schemas under schemas/v1/ are the shared source of truth for both
 # distributions, so the versions move together. CI gate:
 # scripts/check_version_lockstep.py.
-version = "0.7.0"
+version = "0.9.0"
 description = "Python (Pydantic v2) distribution of @intentsolutions/core — the canonical contracts kernel for the Intent Eval Platform. The same entity + predicate contracts the TypeScript package exposes as Zod validators + JSON Schemas, as Pydantic v2 models."
 readme = "README.md"
 requires-python = ">=3.10"

{intent_eval_core-0.7.0 → intent_eval_core-0.9.0}/src/intent_eval_core/__init__.py

@@ -31,6 +31,7 @@ from .models import (
     EvidenceBundle,
     FailureTaxonomy,
     GateResultV1,
+    HumanReview,
     JudgeDecision,
     MatcherMap,
     RegressionPack,
@@ -38,17 +39,21 @@ from .models import (
     RolloutGate,
     RuntimeReceipt,
     SessionTrace,
+    SkillRefinerPassV1,
     SkillSnapshot,
+    SkillVersion,
     ToolInvocation,
+    UsageEvent,
 )
 
 # Kept in lockstep with npm package.json#version by scripts/check_version_lockstep.py
 # (CI-gated). The JSON Schemas under schemas/v1/ are the shared source of truth.
-__version__ = "0.7.0"
+__version__ = "0.9.0"
 
 __all__ = [
     "__version__",
-    # Entities (Blueprint B § 2.1 – § 2.13)
+    # Entities (Blueprint B § 2.1 – § 2.13 + SkillVersion per DR-028 T1
+    # + UsageEvent + HumanReview per ISEDC DR-103 D1)
     "EvalSpec",
     "EvalRun",
     "MatcherMap",
@@ -58,14 +63,18 @@ __all__ = [
     "RegressionPack",
     "RolloutGate",
     "SkillSnapshot",
+    "SkillVersion",
+    "HumanReview",
     "SessionTrace",
     "ToolInvocation",
     "CostRecord",
+    "UsageEvent",
     "FailureTaxonomy",
     # Predicate bodies
     "GateResultV1",
     "RetractionV1",
     "DashboardRenderV1",
+    "SkillRefinerPassV1",
     # Constants
     "GATE_RESULT_V1_URI",
 ]

intent_eval_core-0.9.0/src/intent_eval_core/_generated/human_review_schema.py

@@ -0,0 +1,69 @@
+# generated by datamodel-codegen:
+#   filename:  human-review.schema.json
+
+from __future__ import annotations
+
+from pydantic import BaseModel, ConfigDict
+
+from . import _common_schema as schema
+
+
+class HumanreviewOpenEndedHumanTrustSignalOnAnEvalrunNetNewCanonicalEntityIsedcDr103D1(
+    BaseModel
+):
+    """
+    A single human's open-ended assessment of an EvalRun — the human counterpart to JudgeDecision (whose machine verdict is the `eval-verdict/v1` predicate, NOT `judge-decision/v1`). Three ORTHOGONAL nullable channels capture the Langfuse-style signal set without crushing them into a scalar: `score_text` (open-ended NON-COMPARABLE TEXT, intentionally not an enum/number — ISEDC DR-103 C3 B6.3), `thumbs` (boolean), `annotation` (free text). Append-only ledger; single terminal state `recorded`; a revision is a NEW row citing `supersedes_id`, never a mutation. ANTI-GAMING (ISEDC DR-103 D1 B1.2, machine-enforced): a row MUST pin to a verified source — `session_trace_id` non-null OR `judge_decision_id` non-null (a lone thumb pinned to nothing is refused), AND a service-account-authored row is refused (`reviewer_is_service_account` MUST be false — the `human_reviews` ledger is a HUMAN-only signal), AND at least one of the three signal channels MUST be present. ADDITIVE / one-way-door per ISEDC DR-103 D1: this shape ships in signed @intentsolutions/core entries, so widening or merging it post-publish is irreversible — additive only. `tenant_id` is OPTIONAL-NOT-NULLABLE (DR-085 D5 / deferral-G — byte-identical to EvalSpec/EvalRun/SkillSnapshot/SkillVersion). The non-reproducible human-trust signal does NOT ride the GateResultV1-pinned EvidenceStatement; it gets its own parallel additive HumanReviewStatement (binds subject[0].digest.sha256 === input_hash only) and the `human-review/v1` predicate, which is PERMANENTLY sigstore_staging by design (ISEDC DR-103 D3 B3.2 — a human's open-ended TEXT assessment can never meet rekor_production's reproduce-the-hash bar).
