G.9 — Parity and Benchmark Harness
Status: Stable
G.9:0 — Use this when
- rival method families, method sets, or adaptation paths must be compared under one declared baseline set and freshness window
- you need parity to publish one reproducible report rather than one opaque benchmark score
- downstream selection must recover comparator, normalization, bridge, and evidence pins without relying on one hidden scoring sheet
First move. State whether the comparison concerns exact targets or results about one subject, then fix the baseline, comparison rule and windows in one plan edition. Use A.19.CPM directly for a bounded comparison that needs no reproducible parity report; use this harness when the report and its replay basis are required.
G.9:0.1 — What goes wrong if missed
- benchmark reports present numbers from different windows, baselines, or comparator editions as comparable
- reuse across distinct source-local meanings, a reference-plane crossing, or a normalization mapping stays hidden until a disagreement appears downstream
- parity flattens a partial order into one scalar winner and silently changes what the comparison means
G.9:0.2 — What this buys
- one
ParityPlanRefthat fixes the plan edition, baseline, freshness, comparator, and bridge discipline up front - one
ParityReportthat cites that plan edition and echoes its active baseline binding, pins, outcomes, and evidence trace by value - one harness that downstream selection can consume without inventing a
G.9-local CSLC gate or a shadow governance card
Illumination, coverage and regret are telemetry by default. To use them in dominance, declare the policy that authorizes the promotion and record its ID.
G.9:1 — Intent
Provide a notation‑independent harness that:
- plans parity runs for one explicit subject—either one
EntityOfConcernRefor target refs under their existing subject patterns—with aReferencePlane, scope, window, applicable rules, CSLC comparability and admissibility references, comparator references (CNSpecRef,CGSpecRef,ComparatorSpecRef), and reproducibility pins for editions and policy ids; - executes the comparison with a DRR and SCR evidence trace; when selection is current, G.5 consumes its result and declares the selector outcome;
- publishes an edition-pinned ParityReport suitable for downstream consumption, shipping, refresh wiring, and RSCR.
G.9:2 — Problem frame
Parity claims become non‑reproducible or non‑comparable when any of the following are implicit:
- evidence window and freshness regime,
- comparator semantics, including any normalization or comparability mapping,
- the active C.21 replay basis when DHC coordinates are compared, including the exact Characteristic, Scale, measurement definition, Method, MethodDescription or model, and time or population basis,
- reuse across distinct F.17 source-local meanings or ReferencePlanes (the obtaining relation, crossing pins, and CL penalty placement),
- dominance and
PortfolioModeinterpretation rules, - gate outcomes (why a run abstained or degraded).
G.9 publishes the references needed to recover these inputs and outcomes when disclosing how the comparison was obtained (MOO). This disclosure cites the governing specifications without introducing a new governing specification reference.
G.9:3 — Forces
- Pluralism vs comparability. Support justified comparisons across Traditions while preserving their differences in meaning and admissibility.
- Partial orders. Many targets are only partially ordered; parity reporting must preserve CSLC-admissible outcome shape (often declared set results rather than a single scalar; an archive may supply the compared candidates).
- Edition sensitivity. A reader must be able to tell which measurement and comparator definitions a parity run used, even after those definitions change. When DHC, QD, or OEE modes are used, the required definition pins are introduced only through the corresponding
Extensionsblocks; omit them when unused. - Telemetry versus objectives.
IlluminationSummary, coverage, and regret are report-only telemetry by default. A dominance change needs an explicit CAL policy id recorded in the audit pins. - Crossing visibility. An actual E.18 flow crossing or A.21 gate used by parity retains its required
CrossingBundle, pins and checks. Failure blocks the dependent publication or use. - Cross-sense and reference-plane reuse. When expressions have distinct F.17 source-local meanings, recover both cells and establish the required F.9 relation; a ReferencePlane crossing follows its own declared crossing basis. Each actual correspondence retains its required pins and evidence; any justified numerical loss follows its receiving model and R-channel placement.
- Refreshability. Parity must emit RSCR-relevant causes as canonical ids, with enough pins to identify the affected use and plan any needed rerun.
