universal-dev-standards 5.17.0 → 6.1.0

package/bundled/ai/standards/model-provenance.ai.yaml

@@ -0,0 +1,297 @@
+# Model Provenance Policy Standards - AI Optimized
+# Sources:
+#   v1.0.0 — XSPEC-255 Model Provenance Policy (acceptable model source as an enforced axis)
+#            DEC-075 (neutral mechanism / policy belongs to customer)
+#            DEC-063 (regulated / data-sovereignty customers)
+
+id: model-provenance
+title: Model Provenance Policy Standards (Acceptable Model Source as a Governed, Fail-Closed Axis)
+version: "1.0.0"
+status: Active
+tags: [provenance, model-source, geopolitics, data-sovereignty, governance-gate, fail-closed, audit, allowlist, denylist, compliance]
+created: 2026-06-17
+updated: 2026-06-17
+
+spec_ref: XSPEC-255   # Model Provenance Policy complete spec
+
+references:
+  - XSPEC-255   # v1.0.0 source — Model Provenance Policy
+  - XSPEC-193   # Sibling governance-gate profile — licensing (shared skeleton)
+  - XSPEC-256   # Sibling governance-gate profile — correctness
+  - DEC-075     # Domain-neutral mechanism (policy content belongs to customer/pack)
+  - DEC-063     # Legal & compliance — regulated / sovereignty-sensitive customers
+  - DEC-064     # Customer IP isolation (per-customer policy scoping)
+  - DEC-066     # Telemetry-driven evolution — audit/telemetry allowlist
+  - DEC-076     # Self-serve north star — provenance = a self-serve ceiling
+  - XSPEC-254   # Positioning pre-mortem (cost-structure / origin axis context)
+
+summary: |
+  Model selection is not only "cost x quality" — there is a third axis: whether the model's
+  SOURCE is acceptable to the customer (geopolitics / data sovereignty / regulation). A
+  regulated or geopolitically sensitive buyer may be unable to use models from a given
+  jurisdiction; the strictest reject even locally-run open weights of that origin. This
+  standard makes model provenance a first-class, enforced, auditable property: declare an
+  acceptable-source policy, block non-compliant model choices fail-closed, and emit
+  auditable evidence ("this delivery used no models of the denied origin").
+
+  It is the PROVENANCE (source) member of the governance-gate family — alongside
+  license-compliance (XSPEC-193, licensing) and verification-oracle (XSPEC-256, correctness)
+  — sharing the same shape: registry/check -> fail-closed gate -> audit evidence ->
+  human-escalation ceiling. The three are profiles of one mechanism over three axes:
+  licensing / provenance / correctness.
+
+  REQ-001~007: a model origin registry (model -> vendor -> jurisdiction/origin class); a
+  provenance policy schema (allow/deny origins, local-weights distinction, mode,
+  unknown-origin fail-closed); fail-closed enforcement at every model-selection point;
+  auditable provenance evidence; reference policy packs (content customer-owned);
+  unknown-origin fail-closed; and policy content sovereignty (mechanism neutral, policy
+  owned by the customer).
+
+scope:
+  applies_to:
+    - Any system that selects an LLM/model where the customer constrains acceptable model source
+      (regulated / government / geopolitically sensitive buyers)
+    - Model-selection points: per-agent config, routing decisions, BYOK
+    - Provenance audit reports proving no denied-origin model was used
+  excludes:
+    - The policy CONTENT itself (which origins are acceptable) — owned by the customer / a compliance pack
+    - Enforcement-engine wiring (e.g. VibeOps provider abstraction / router gate) — downstream adoption concern
+    - Cost/quality model right-sizing — that is model-selection; provenance is a constraint applied first
+
+origin_registry:
+  description: |
+    Each known model is classified by vendor -> jurisdiction -> origin class. Like the
+    license-compliance license DB, it is maintainable and extensible by the ecosystem/customer.