+    """
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+    id: schema.Uuidv7
+    """
+    UUIDv7 PK. The `human_review_id` the human-review/v1 predicate body references.
+    """
+    eval_run_id: schema.Uuidv7
+    """
+    FK -> EvalRun.id (required). The EvalRun this human is assessing.
+    """
+    session_trace_id: schema.Uuidv7 | None = None
+    """
+    FK -> SessionTrace.id, NULLABLE. ANTI-GAMING (DR-103 D1 B1.2): a review MUST pin to a verified source — this OR `judge_decision_id` MUST be non-null. A review pinned to nothing is forgery-cost-zero and is refused at all three layers.
+    """
+    judge_decision_id: schema.Uuidv7 | None = None
+    """
+    FK -> JudgeDecision.id, NULLABLE. The machine verdict the human is agreeing-with / overriding (JudgeDecision's signed predicate is `eval-verdict/v1`). ANTI-GAMING (DR-103 D1 B1.2): this OR `session_trace_id` MUST be non-null (pin to something verified).
+    """
+    supersedes_id: schema.Uuidv7 | None = None
+    """
+    Convenience pointer to the prior HumanReview this row revises (HumanReview -> HumanReview). `null` for the first review. No record is ever mutated — a revision is a NEW row citing the prior (append-only ledger).
+    """
+    reviewer_identity: schema.ActorIdentity
+    """
+    Identity of the human reviewer (engineer email, account name). The human-trust signal's authorship anchor. DR-103 D3 B3.2 names verified reviewer identity + the pinned session as `human-review/v1`'s REAL trust criterion (the predicate being permanently staging is by design, NOT a half-finished state).
+    """
+    reviewer_is_service_account: Literal[False]
+    """
+    ANTI-GAMING (DR-103 D1 B1.2): the `human_reviews` ledger is a HUMAN-only signal, so a service-account-authored row is EXCLUDED structurally — this MUST be `false`. A service-account assessment is not a human-trust signal and cannot be laundered into the human-signal counts. Carried as a `const false` field (not omitted) so the exclusion is an explicit, audited claim on every row rather than an implicit assumption.
+    """
+    score_text: str | None = None
+    """
+    Open-ended human assessment as NON-COMPARABLE free TEXT — intentionally NOT an enum or number (DR-103 C3 B6.3). Consumers MUST NOT parse this into a scalar or aggregate it; a numeric channel that needs comparison is a SEPARATE structured field, never this string. Nullable: a review may carry only a thumb and/or an annotation.
+    """
+    thumbs: bool | None = None
+    """
+    Coarse up/down signal. `true` = thumbs-up, `false` = thumbs-down, `null` = no thumb on this review. One of the three ORTHOGONAL channels (DR-103 C3 B6.3) — never averaged with the others.
+    """
+    annotation: str | None = None
+    """
+    Free-text annotation / comment from the reviewer. Distinct from `score_text` (which is the open-ended SCORE channel) — an annotation is contextual prose, a score_text is the human's open-ended assessment. Nullable. One of the three ORTHOGONAL channels (DR-103 C3 B6.3).
+    """
+    input_hash: schema.Sha256
+    """
+    BARE 64-hex sha256 of the canonical input the human reviewed. The parallel HumanReviewStatement binds `subject[0].digest.sha256 === input_hash` (DR-103 D1 B1.3) — that pin IS the forgery cost. BARE alphabet (aligned to the entity-hash convention, e.g. SkillSnapshot.combined_sha); the in-toto wire layer strips/adds the `sha256:` prefix at the boundary.
+    """
+    created_at: schema.Rfc3339
+    tenant_id: schema.Uuidv7 | None = None
+    """
+    RESERVED multi-tenancy slot (deferral-G, bd_000-projects-k0fj; DR-085 D5), OPTIONAL-NOT-NULLABLE and byte-identical to EvalSpec/EvalRun/SkillSnapshot/SkillVersion. An absent tenant is a first-class single-tenant/global state, NEVER pooled into a cross-tenant aggregate (DR-103 D2 B2.2). v1 single-tenant reviews omit it; tenant-isolation semantics deferred to a future DR.