G.9:4 — Solution
G.9:4.0 — G.Core linkage (normative)
This pattern binds to G.Core through the following manifest.
GCoreLinkageManifest (G.9) (normative; expands per G.Core:4.2)
Effective obligations/pins/triggers are computed as union(expand(sets), explicit deltas) under Nil‑elision.
-
CoreConformanceProfileIds:= {GCoreConformanceProfileId.PartG.AuthoringBase,GCoreConformanceProfileId.PartG.TriStateGuard,GCoreConformanceProfileId.PartG.ShippingBoundary,GCoreConformanceProfileId.PartG.UTSWhenPublicIdsMinted} -
RSCRTriggerSetIds:= {GCoreTriggerSetId.CGSpecGate} -
RSCRTriggerKindIds:= {RSCRTriggerKindId.EvidenceSurfaceEdit,RSCRTriggerKindId.PenaltyPolicyEdit,RSCRTriggerKindId.BaselineBindingEdit,RSCRTriggerKindId.TelemetryDelta} (Pattern-local deltas; cross-tradition or Bridge-calibration causes are wired viaG.9:Ext.CrossTraditionParityand MUST NOT over-trigger parity runs that use one already recovered meaning and ReferencePlane.) -
DefaultsConsumed:= {DefaultId.DominanceRegime,DefaultId.PortfolioMode,DefaultId.GammaFoldForR_eff} (Defaults are cited throughG.Core.DefaultGoverningDefinitionIndex(not restated here); the expected default governing definitions areCC‑G5.28,CC‑G5.23, andCC‑G5.4respectively.) -
CorePinSetIds:= {GCorePinSetId.PartG.AuthoringMinimal,GCorePinSetId.PartG.CrossingVisibilityPins} -
CorePinsRequired(pattern delta; pin names only; all are id‑valued unless noted) := {ComparatorSpecRef.edition,TaskSignatureRef?,TaskMapRef?, (when a selector consumes the parity result; TaskMapRef only when its G.4 CAL gates are current)entityOfConcernRef?,targetRefs[]?, (exactly one subject branch)ClaimScope,EvaluationWindow,FreshnessWindows,BaselineSet,BaselineBindingRef,ParityPinSet,PlannedFillingRows[]?,EvidenceGraphId,Budgeting?,EpsilonDominance?,UNM_id?,NormalizationMethodId[]?,NormalizationMethodInstanceId[]?,SCPRef.edition?,MinimalEvidenceRef.edition?} (Nil‑elision applies; mode‑specific definition pins are introduced only by the correspondingGPatternExtensionblocks.) -
TriggerAliasMapRef:=∅
G.9:4.1 — Objects and publication records
All objects below are notation‑independent; serialisations (if any) are handled in shipping and interop publication forms, not here.
(1) ParityPlan (one U.WorkPlan episteme; ParityPlan is the local application name)
A plan that fixes what is being compared and under what pinned conditions.
Minimal fields (conceptual; ids/pins only):
ParityPlan := ⟨
ParityPlanId(UTS), // continuing plan lineage
planEdition, // one immutable edition
CGFrameId?, // exact cited CG frame when the plan depends on one
TaskSignatureRef?, TaskMapRef?, // conditional G.5 input; TaskMap only for current G.4 CAL gates
entityOfConcernRef? := EntityOfConcernRef, // one-EntityOfConcern branch only
targetRefs[]?, // exact-target branch only; existing kinds and editions
groundingHolonRef := GroundingHolonRef,
referencePlaneRef := ReferencePlane,
claimScopeRef := ClaimScope,
EvaluationWindow,
UNM_id?, NormalizationMethodId[]?, NormalizationMethodInstanceId[]?, // when “normalize, then compare” is required (ids only; semantics come from CN‑Spec / UNM)
EpsilonDominance?, // optional ε-front thinning (ε≥0; id/param; pinned when used)
PortfolioMode?, DominanceRegime?, // may be explicit or inherited via DefaultGoverningDefinition (semantics follow G.5)
BaselineSet, // exact method-family or generator-family targets (ids; notation-independent)
BaselineBindingRef, // evidence-backed baseline-set reference that says what counts as baseline
FreshnessWindows,
CNSpecRef.edition, CGSpecRef.edition, ComparatorSpecRef.edition, // edition-pinned refs
SCPRef.edition?, // optional (when a specific SCP profile must be pinned/cited)
MinimalEvidenceRef.edition?, // optional (when CG-Spec exposes minima profiles by ref)
Budgeting?,
ParityPinSet,
EvidenceGraphId, PathId[], PathSliceId?,
PlannedFillingRows[]? // declaration-local A.15.3 content inside this WorkPlan; no independent row refs
⟩
ParityPlanRef := <ParityPlanId, planEdition> designates one immutable plan edition. Changing its subject, baseline binding, comparator edition, or another active value that can change the run or its interpretation creates a new planEdition.