+  origin_classes:
+    - western-frontier   # e.g. Claude (Anthropic/US), GPT/o-series (OpenAI/US), Gemini (Google/US)
+    - western-open       # e.g. Llama (Meta/US), Mistral (Mistral/EU), Gemma (Google/US)
+    - china-origin       # e.g. DeepSeek (DeepSeek/CN), MiniMax (MiniMax/CN), Qwen (Alibaba/CN)
+  example_entries:
+    - { model: Claude, vendor: Anthropic, jurisdiction: US, origin_class: western-frontier }
+    - { model: GPT, vendor: OpenAI, jurisdiction: US, origin_class: western-frontier }
+    - { model: Gemini, vendor: Google, jurisdiction: US, origin_class: western-frontier }
+    - { model: Llama, vendor: Meta, jurisdiction: US, origin_class: western-open }
+    - { model: Mistral, vendor: Mistral, jurisdiction: EU, origin_class: western-open }
+    - { model: Gemma, vendor: Google, jurisdiction: US, origin_class: western-open }
+    - { model: DeepSeek, vendor: DeepSeek, jurisdiction: CN, origin_class: china-origin }
+    - { model: MiniMax, vendor: MiniMax, jurisdiction: CN, origin_class: china-origin }
+    - { model: Qwen, vendor: Alibaba, jurisdiction: CN, origin_class: china-origin }
+
+policy_schema:
+  description: |
+    Per-project/per-customer provenance policy. Mechanism only — concrete values are
+    customer/pack-owned (DEC-075).
+  fields:
+    allow_origins: "list of origin classes that are permitted"
+    deny_origins: "list of origin classes that are blocked"
+    china_local_weights: "deny | allow_local_only — distinguishes foreign open weights run locally; strict buyers deny"
+    mode: "block | warn | require_human"
+    unknown_origin: "block (fail-closed default) | warn | require_human"
+  example: |
+    {
+      "allow_origins": ["western-frontier", "western-open"],
+      "deny_origins": ["china-origin"],
+      "china_local_weights": "deny",
+      "mode": "block",
+      "unknown_origin": "block"
+    }
+
+principles:
+  - id: P-1
+    name: Provenance is a Third Axis
+    description: |
+      Model choice is cost x quality x acceptable source. The source axis MUST be a
+      first-class constraint, not an afterthought, for regulated/sovereignty-sensitive buyers.
+
+  - id: P-2
+    name: Registry First
+    description: |
+      Every model MUST carry vendor/jurisdiction/origin class before a policy can act on it.
+      An unclassified model is "unknown origin" and falls under fail-closed handling (P-3).
+
+  - id: P-3
+    name: Fail-Closed
+    description: |
+      A denied or unknown origin MUST block or escalate. The check MUST NOT silently pass —
+      mirroring the license-compliance blocklist and verification-oracle non-reproduction block.
+
+  - id: P-4
+    name: Evidence-Based
+    description: |
+      Every generation MUST record the model used and its origin class, so a compliance report
+      ("no denied-origin model used") is producible with a reproducible trace and audit id.
+
+  - id: P-5
+    name: Human Ceiling
+    description: |
+      Ambiguous provenance or a require_human mode MUST escalate to a human (DEC-076 ceiling)
+      rather than auto-resolve — a product-enforced governance gate.
+
+  - id: P-6
+    name: Content Sovereignty
+    description: |
+      The mechanism is domain-neutral (DEC-075); the acceptable-source policy is owned by the
+      customer (DEC-063). Policies SHOULD be customer-scoped and isolated (DEC-064).
+
+requirements:
+  - id: REQ-001
+    title: Model Origin Registry (first-class artifact)
+    description: |
+      Each known model MUST be classified by vendor -> jurisdiction -> origin class. The
+      registry is maintainable/extensible by the ecosystem/customer (like the license DB).
+      At minimum origin classes distinguish western-frontier, western-open, and other origins
+      so a policy can be expressed against them.