+    """

intent_eval_core-0.9.0/src/intent_eval_core/_generated/skill_refiner_pass_schema.py

@@ -0,0 +1,154 @@
+# generated by datamodel-codegen:
+#   filename:  skill-refiner-pass.schema.json
+
+from __future__ import annotations
+
+from enum import Enum
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, Field, confloat, constr
+
+from . import _common_schema as schema
+
+
+class Verdict(Enum):
+    """
+    Decision verdict. CLOSED enum — a row is EMITTED on accept, so verdict is the accept-record discriminator (DR-082 Q2). Widening this enum is a /v2 trigger (CISO binding: a consumer that hard-codes the closed set must never silently encounter a new verdict on an immutable row).
+    """
+
+    accept = "accept"
+    reject = "reject"
+
+
+class EvalSetRef(BaseModel):
+    """
+    Reference to the FROZEN eval-set the verdict was derived against — the entire epistemic basis of the claim. The hash pins exact content; version + lineage_id pin which published eval-set and its provenance lineage.
+    """
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+    hash: schema.Sha256Prefixed
+    """
+    sha256-prefixed digest of the frozen eval-set content.
+    """
+    version: constr(min_length=1)
+    """
+    Version identifier of the frozen eval-set.
+    """
+    lineage_id: schema.Uuidv7
+    """
+    UUIDv7 of the eval-set lineage the frozen version belongs to.
+    """
+
+
+class NamedDimensionDelta(BaseModel):
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+    id: schema.KebabSlug
+    """
+    Named-dimension identifier (kebab-slug).
+    """
+    delta: float
+    """
+    Observed delta on this named dimension.
+    """
+    non_regressed: bool
+    """
+    Whether this dimension cleared the non-regression bar at the stated alpha. For an accept verdict every entry MUST be true.
+    """
+
+
+class ReplayFidelityLevel(Enum):
+    """
+    OPTIONAL replay-fidelity claim for the refiner run (RF-0 bit-exact … RF-4 non-reproducible), mirroring gate-result/v1's iel-E11 levels.
+    """
+
+    RF_0 = "RF-0"
+    RF_1 = "RF-1"
+    RF_2 = "RF-2"
+    RF_3 = "RF-3"
+    RF_4 = "RF-4"
+
+
+class SkillRefinerPassV1InTotoPredicateBodyAttestingASkillversionClearedTheSkillRefinerAcceptanceGate(
+    BaseModel
+):
+    """
+    Predicate body of an in-toto Statement v1 whose predicateType is https://evals.intentsolutions.io/skill-refiner-pass/v1. Attests that a real SkillVersion cleared the @j-rig/refiner-core acceptance gate: significant Pareto-dominance on the behavioral dimension AND non-regression on every named dimension, by a one-sided z-test at the stated alpha. The signed body carries exactly the accept DETERMINANTS — the fields a verifier needs to independently re-derive the accept decision from immutable inputs (DR-082 Q2): strip any one and the PASS becomes an unfalsifiable 'trust me, it passed'. Provenance bulk (raw per-eval-item transcripts, human-readable diff prose, dashboard-render metadata) is OUT of the signed body, referenced by hash via the EvidenceBundle side-channel. This schema validates ONLY the predicate body — the enclosing in-toto Statement envelope (_type, subject, predicateType) is validated separately; the in-toto subject[].digest.sha256 MUST equal result_snapshot_hash without the sha256: prefix (DR-085 D4: the subject is the POST-EDIT artifact being attested; the wire-discipline mirror of gate-result/v1's input_hash binding, DR-082 Q4). Each row independently verifiable; NO top-level bundle signature per Blueprint B § 7 line 754.
+
+    The skill_version_id / parent_version_id / source_snapshot_hash / result_snapshot_hash provenance set references the SkillVersion by the kernel's EXISTING id/hash primitives (UUIDv7 + sha256-prefixed) — it does NOT define a SkillVersion kernel entity (DR-028 T1 one-way-door: SkillVersion stays a separate entity, minted elsewhere). parent_version_id is NULLABLE (DR-085 D3) so a root SkillVersion attests without forging a parent. Net-new predicate URI, ADDITIVE per § 7.2 backward-compat — no v1 contract is changed.