When a G.5 selector consumes this comparison, pin its TaskSignatureRef in the plan. If that selector uses G.4 CAL gates, also pin its TaskMapRef and require the map’s task-signature reference to match. Omit these selector inputs from a parity run that needs no selector.
A later edition may retain the lineage ID only if it continues the same plan. Earlier ParityPlanRef values still resolve their original editions.
Exactly one subject branch is present. Use entityOfConcernRef when the report compares results about one EntityOfConcern. Use targetRefs[] when the targets themselves are compared; each ref keeps the kind and edition defined by its existing subject pattern. In particular, a G.5 method-family target is an exact MethodFamilyRowRef, and a generator-family target is an exact GeneratorFamilyRowRef.
For example, a comparison of <ThresholdTrendReview-local, R3> and <SpectralResidualReview-local, R2> puts those two row refs in targetRefs[]; the plan may explicitly use the same two refs as its BaselineSet. A comparison of their results for Pump-P17 instead puts Pump-P17 in entityOfConcernRef, leaves targetRefs[] absent, and uses BaselineBindingRef to say how the two method rows supply results about that pump.
BaselineSet names the alternatives treated as the comparison baseline; it supplies targetRefs[] only when the plan explicitly says that the same exact refs serve both purposes. Otherwise the subject and baseline remain separate, and BaselineBindingRef records how that baseline applies to the named subject.
These values determine what is compared and when the parity claim is usable; do not add ParityContextId. If the plan relates expressions with distinct source-local meanings, first recover the exact F.17 cells and establish the required F.9 relation. A shared label, source note, or generic context identifier does not establish comparability.
(2) ParityPinSet (pin set)
A declared set of pins required for reproducibility and audit (editions + policy‑ids + UTS/Path pins).
G.9 specifies the contents of ParityPinSet under the common pin rules in G.Core.
(3) ParityReport (UTS publication record; work-result or audit-facing publication record only when the neighboring source exists)
A UTS-publishable parity publication record produced by a parity run under one ParityPlanRef. Work or audit occurrence claims use A.15/A.15.1; evidence-path claims use A.10/G.6; a separate named assurance claim uses B.3; and a gate decision uses A.21 under its applicable profile. Keep each such occurrence, path, or result distinct from the report.