+    level: MUST
+    examples:
+      - "Claude -> Anthropic / US / western-frontier"
+      - "Qwen -> Alibaba / CN / china-origin"
+      - "Llama -> Meta / US / western-open"
+
+  - id: REQ-002
+    title: Provenance Policy Schema
+    description: |
+      A per-project/per-customer policy MUST express allow/deny origins (by class), a
+      local-weights distinction (foreign API/data-egress vs locally-run foreign open weights),
+      an enforcement mode (block/warn/require_human), and an unknown_origin behaviour that is
+      fail-closed by default.
+    level: MUST
+    schema_fields:
+      - "allow_origins / deny_origins (by origin class)"
+      - "china_local_weights: deny | allow_local_only (strict buyers deny even local foreign weights)"
+      - "mode: block | warn | require_human"
+      - "unknown_origin: block (fail-closed default)"
+
+  - id: REQ-003
+    title: Fail-Closed Enforcement (every model-selection point)
+    description: |
+      At every model-selection point (per-agent config / routing / BYOK) the chosen model's
+      origin MUST be checked against the active policy. A violation MUST block or escalate
+      fail-closed — never silently pass. Enforcement runs before routing so a disallowed
+      source is swapped for an acceptable model.
+    level: MUST
+    checks:
+      - "Apply TW-finance-gov-style policy -> selecting DeepSeek/MiniMax/Qwen (incl. local) is blocked"
+      - "Selecting Llama/Claude under the same policy is allowed"
+      - "Injection test: route a denied-origin model -> gate blocks (never silent pass)"
+
+  - id: REQ-004
+    title: Audit Evidence
+    description: |
+      Each generation MUST record the model used + its origin class, enabling a compliance
+      report ("this delivery used no models of the denied origin"). It plugs into the audit
+      logger / hash chain, verification-oracle correctness evidence (XSPEC-256), and the
+      telemetry allowlist (DEC-066). This report is the outward proof of the provenance moat.
+    level: MUST
+    evidence_required:
+      - "per generation: model id + vendor + origin class"
+      - "policy applied (allow/deny snapshot)"
+      - "denied-origin count (expected 0 under a deny policy)"
+      - "content hash (sha256) of the report"
+      - "audit-trail id (hash-chain entry)"
+
+  - id: REQ-005
+    title: Reference Policy Pack(s)
+    description: |
+      The mechanism SHOULD ship a small number of reference policy packs that demonstrate it
+      (e.g. a pack denying a given origin including local weights). The mechanism is neutral;
+      concrete policy content is owned by the customer or a compliance partner.
+    level: SHOULD
+    examples:
+      - "Reference pack 'TW-finance-gov': deny china-origin including local weights"
+
+  - id: REQ-006
+    title: Unknown-Origin Fail-Closed
+    description: |
+      A model whose origin is not in the registry (unknown provenance) MUST NOT be silently
+      allowed. The system MUST treat unknown origin as a fail-closed block or escalation
+      (REQ-002 unknown_origin), consistent with the governance-gate no-silent-pass rule.
+    level: MUST
+    checks:
+      - "Unknown-origin model under unknown_origin:block -> blocked/escalated, not auto-allowed"
+      - "Injection test: present an unregistered model -> gate does not silently pass"
+
+  - id: REQ-007
+    title: Provenance Policy Content Sovereignty
+    description: |
+      The mechanism is domain-neutral (DEC-075); the acceptable-source policy is
+      customer-owned (DEC-063). Policies SHOULD be customer-scoped and isolated (DEC-064).
+    level: SHOULD
+    checks:
+      - "Policy scoping is per-customer; provenance verdicts attribute policy to the customer"
+      - "Mechanism carries no jurisdiction-specific allow/deny baked in (only the registry classes)"