+
+    DR-085 D5 machine-enforced accept invariant: when verdict === 'accept', EVERY named_dimension_deltas[].non_regressed MUST be true (a 'pass' that claims acceptance while a named dimension regressed is a forgeable falsehood; enforced via the if/then below + Zod .superRefine + Pydantic model_validator).
+
+    Runs in sigstore_staging (the canonical staging SigningMode, DR-085 D4 — retires the legacy 'ln' label) until ALL FOUR DR-082 Q3 production triggers hold, AND-gated: (1) the skill-refiner-pass/v1 SPEC.md normative section lands (RESERVED→ACTIVE); (2) the DNSSEC + CAA pre-flight is green on evals.intentsolutions.io; (3) the authoring chamber's SEPARATE signing trust root is provisioned-and-live (DR-081 no-shared-root); (4) ≥1 REAL SkillVersion clears the behavioral gate on a FROZEN, signed eval-set. Until then signing_mode=staging, rekor_log_index=null.
+    """
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+    verdict: Verdict
+    """
+    Decision verdict. CLOSED enum — a row is EMITTED on accept, so verdict is the accept-record discriminator (DR-082 Q2). Widening this enum is a /v2 trigger (CISO binding: a consumer that hard-codes the closed set must never silently encounter a new verdict on an immutable row).
+    """
+    reason: list[str] = Field(..., min_length=1)
+    """
+    Structured reason entries — non-empty always, since this predicate is emitted on a real verdict. Reasons SHOULD be structured codes rather than free prose to avoid leaking skill content onto a public transparency log (CISO + GC binding).
+    """
+    refiner_strategy_id: constr(min_length=1)
+    """
+    Identifier of the RefinerStrategy that produced this verdict. REQUIRED in the signed body per the DR-028 CISO Session-7 binding (mechanism-swappable must not become mechanism-untraceable). Strategy ids are append-only-registered — a retired id is burned forever, never reused for a different mechanism (DR-082 Q5 CISO clarification).
+    """
+    skill_version_id: schema.Uuidv7
+    """
+    UUIDv7 of the accepted SkillVersion (the 14th entity per DR-028 T1). Referenced by the kernel's EXISTING UUIDv7 primitive — this predicate does NOT define a SkillVersion entity (DR-028 one-way-door).
+    """
+    parent_version_id: schema.Uuidv7 | None = None
+    """
+    DR-085 D3: NULLABLE UUIDv7 of the parent SkillVersion the accepted version was refined from — `null` for a ROOT SkillVersion (the first refinement in a lineage), so a root attests a valid pass WITHOUT forging a parent (root emission is provably zero-forgery — DR-085 D3 DevRel/CISO binding). When non-null, binds parent→child so a refiner cannot launder an unrelated skill through a forged lineage (CISO binding). REQUIRED key, but its value MAY be null (matching the SkillVersion entity's nullable parent_version_id).
+    """
+    source_snapshot_hash: schema.Sha256Prefixed
+    """
+    DR-085 D4: sha256-prefixed content hash of the PRE-EDIT SkillSnapshot — the INPUT to the refinement (re-defined to mean pre-edit input consistently with the SkillVersion entity's source_snapshot_hash; the prior 'post-edit' meaning collided semantically with the entity's meaning, the very collision D4 corrects). The POST-EDIT output is now the separate result_snapshot_hash. References the SkillSnapshot content by the kernel's EXISTING sha256-prefixed primitive.
+    """
+    result_snapshot_hash: schema.Sha256Prefixed
+    """
+    DR-085 D4: sha256-prefixed content hash of the POST-EDIT output SkillSnapshot — the RESULT of the refinement (the artifact the pass attests). The in-toto subject[].digest.sha256 for the row MUST equal this value WITHOUT the sha256: prefix (the authoring-chamber analogue of gate-result/v1's input_hash === subject digest binding, DR-082 Q4 — the subject is the post-edit artifact being attested, so it binds to result_snapshot_hash, not the pre-edit source_snapshot_hash). References the SkillSnapshot content by the kernel's EXISTING sha256-prefixed primitive.
+    """
+    eval_set_ref: EvalSetRef
+    """
+    Reference to the FROZEN eval-set the verdict was derived against — the entire epistemic basis of the claim. The hash pins exact content; version + lineage_id pin which published eval-set and its provenance lineage.