ParityReport := ⟨
ParityReportId(UTS),
parityPlanRef := ParityPlanRef,
entityOfConcernRef?, targetRefs[]?, // exactly one subject branch is present
TaskSignatureRef?, TaskMapRef?, // echoed when the selector branch is used
groundingHolonRef, referencePlaneRef,
claimScopeRef, EvaluationWindow,
BaselineSet, BaselineBindingRef, FreshnessWindows,
CNSpecRef.edition, CGSpecRef.edition, ComparatorSpecRef.edition,
SCPRef.edition?, MinimalEvidenceRef.edition?, // echoed iff used/pinned in the plan
UNM_id?, NormalizationMethodId[]?, NormalizationMethodInstanceId[]?, // echoed iff used in the plan
OutcomeRefs, // parity/comparison results; G.5 output refs only when selection is current
EpsilonDominance?, // echoed when used
AbstainReasons[]?, // ids/labels (policy-bound) for abstain/degrade; refusal paths included
TelemetrySummary? := ⟨IlluminationSummary?, coverage?, regret?⟩, // report-only by default; promotion requires CAL policy-id pins
GuardOutcomeTraceRef?, // pass/degrade/abstain trace + cited reasons (policy-bound)
EvidenceTrace := ⟨EvidenceGraphId, PathId[], PathSliceId?⟩,
CrossingPins?, // Bridge/CL/Φ/Ψ/Φ_plane pins, when crossings are invoked
EditionPinsDelta?, // explicit list of edition pins actually active during the run
PolicyPinsDelta?, // explicit list of policy-ids actually active during the run
RSCRRefs[] // parity RSCR test ids / trigger emissions
⟩
The report carries the ParityPlanRef and echoes the BaselineBindingRef used in that edition. For example, if <PumpParityPlan, E4> used PumpBaselineBinding-E7 and a later E5 changes the binding or comparator, an old report still resolves E4 and PumpBaselineBinding-E7. A missing historical plan edition or binding is an unresolved required input; it is never replaced with the current value.
Naming discipline.
- Head names follow the existing kind definitions and LEX discipline.
- The older labels
ParityPlan@ContextandParityReport@Contextare retired. The suffix did not identify the object or the comparison basis. UseParityPlanandParityReport, with all operative conditions carried in explicit fields and exact refs. - Tech/Plain twins follow E.10 rules (no drift‑inducing synonyms in Tech).
G.9:4.2 — Parity planning (one U.WorkPlan)
Planning records the conditions needed to repeat and interpret the parity run:
-
Fix the baseline set. Choose the exact
BaselineSet(MethodFamilies, and optionally GeneratorFamilies) used as the comparison baseline. When SoS-log or source-maturity values change baseline eligibility or interpretation, citeSoS‑LOGBundleId?and the source-maturity ids by reference; acceptance-gate thresholds remain inG.4Acceptance. -
Bind subject, scope, and evaluation window. Choose exactly one subject branch: one
entityOfConcernRef, or exacttargetRefs[]under their existing kinds and editions. For G.5 families, useMethodFamilyRowReforGeneratorFamilyRowRef, not a bare lineage id. Then recordgroundingHolonRef,referencePlaneRef = ReferencePlane, the chosenClaimScopeandEvaluationWindow. Keep the EntityOfConcern distinct from its grounding holon, and preserve the stated scope and window. -
Define baseline-set reference. Declare what counts as the baseline and how it applies to the selected subject in
BaselineBindingRef(for example, through an EvidenceGraph path slice or an upstream shipped package or publication-record id). IfBaselineSetalso supplies the exact compared targets, say so and use the same refs by value; otherwise keep baseline and subject refs distinct. -
Equalise window (and budget, if pinned). Declare a single
FreshnessWindowsand apply it across all baselines; ifBudgetingis used/pinned, it MUST be shared/pinned across baselines as well.When specialization is part of the parity claim, the same plan should also hold constant the declared task family or target scope cut, the work-measure threshold target, adaptation budget, prior exposure declaration, and freshness window. Identify what each candidate performed and what assistance it had. To compare the candidates’ own acquisition, account for differences in support that could explain the result. If the arrangements themselves are being compared, retain each arrangement as the bearer of its performance. When transfer, retention, later efficiency, downside or corridor entry matters, keep the observations and the baseline supporting that claim under C.22.1:8.
-
Pin governance, CSLC comparability and admissibility references, and comparator references.