+
+# ─────────────────────────────────────────────────────────────
+# Governance-gate family (shared shape with XSPEC-193 / XSPEC-256)
+# ─────────────────────────────────────────────────────────────
+
+governance_gate_family:
+  description: |
+    model-provenance is the PROVENANCE (source) profile of one governance-gate mechanism.
+    All profiles share: registry/check -> fail-closed gate -> audit evidence ->
+    customer-override-telemetered -> human-escalation ceiling.
+  profiles:
+    - profile: licensing
+      standard: license-compliance
+      spec: XSPEC-193
+      registry: blocklist/allowlist/greylist
+      block_trigger: prohibited license
+    - profile: provenance
+      standard: model-provenance
+      spec: XSPEC-255
+      registry: model origin registry + policy
+      block_trigger: denied / unknown origin
+    - profile: correctness
+      standard: verification-oracle
+      spec: XSPEC-256
+      registry: ground-truth registry
+      block_trigger: non-reproduction / drift
+
+# ─────────────────────────────────────────────────────────────
+# Gate timing
+# ─────────────────────────────────────────────────────────────
+
+gate_timing:
+  pre_routing: REQ-003 fail-closed provenance check at every model-selection point
+  on_denied_or_unknown: REQ-006 / P-5 block or escalate to human (no silent pass)
+  on_allowed: REQ-004 record model + origin
+  after_run: REQ-004 emit provenance audit / compliance report (outward moat proof)
+
+# ─────────────────────────────────────────────────────────────
+# Telemetry (DEC-066 / XSPEC-189 v2 envelope)
+# ─────────────────────────────────────────────────────────────
+
+telemetry:
+  required_events:
+    - model_provenance_check      # origin check outcome per model-selection point
+    - model_provenance_block      # fail-closed block detail (REQ-003)
+    - model_provenance_escalation # require_human / ambiguous provenance escalated (P-5)
+    - unknown_origin_detected     # unregistered model encountered (REQ-006)
+    - human_override_block        # human override of a block (requires reason)
+  envelope_reference: XSPEC-189   # Telemetry Schema v2 envelope
+  event_type: quality
+  event_subtype_examples:
+    - model_provenance_check
+    - gate_pass / gate_fail       # when the provenance check acts as a selection gate
+
+# ─────────────────────────────────────────────────────────────
+# Integration with existing standards
+# ─────────────────────────────────────────────────────────────
+
+integration:
+  license-compliance: sibling profile; shared allowlist/denylist + fail-closed + audit + override skeleton
+  verification-oracle: sibling profile; provenance evidence sits alongside correctness evidence in one report
+  model-selection: provenance is a constraint applied before cost/quality right-sizing (unacceptable source removed from candidate set)
+  security-standards: foreign-API data egress + local-weights distinction are data-sovereignty controls
+  verification-evidence: the provenance report is a kind of verification evidence (model + origin + no-denied-origin trace)
+
+# ─────────────────────────────────────────────────────────────
+# Adoption guidance
+# ─────────────────────────────────────────────────────────────
+
+adoption_guidance:
+  uds_install_path: ai/standards/model-provenance.ai.yaml
+  notes:
+    - "v1.0.0 defines the provenance mechanism and audit evidence. Policy content (acceptable origins) is customer-owned (DEC-075/DEC-063)."
+    - "Adopters MAY enforce the check in CI / at the provider abstraction without a bespoke engine."
+    - "Distinguish foreign-API access (data egress) from locally-run foreign open weights — strict buyers deny even local weights."
+    - "Group with license-compliance and verification-oracle as one governance-gate subsystem (three profiles)."