+    """
+    edit_proposal_hash: schema.Sha256Prefixed
+    """
+    sha256-prefixed hash of the EditProposal (the bounded edit-ops) that earned the pass — binds WHAT changed.
+    """
+    behavioral_delta: float
+    """
+    Observed delta on the behavioral dimension the accept gate requires significant Pareto-dominance on. A determinant of the accept decision (DR-082 Q2).
+    """
+    named_dimension_deltas: list[NamedDimensionDelta]
+    """
+    Per-named-dimension observed deltas — the non-regression surface of the accept gate. Each entry is independently re-checkable: a verifier re-runs the one-sided z-test on these deltas at the stated alpha. MAY be empty when the skill declares no named dimensions beyond the behavioral one.
+    """
+    alpha: confloat(lt=1.0, gt=0.0)
+    """
+    The significance level (α) the one-sided z-test was evaluated at — the falsifiability anchor. A 'pass' with no published alpha is an unfalsifiable assertion (DR-082 Q2). Bound in (0, 1).
+    """
+    test_statistic_kind: Literal["one-sided-z"]
+    """
+    Statistical-test family identifier. CONST for v1 — the acceptance gate is a one-sided z-test (DR-028 + DR-082 Q2). Changing the test family is a SEMANTIC change that mints /v2 (the same deltas/alpha would no longer mean the same verdict).
+    """
+    cost_record_ref: schema.Uuidv7 | None = None
+    """
+    OPTIONAL FK → CostRecord.id for cost attribution of the refiner run. Descriptive — not a determinant of accept.
+    """
+    replay_fidelity_level: ReplayFidelityLevel | None = None
+    """
+    OPTIONAL replay-fidelity claim for the refiner run (RF-0 bit-exact … RF-4 non-reproducible), mirroring gate-result/v1's iel-E11 levels.
+    """
+    signing_downgrade_reason: constr(min_length=1) | None = None
+    """
+    OPTIONAL structured reason recorded ONLY when the signing_mode was downgraded for this row (e.g. production→staging fallback). Absent on a normally-signed row.
+    """

intent_eval_core-0.9.0/src/intent_eval_core/_generated/skill_version_schema.py

@@ -0,0 +1,70 @@
+# generated by datamodel-codegen:
+#   filename:  skill-version.schema.json
+
+from __future__ import annotations
+
+from enum import Enum
+
+from pydantic import BaseModel, ConfigDict, constr
+
+from . import _common_schema as schema
+
+
+class VersionKind(Enum):
+    """
+    Load-bearing signable discriminator (DR-028 T1 line 90, P0-RATIFY-2 line 233). CLOSED enum — widening it after a signed publish is a /v2 trigger (one-way door). `edit` = a refiner-proposed forward change; `revert` = undo to a prior version's content; `restore` = re-instate an archived version. Distinct from SkillSnapshot, which has no discriminator (single terminal `created`).
+    """
+
+    edit = "edit"
+    revert = "revert"
+    restore = "restore"
+
+
+class SkillversionRefinementLineageRecordOfASkill14ThCanonicalEntityDr028T1Discriminator(
+    BaseModel
+):
+    """
+    The 14th canonical entity. SEPARATE from SkillSnapshot (which content-addressed-pins a skill's source state); SkillVersion captures the REFINEMENT LINEAGE produced by the Skill Refiner. DR-028 T1 DISCRIMINATOR resolution: `version_kind` is a load-bearing signable discriminator; `parent_version_id` references the prior SkillVersion (lineage internal, SkillVersion -> SkillVersion, NOT an FK to SkillSnapshot); `source_snapshot_hash` is a READ-ONLY REFERENCE (content hash) to the PRE-EDIT SkillSnapshot at the moment of refinement — explicitly NOT a relational foreign key (no enforced relational integrity; it is a content-hash reference). DR-085 D3 lineage-integrity correction: this entity now carries a self `content_hash` (its own post-edit content) + a `parent_content_hash` (the parent's `content_hash`, the tamper-evident lineage anchor); all three hash fields use the BARE 64-hex `sha256` alphabet, ALIGNED to `SkillSnapshot.combined_sha` (one hash alphabet platform-wide — no sha256:-prefixed vs bare mismatch). `parent_content_hash` is null iff `parent_version_id` is null (a root version forges no parent). ADDITIVE / one-way-door: this shape ships in signed @intentsolutions/core entries, so widening or merging it post-publish is irreversible — additive only. The state-machine formalism + status/signing cross-field invariants (P0-RATIFY-2 `pending_production`/`active`, `rekor_log_index iff signing_mode='production' AND status='active'`) are DEFERRED to a follow-up v0.4.0 Decision Record per DR-028 T1 binding minority constraint (lines 105/108) reconciled by DR-085 D1: "Phase C ships entity + discriminator + parent_version_id only." Predicate URI: skill-refiner-pass/v1 references this entity by id/hash (already shipped); SkillVersion does NOT define that predicate.