CNSpecRef,CGSpecRef, andComparatorSpecRefare referenced with explicit edition pins. -
Pin measurement/comparator definitions (conditional). Where parity depends on mode‑specific definition records (e.g., DHC/QD/OEE), pin the relevant definition ids/editions/policies. The minimum required pins are declared by the applicable
Extensionsblocks (e.g.,G.9:Ext.DHCParityPins,G.9:Ext.QDArchiveParity,G.9:Ext.OEEParity) and the referenced records they cite. -
Bind comparator choice to CG-Spec (CSLC comparability and admissibility). Any numeric comparison or aggregation MUST be CSLC‑admissible and cite the corresponding CG‑Spec entry (via
ComparatorSpecRef). If Characteristics differ by unit, scale, or space, the plan MUST declare the ids used for “normalize, then compare” (UNM_id?,NormalizationMethodId[]?,NormalizationMethodInstanceId[]?) — ids only; semantics are defined elsewhere. -
Declare order & PortfolioMode semantics. Parity MUST preserve set‑return semantics;
PortfolioModeandDominanceRegimeare either explicitly pinned or cited throughG.Core.DefaultGoverningDefinitionIndex. IlluminationSummary/coverage/regret remain telemetry unless a CAL policy explicitly promotes them (policy‑id pinned & recorded). -
Attach planned fillings when applicable. If parity depends on planned slot fillings, this WorkPlan contains the relevant A.15.3 rows in
PlannedFillingRows[]; each row points to a declaration member defined by its own pattern and has no independent reference or identity. Omit the field when no such row is needed. -
Publish the relation and its receiving use (when invoked). For distinct recovered F.17 meanings, establish the F.9 relation and the separate bounded-use claim and matching reliance. Cite G.7 calibration, CL and policy pins when that account uses them; its summary does not decide parity admissibility. Kind correspondence also requires receiving admissibility and fresh target classification under C.3.3. A plane relation keeps its own basis. Any numerical assurance penalty follows the applicable B.3 model for the named use.
G.9:4.3 — Execution protocol (run‑time / selector‑adjacent)
For one run under the pinned plan, follow these steps:
-
Validate CSLC references and pins. Validate the cited CSLC comparability and admissibility references, active pins, and witnesses. Apply the eligibility or acceptance checks required by the pinned comparison and refuse or abstain on non-admissible operations. When using G.5, record its branch trace in its own result vocabulary and use the planned
TaskSignatureRef. Add the matchingTaskMapRefwhen that selector uses G.4 CAL gates. If a liveA.21gate consumes this check, cite itsGateDecisionResultand add aDecisionLogonly when needed for audit or reuse. The profile application and its action consequence remain governed by A.21; do not create aG.9-local CSLC gate. -
Run the comparison and any required selection. Apply the pinned comparator to the admissible comparison inputs and retain its lawful result shape in
ParityReport. If a selector-facing result is required, apply G.5 under the plan’s exact task and conditional map refs and publish its declared outcome. An ordinary parity or benchmark report ends without G.5 when no selector result is needed.For a bounded-specialization comparison, the report should include the active specialization profiles or equivalent pins so readers can recover the work-measure threshold target, prior exposure, budget-to-threshold, post-threshold efficiency when relevant, transfer, retention, downside field, and any corridor-entry baseline or evidence note from the report.
-
Record the comparability mapping when used. If
UNM_id?,NormalizationMethodId[]?, orNormalizationMethodInstanceId[]?was declared, echo it inParityReportor its explicit pins delta. Record the ids and any scoped notes required by the cited specification in the audit pins and SCR; cite the applicablePathIdvalues. -
Publish trace. Emit
ParityReportwith theParityPlanRef, itsBaselineBindingRef, conditional selector input refs, EvidenceGraph citations, and all active edition and policy-id pins, so the run can be checked and run again. -
Emit telemetry for refresh when produced. Publish the telemetry pins or events used for refresh. Treat the values as report-only unless a CAL policy promotes them and its ID is recorded.
G.9:4.3a — Worked parity slice
Ordinary case: compare two pump-triage method rows. The team from the G.5 example wants a reproducible comparison of the two exact selector-row editions, not a claim that one Method is universally better. Both rows use the same scheme, units, and ReferencePlane, so no normalization or crossing branch is needed.