package/bundled/ai/standards/model-selection.ai.yaml

@@ -8,8 +8,8 @@ standard:
   description: AI 模型分級選擇策略
 
   meta:
-    version: "2.0.0"
-    updated: "2026-04-13"
+    version: "2.0.1"
+    updated: "2026-06-10"
     source: core/model-selection.md
     description: "三層模型分級選擇策略 + 多模型池能力管理（XSPEC-027）"
     inspired_by: superpowers/subagent-driven-development
@@ -176,6 +176,15 @@ standard:
         action: "自動切換備用模型，上報 P1 Issue"
         priority: critical
 
+  navigation_integration:
+    description: >
+      The ai-response-navigation standard (Rule R6, optional) allows next-step suggestion
+      options to carry a Tier annotation (〔model: Fast〕 / 〔model: Standard〕 / 〔model: Capable〕).
+      The complexity signals in this standard serve as the judgment basis for those annotations.
+    vendor_neutral_principle: >
+      Tier names (Fast / Standard / Capable) are vendor-agnostic labels.
+      Each platform independently maps these tiers to its available models.
+
 physical_spec:
   type: checklist
   validator:

package/bundled/ai/standards/observability-standards.ai.yaml

@@ -8,6 +8,16 @@ meta:
   source: core/observability-standards.md
   description: Three pillars observability framework (Logs, Metrics, Traces) with Golden Signals and maturity model
 
+# XSPEC-070 Phase 2 — skill↔standard content-coverage audit.
+# Consumed by scripts/check-skill-content-coverage.ts. The anchor skill
+# (observability-assistant) is a thin launcher; require it to carry at least 30%
+# of the standard's depth so a stale skeleton is flagged (CONTENT-003).
+audit:
+  min_skill_ratio: 0.30
+  mandatory_keywords:
+    - golden signals
+    - maturity
+
 three_pillars:
   logs: "Discrete events — debugging, audit, error details (see logging.ai.yaml)"
   metrics: "Numerical measurements over time — dashboards, alerting, capacity"

package/bundled/ai/standards/performance-standards.ai.yaml

@@ -1,7 +1,7 @@
 # Performance Standards (AI-Optimized)
 # Source: core/performance-standards.md
 # Guide: core/guides/performance-guide.md
-# Version: 1.0.0
+# Version: 1.2.0
 
 iso25010_efficiency:
   time_behavior: "Response time, throughput"
@@ -127,3 +127,48 @@ capacity_planning:
     cpu: "Plan at 70%"
     memory: "Plan at 80%"
     storage: "Plan at 70%"
+
+# XSPEC-286: Migration Non-Functional Parity (axis 6 of XSPEC-284 completeness matrix)
+# Covers "function correct, data correct, but RUNS differently" omissions in a legacy rewrite.
+# Differential oracle: legacy is the baseline, "no regression beyond declared tolerance" is pass.
+migration_non_functional_parity:
+  axis: "6 (non-functional) of XSPEC-284 migration completeness matrix"
+  principle: "Legacy is the baseline; declared tolerance, not absolute target. Divergence self-reports."
+  why_gap: "General baseline/drift compares a version to its OWN prior baseline, not new-vs-legacy. Per-request and behavior-snapshot parity both pass while the system silently runs slower / loses a lock semantic."
+  derive: # R3 mechanized baseline capture
+    sources:
+      key_routes: "p50/p95/p99 latency per critical path"
+      hot_queries: "per-query latency, rows scanned (slow-query log)"
+      background_jobs: "job duration, throughput records/min"
+      connection_pool: "min/max connections, idle/connection timeout"
+      timeout_ratelimit: "per-endpoint timeout, requests/min, burst"
+      transaction_config: "isolation level"
+    note: "Captured set is both the regression baseline AND the resource/limit parity checklist."
+  regression_gate: # R1 performance regression differential
+    oracle: "measure new-system path, compare vs legacy baseline within declared tolerance"
+    default_tolerance:
+      p50_latency: "<= +10% vs legacy"
+      p95_latency: "<= +15% vs legacy"
+      p99_latency: "<= +20% vs legacy"
+      throughput: "<= -10% decrease vs legacy"
+    per_path_configurable: true
+    rollout: "shadow (measure-only) first to calibrate per-path tolerance, then promote to blocking"
+    gate_timing: [pre-UAT, post-cutover]
+  concurrency_isolation: # R2
+    method: "drive a race with concurrent stress, then assert DOMAIN INVARIANTS (not one-at-a-time output checks)"
+    invariants:
+      conservation: "SUM(balance) unchanged after N concurrent transfers"
+      uniqueness: "no duplicate sequence number under concurrent allocation"
+      no_lost_update: "committed write not silently replaced"
+      ordering: "events applied in guaranteed order under concurrency"
+    blocks_on: "isolation downgrade (e.g. SERIALIZABLE -> READ COMMITTED, removed row lock)"
+    boundary: "state-transition legality and temporal invariants owned by XSPEC-287 (axis 8); concurrency race owned here (axis 6)"
+    gate_timing: [pre-UAT]
+  resource_limit_parity: # R4
+    checklist: [request_timeout, rate_limit, connection_pool, batch_size, transaction_isolation]
+    rule: "each item matched OR has a declared justified delta; undeclared delta blocks cutover"
+  completeness_declaration:
+    derive: "Step1 mechanized baseline"
+    oracle: "Step2 regression differential + Step3 invariant assertions"
+    gate_timing: "pre-UAT + post-cutover"
+    reuse: "general baseline/drift + observability-assistant measurement; adds only migration before/after + concurrency-isolation semantics"