+    """
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+    id: schema.Uuidv7
+    """
+    UUIDv7 PK. The `skill_version_id` the skill-refiner-pass/v1 predicate body references.
+    """
+    skill_id: schema.KebabSlug
+    """
+    Logical skill identifier — stable across versions of the same skill (mirrors SkillSnapshot.skill_id).
+    """
+    version_kind: VersionKind
+    """
+    Load-bearing signable discriminator (DR-028 T1 line 90, P0-RATIFY-2 line 233). CLOSED enum — widening it after a signed publish is a /v2 trigger (one-way door). `edit` = a refiner-proposed forward change; `revert` = undo to a prior version's content; `restore` = re-instate an archived version. Distinct from SkillSnapshot, which has no discriminator (single terminal `created`).
+    """
+    parent_version_id: schema.Uuidv7 | None = None
+    """
+    HUMAN-READABLE CONVENIENCE POINTER to the PRIOR SkillVersion in the refinement lineage (SkillVersion -> SkillVersion). Lineage is internal — this is NOT an FK to SkillSnapshot (DR-028 T1 line 91). `null` for a ROOT version (the first refinement in a lineage). No record is ever mutated; a rollback is a NEW SkillVersion citing the prior (append-only event log, AC-2 / F-MK-001). DR-085 D3: this UUID is reassignable, so it is a convenience pointer ONLY — the tamper-evident integrity anchor is `parent_content_hash` (a content hash a reviewer cannot silently re-point).
+    """
+    content_hash: schema.Sha256
+    """
+    DR-085 D3: this version's OWN post-edit content hash — the integrity anchor a child SkillVersion cites as its `parent_content_hash`. BARE 64-hex `sha256`, the SAME alphabet as `SkillSnapshot.combined_sha` (one hash alphabet platform-wide). Append-only content-hash chain: child.parent_content_hash == parent.content_hash makes the lineage tamper-evident — re-pointing `parent_version_id` (a reassignable UUID) cannot forge an append-only claim a verifier can falsify (DR-085 D3 CSO/CISO binding).
+    """
+    parent_content_hash: schema.Sha256 | None = None
+    """
+    DR-085 D3: the parent SkillVersion's `content_hash` — the tamper-evident lineage integrity anchor (the real append-only proof, vs the reassignable `parent_version_id` convenience pointer). BARE 64-hex `sha256`, ALIGNED to `SkillSnapshot.combined_sha`. MUST be `null` EXACTLY when `parent_version_id` is null (a ROOT version): root emission is provably zero-forgery (null parent_version_id + null parent_content_hash) — an implementer never forges a fake parent on a root (DR-085 D3 DevRel/CISO binding). This iff is machine-enforced at all three layers (if/then below + Zod + Pydantic).
+    """
+    source_snapshot_hash: schema.Sha256
+    """
+    READ-ONLY REFERENCE (content hash) to the PRE-EDIT SkillSnapshot at the moment of refinement — the INPUT to the refinement (DR-028 T1 line 92; DR-085 D4 pre-edit-input clarification, matching the skill-refiner-pass/v1 predicate's now-pre-edit `source_snapshot_hash`). EXPLICITLY NOT A RELATIONAL FOREIGN KEY — there is no enforced relational integrity; it is a content-hash reference (the SkillSnapshot it names need not exist in any table the kernel knows about). DR-085 D3: BARE 64-hex `sha256`, ALIGNED to `SkillSnapshot.combined_sha` + `content_hash` + `parent_content_hash` (one hash alphabet platform-wide — the prior sha256:-prefixed form was the very mismatch DR-085 D3 corrects).