package/bundled/ai/standards/pii-classification.ai.yaml

@@ -3,7 +3,7 @@
 
 id: pii-classification
 title: PII Classification and Handling Standards
-version: "1.0.0"
+version: "1.1.0"
 status: Active
 tags: [compliance, privacy, pii, gdpr, data-protection, security]
 summary: |
@@ -107,3 +107,17 @@ requirements:
       - "New feature: 'Save payment method' → PIA required (TIER-1 card data)"
       - "PIA template: docs/templates/privacy-impact-assessment.md"
       - "PIA outcome: fingerprint auth approved with biometric data stored only on-device"
+
+# Hand-off contract (XSPEC-292 T16): how a classification reaches its consumers
+hand_off:
+  trigger:
+    - "New feature design: REQ-006 PIA classifies new PII fields before implementation"
+    - "PR/scan time: a PII field with no tier is a blocking finding"
+    - "Data-dictionary change: tier added/removed -> re-notify consumers"
+  interface: "PII classification registry entry per field {field, tier, purpose, legal_basis, masking_rule, retention}; consumers register per data-contract REQ-005"
+  consumers:
+    - "logging-standards: applies masking_rule to TIER-1/2 in non-prod"
+    - "audit-trail: records access/export/erasure events"
+    - "security-standards: enforces encryption for TIER-1"
+    - "privacy/compliance owner: notified on new TIER-1 field or classification change"
+  state: "Found -> Classified -> Handled (masked/encrypted/retention) -> Verified; a field stuck in Found blocks release"

package/bundled/ai/standards/pipeline-security-gates.ai.yaml

@@ -13,6 +13,7 @@ standard:
     - "Secrets scan at pre-commit — block ALL commits containing secrets"
     - "SAST after build — block Critical/High findings"
     - "SCA + SBOM generation at package stage"
+    - "Attestation verification before deploy — verify artifact checksum, cosign signature, SLSA provenance (consume supply-chain-attestation output; never generate SBOM/signature without verifying it)"
     - "DAST after staging deploy — warn on findings, require approval for Critical"
     - "Never skip security gates with --force or emergency bypass without audit trail"
     - "Treat security gate failures as pipeline failures, not warnings"
@@ -51,6 +52,14 @@ gate_positions:
     block_on: ["Critical CVE with fix available"]
     warn_on: ["High CVE", "outdated dependencies"]
     sbom_format: ["spdx", "cyclonedx"]
+  pre_deploy_attestation:
+    type: attestation_verification
+    tools: ["cosign", "sha256sum", "slsa-verifier"]
+    scope: "packaged release artifact (archive / container image)"
+    block_on: ["signature verification failure", "checksum mismatch", "missing SLSA provenance"]
+    warn_on: ["SLSA level below target (>= L2 recommended for public release)"]
+    consumes: supply-chain-attestation
+    note: "Runs after package_stage, before deploy — closes the generate-but-never-verify gap"
 
 bypass_policy:
   allowed: false
@@ -66,6 +75,7 @@ failure_behavior:
 integration_points:
   secrets_vault: "Integrate with HashiCorp Vault or AWS Secrets Manager for secrets injection"
   sbom_registry: "Upload SBOM to dependency-track or grype-db for continuous monitoring"
+  artifact_signing: "Verify SBOM/provenance signatures and checksum with cosign before release (see supply-chain-attestation)"
   incident_response: "Critical findings automatically create incident tickets"
 
 physical_spec:

package/bundled/ai/standards/privacy-standards.ai.yaml

@@ -1,5 +1,5 @@
 id: privacy-standards
 meta:
-  version: "1.0.0"
-  updated: "2026-04-01"
+  version: "1.1.0"
+  updated: "2026-06-18"
   source: core/privacy-standards.md

package/bundled/ai/standards/project-context-memory.ai.yaml

@@ -12,8 +12,8 @@ standard:
     - "Use identified paths and specific names (unlike Developer Memory)"
 
   meta:
-    version: "1.0.0"
-    updated: "2026-02-09"
+    version: "1.2.0"
+    updated: "2026-06-18"
     source: core/project-context-memory.md
     description: System for capturing, retrieving, and enforcing project-specific context, architectural decisions, and domain knowledge
 

package/bundled/ai/standards/push-standards.ai.yaml

@@ -1,7 +1,7 @@
 id: push-standards
 meta:
-  version: "1.0.0"
-  updated: "2026-04-23"
+  version: "1.1.0"
+  updated: "2026-07-09"
   source: skills/push/SKILL.md
   description: "Git push safety gates, protected branch detection, force-push guardrails, and push receipt audit trail"
 
@@ -14,6 +14,14 @@ config:
     default: ["lint", "test"]
     optional: ["ac-coverage", "type-check", "security-scan"]
     description: "Quality gates to run before pushing"
+    test_scope:
+      default: "full"
+      choices: ["full", "changed"]
+      description: >-
+        "changed" runs only the diff-affected test subset via the project's
+        test impact analysis tooling (e.g. `vitest --changed`), if configured.
+        Fast local pre-check only — never a substitute for the project's own
+        CI/full-suite run, which remains the authoritative gate.
 
   receipt:
     output:
@@ -58,7 +66,10 @@ rules:
     instruction: |
       Run all configured push_gates in sequence:
       - lint: run project lint command (from uds.project.yaml or detected)
-      - test: run project test command
+      - test: run project test command, respecting config.push_gates.test_scope —
+        "full" (default) runs the complete suite; "changed" runs only the
+        diff-affected subset via the project's test impact tooling if configured.
+        A "changed" pre-check does not replace the project's CI/full-suite run.
       If any gate fails: abort push, display which gate failed, suggest fix.
       If --skip-gates flag: run push without gates, mark gates_skipped: true in receipt.
 

package/bundled/ai/standards/refactoring-standards.ai.yaml

@@ -3,8 +3,8 @@
 
 id: refactoring-standards
 meta:
-  version: "2.0.0"
-  updated: "2026-01-21"
+  version: "2.2.0"
+  updated: "2026-07-01"
   source: core/refactoring-standards.md
   guide: core/guides/refactoring-guide.md
   description: Refactoring guidelines with tactical, strategic, and safety strategies including decision matrix
@@ -292,9 +292,14 @@ metrics:
     - metric: Coupling
       measurement: Module dependencies count
       target: Reduce dependencies
-    - metric: Code Duplication
+    - metric: Code Duplication (textual)
       measurement: Duplicate percentage
       target: "< 3%"
+      note: Detects textual duplication only; cannot detect semantic duplication (see semantic_duplication)
+    - metric: Semantic Duplication
+      measurement: Golden / architecture test
+      target: 0 unmanaged domain facts
+      note: Static analysis cannot detect the same domain fact re-implemented across call sites
 
   test_quality:
     - metric: Test Coverage
@@ -314,6 +319,40 @@ metrics:
       source: Rollback count
       expected: Should decrease
 