+    """
+    refiner_strategy_id: constr(min_length=1)
+    """
+    Identity of the RefinerStrategy that produced this version (DR-028 P0-RATIFY-5 line 270, CISO signed-manifest binding: mechanism-swappable must NOT become mechanism-untraceable). Signed in the skill-refiner-pass/v1 predicate payload, where it is also `refiner_strategy_id`.
+    """
+    created_at: schema.Rfc3339
+    created_by: schema.ActorIdentity
+    tenant_id: schema.Uuidv7 | None = None
+    """
+    RESERVED multi-tenancy slot (deferral-G, bd_000-projects-k0fj), mirroring SkillSnapshot. OPTIONAL + additive per Blueprint B § 7.2; tenant-isolation semantics deferred to a future DR. v1 single-tenant versions omit it.
+    """

intent_eval_core-0.9.0/src/intent_eval_core/_generated/usage_event_schema.py

@@ -0,0 +1,93 @@
+# generated by datamodel-codegen:
+#   filename:  usage-event.schema.json
+
+from __future__ import annotations
+
+from enum import Enum
+
+from pydantic import BaseModel, ConfigDict, conint
+
+from . import _common_schema as schema
+
+
+class Meter(Enum):
+    """
+    The metered PRODUCT-METER dimension. CLOSED pricing enum (DR-103 D1 B1.4, safest for billing). HETEROGENEOUS — distinct meters carry distinct `unit`s and MUST NOT be cross-summed (C3). FREEZES at the first production-Rekor signature; widening after is a /v2 trigger. `dashboard_render` is intentionally absent (a predicate body, not a lifecycle entity). `api_call` is the ONLY meter exempt from the verified-source binding (the leaf action has no gated parent session).
+    """
+
+    api_call = "api_call"
+    eval_run = "eval_run"
+    skill_invocation = "skill_invocation"
+    judge_decision = "judge_decision"
+    gate_evaluation = "gate_evaluation"
+    report_render = "report_render"
+
+
+class Unit(Enum):
+    """
+    The business unit `quantity` is expressed in. MIXED across meters — a `tokens` count and a `count` are NON-COMPARABLE, which is exactly why a cross-`(meter, unit)` SUM is out of contract (C3).
+    """
+
+    count = "count"
+    tokens = "tokens"
+    seconds = "seconds"
+    bytes = "bytes"
+
+
+class SourceEntityType(Enum):
+    """
+    The gated entity TYPE this row meters. Non-null for every metered (non-`api_call`) row (anti-gaming, DR-103 D1 B1.2, enforced in the allOf/if-then below); `null` permitted ONLY for `meter === 'api_call'`.
+    """
+
+    eval_run = "eval_run"
+    session_trace = "session_trace"
+    judge_decision = "judge_decision"
+    tool_invocation = "tool_invocation"
+    skill_version = "skill_version"
+
+
+class UsageeventAppendOnlyProductMeteringLedgerRow15ThCanonicalEntityDr103D1(BaseModel):
+    """
+    The 15th canonical entity. Append-only product-metering LEDGER, DISTINCT from CostRecord: CostRecord (§ 2.12) attributes PROVIDER SPEND (money in micro-USD + tokens); UsageEvent meters PRODUCT-METER COUNTS in BUSINESS UNITS (`meter` + `quantity` + `unit`). The nullable `cost_record_ref` back-reference is the SEAM proving they are SEPARATE tables — a metering row MAY cite the CostRecord that priced the same action, but the two are independently authored and never merged. Single terminal state `recorded`; the row is immutable at creation and corrections are NEW rows, never mutation (append-only). ANTI-GAMING (DR-103 D1 B1.2, machine-enforced): every metered (non-`api_call`) row MUST bind to a VERIFIED gated source — non-null `source_entity_type` + `source_entity_id` AND `source_verified === true` (the marker the runtime sets ONLY after the source session clears its quality gate) — so an arbitrary hand-supplied `quantity` with no verified provenance is refused; enforced here via `allOf`/`if-then`, in Zod via `.superRefine`, and in Pydantic via `model_validator` (one rule, three layers). The spec's original "ZERO hand-validators needed" claim is DELETED. C3 (DR-103 C3 B6.1): the `meter` enum is HETEROGENEOUS and the `unit` is mixed across meters; a consumer MUST NOT SUM/aggregate `quantity` across distinct `(meter, unit)` pairs into one scalar — rollups are per-`(meter, unit)` vectors only, enforced structurally (there is no rolled-total field on the row by construction). `quantity` is a COUNTED action, NEVER an assigned utility/weight/score (DR-103 epic Rule 1). FREEZE discipline (DR-103 binding constraint 1): the `meter` enum AND the entity name FREEZE at the first production-Rekor signature (NONE in this PR); until then both stay malleable. CMO dissent preserved (`usage_events` is a billing name) — the rename-before-production-sign debt is bound in DR-103. ADDITIVE / one-way-door: this shape ships in signed @intentsolutions/core entries, so widening or merging it post-publish is irreversible — additive only. `tenant_id` is OPTIONAL-NOT-NULLABLE (DR-103 D2 / deferral-G, bd_000-projects-k0fj), byte-identical to EvalSpec/EvalRun/SkillSnapshot/SkillVersion. A UsageEvent rides INSIDE the `extensions` field of an EvidenceStatement; it does NOT become a gate-result/v1 statement. NO production-Rekor signing is wired here.