+# Semantic Duplication & Copy-Drift (issue #142)
+semantic_duplication:
+  anti_pattern:
+    id: copy-drift
+    name: Copy-Drift
+    definition: Same domain fact implemented in multiple places, OR a derived value persisted without an enforced binding to its source, so the copies drift apart independently
+    why_undetected: Copies are not byte-identical (paraphrased logic, formula right in one place and wrong in another), so textual-duplication metrics report nothing
+    symptoms:
+      - List view disagrees with detail view
+      - Stored counter double-counts because one update path's decrement/rollback branch was skipped or commented out
+      - Same formula gives different answers in different modules
+      - Scope over-reach where one path adjusts an aggregate others also touch
+
+  pattern:
+    id: single-source-of-truth
+    name: Single Source of Truth / Derive-Don't-Duplicate
+    rules:
+      - one_unit_per_fact: Extract each domain fact into a single classifier/calculator/resolver; every call site calls it, no re-implementations or paraphrases
+      - derive_dont_store: Prefer computing on read over persisting a second copy; a value always derived cannot drift
+      - aggregate_as_projection: If an aggregate must be stored, treat it as a projection of its detail rows and recompute it at a single choke point (e.g. SaveChanges interceptor or one shared recompute() function) so every write path (incl. async jobs and failure write-backs) stays consistent
+      - never_incremental_adjust: Never let an individual path apply incremental "col = col + delta" to a stored aggregate; that is how decrement/rollback branches get skipped and counters diverge
+      - lock_with_tests: Use a golden test to pin the single source of truth, and an architecture test that fails if the rule is redefined anywhere else
+
+  intentional_divergence_registry:
+    when: New system intentionally diverges from legacy behavior (e.g. legacy function returns X but legacy pipeline reconciles it to Y; the port folds that reconciliation into the single unit so it returns Y directly)
+    actions:
+      - Register each intentional divergence (legacy behavior, new behavior, rationale, owning fact)
+      - Pin the new behavior with a golden test so a future "correction" back to legacy fails the build
+      - Feed the registry into any differential test against legacy so the divergence is treated as expected, not a regression
+
+  adopter_notes:
+    - Per-function legacy-vs-new diffing produces candidates, not verdicts; adjudicate each against the ENTIRE legacy pipeline (which may already reconcile the difference downstream) before calling it a bug
+    - Extract legacy function bodies programmatically / token-based (not by hand) to avoid reintroducing transcription errors
+
 # Team Collaboration
 team_collaboration:
   pr_size_limits:
@@ -456,8 +495,34 @@ rules:
     instruction: Keep PRs under 200 lines; split larger refactorings
     priority: recommended
 
+  - id: single-source-of-truth
+    trigger: same domain fact (classification, rule, calculation) needed at multiple call sites
+    instruction: Extract one shared classifier/calculator/resolver and call it everywhere; never re-implement or paraphrase the fact
+    priority: required
+
+  - id: derive-dont-duplicate
+    trigger: a derived value (aggregate, counter, denormalized field) is being persisted
+    instruction: Prefer deriving on read; if it must be stored, recompute it as a projection at a single choke point and never apply incremental col = col + delta adjustments per write path
+    priority: required
+
+  - id: lock-single-source-with-tests
+    trigger: after establishing a single source of truth for a domain fact
+    instruction: Add a golden test to pin it and an architecture test that fails if the rule is redefined elsewhere
+    priority: recommended
+
+  - id: register-intentional-divergence
+    trigger: new system intentionally diverges from legacy behavior during a port/migration
+    instruction: Record it in an intentional-divergence registry, pin with a golden test, and feed differential tests so it is treated as expected not a regression
+    priority: required
+
+  - id: adjudicate-diff-against-pipeline
+    trigger: reviewing per-function legacy-vs-new diffs
+    instruction: Treat diffs as candidates only; adjudicate each against the entire legacy pipeline before calling it a bug, and extract legacy bodies programmatically to avoid transcription errors
+    priority: recommended
+
 related_standards:
   - test-driven-development.md
   - testing-standards.md
+  - test-governance.md
   - code-review-checklist.md
   - checkin-standards.md