+    """
+
+    model_config = ConfigDict(
+        extra="forbid",
+    )
+    id: schema.Uuidv7
+    """
+    UUIDv7 PK.
+    """
+    meter: Meter
+    """
+    The metered PRODUCT-METER dimension. CLOSED pricing enum (DR-103 D1 B1.4, safest for billing). HETEROGENEOUS — distinct meters carry distinct `unit`s and MUST NOT be cross-summed (C3). FREEZES at the first production-Rekor signature; widening after is a /v2 trigger. `dashboard_render` is intentionally absent (a predicate body, not a lifecycle entity). `api_call` is the ONLY meter exempt from the verified-source binding (the leaf action has no gated parent session).
+    """
+    quantity: conint(ge=0)
+    """
+    The metered count, expressed in `unit`. A non-negative COUNT of a verified action — NEVER an assigned utility/weight/score (DR-103 epic Rule 1). Any derived value is computed downstream from the ledger at read time, never stored on the row.
+    """
+    unit: Unit
+    """
+    The business unit `quantity` is expressed in. MIXED across meters — a `tokens` count and a `count` are NON-COMPARABLE, which is exactly why a cross-`(meter, unit)` SUM is out of contract (C3).
+    """
+    source_entity_type: SourceEntityType | None = None
+    """
+    The gated entity TYPE this row meters. Non-null for every metered (non-`api_call`) row (anti-gaming, DR-103 D1 B1.2, enforced in the allOf/if-then below); `null` permitted ONLY for `meter === 'api_call'`.
+    """
+    source_entity_id: schema.Uuidv7 | None = None
+    """
+    Reference (UUIDv7) to the gated source entity. Non-null for every metered (non-`api_call`) row (anti-gaming); `null` only for `api_call`. A content/identity reference, NOT an enforced relational FK — the kernel persists nothing.
+    """
+    source_verified: bool
+    """
+    The verified marker — `true` iff the runtime confirmed the source session cleared its quality gate. MUST be `true` for every metered (non-`api_call`) row (anti-gaming, DR-103 D1 B1.2): a metered row with an unverified source is REFUSED. The runtime sets this; a developer never hand-authors `true` (DR-103 D1 B1.6 honest-emit-path binding). For `api_call` it MAY be `false`.
+    """
+    cost_record_ref: schema.Uuidv7 | None = None
+    """
+    NULLABLE back-reference to the CostRecord that priced the SAME action — the SEAM proving UsageEvent and CostRecord are SEPARATE tables (DR-103 D1). `null` in the common case (most metered actions are not provider-priced). A reference, NOT an enforced relational FK.
+    """
+    recorded_at: schema.Rfc3339
+    tenant_id: schema.Uuidv7 | None = None
+    """
+    RESERVED multi-tenancy slot (DR-085 D5 / deferral-G, bd_000-projects-k0fj), byte-identical to EvalSpec/EvalRun/SkillSnapshot/SkillVersion. OPTIONAL + NOT NULLABLE — omitted from `required`; an explicit `null` is rejected (the schema $ref's uuidv7 with no null branch). A present tenant_id is an attested tenant claim; an absent one is single-tenant/global and MUST NOT be pooled into a cross-tenant aggregate (DR-103 D2 B2.2). Tenant-isolation semantics deferred to a future DR; v1 single-tenant rows omit it.
+    """