ParityPlanRef = <PumpTriageParity, E1>
targetRefs = [<ThresholdTrendReview-local, R3>,
<SpectralResidualReview-local, R2>]
entityOfConcernRef = absent
groundingHolonRef = PumpMaintenanceProgram-H1
referencePlaneRef = PumpVibrationTriage-RP1
claimScopeRef = PumpFleet-F7-VibrationTriageClaims-E1
EvaluationWindow = 2026-08-01T00:00Z .. 2026-08-07T23:59Z
BaselineSet = [<ThresholdTrendReview-local, R3>,
<SpectralResidualReview-local, R2>]
BaselineBindingRef = PumpTriageBaselineBinding-E1
FreshnessWindows = { sensorSeries: at-most-24h-old-at-run,
evidencePath: at-most-72h-old-at-run }
CNSpecRef.edition = PumpCN-E2
CGSpecRef.edition = PumpCG-E4
ComparatorSpecRef.edition = PumpTriageComparator-E3
TaskSignatureRef = PumpVibrationTriage-T1
TaskMapRef = absent // this G.5 case uses no G.4 CAL gate
ParityPinSet = [PumpCN-E2, PumpCG-E4, PumpTriageComparator-E3, PumpVibrationMeasureSpec-E2]
EvidenceGraphId = PumpTriageEvidence-E5
PathId[] = [PumpReadings-P7, ComparatorRun-P3]
PathSliceId = PumpParitySlice-S2
expected selector result = unordered Shortlist
The plan explicitly says that BaselineSet and targetRefs[] contain the same two row refs. EvaluationWindow bounds the observations and results included in the comparison. FreshnessWindows asks a different question at run or reuse time: whether each required input and evidence path is still recent enough to rely on.
A report from this run carries parityPlanRef=<PumpTriageParity, E1>, BaselineBindingRef=PumpTriageBaselineBinding-E1, the same evidence path, and the unordered Shortlist; it does not invent a scalar winner.
Conditional specialization case. Loop-engineering parity may add further pins after the ordinary comparison boundary above is complete. An evaluation program, benchmark script, or dashboard is part of the evaluation or comparison procedure; it is not the Characteristic being improved.
- Two agentic search setups both claim bounded specialization on the same declared task family.
- Their
ParityPlanalso pins the same threshold target, adaptation budget, prior-exposure declaration, and corridor-entry baseline. One setup reaches threshold sooner but shows low retention and no transfer. The other reaches threshold later, but carries reusable transfer and lower downside field. - Their CSLC-admissible
ParityReportstates what was held constant, which signals remained telemetry, and why the outcome stays a selected set or partial order rather than collapsing into a scalar winner. These specialization values extend the complete comparison boundary; they do not replace its subject, scope, windows, baseline binding, comparator editions, or evidence.
G.9:4.3b — Conditional causal method rung parity
Use this extension only when a parity report compares causal methods or causal-use claims. Start with a cheap screen and stop at degraded parity or abstention when the methods answer different questions.
CausalRungParityScreen:
comparedMethodsRef
targetCausalityLadderRungSet
causalSupportComponentTypeSet
sameEstimand: yes | no | unclear
sameOutcomeWindow: yes | no | unclear
sameTransportEndpoints: yes | no | unclear
cheapParityStop:
comparableEnoughForFullRecord |
crossRungDegrade |
crossSupportComponentsDegrade |
differentEstimandAbstain |
differentOutcomeWindowAbstain |
differentEndpointsAbstain |
returnToC28
causalSupportComponentTypeSet records which methods rely on evidence paths/data regimes, identification, estimates, direct counterfactual sampling, simulation, or transport. Expose the difference and justify the proposed comparison under its causal-support conditions; a shared label does not make unlike components equivalent.
Open the full record only when comparison remains meaningful:
CausalMethodRungParityRecord:
comparedMethodsRef
causalUseQuestionRef?: CausalUseQuestionRef
targetCausalUseClaimKind: CausalUseClaimKind
targetCausalityLadderRung: CausalityLadderRung
causalEstimandRef: CausalEstimandRef
declaredCausalityLadderBridgeOrLossRef?
interventionBudgetOrActionSetRef?
causalSupportComponentRefs: CausalSupportComponentRefs
declaredCausalSupportLossRef?
causalUseSupportResultRef?: CausalUseSupportResultRef
causalFollowUpWindowRef
outcomeMeasureRef
sourcePopulationRef?
targetPopulationRef?
sourceDomainRef?
targetDomainRef?
sourceEnvironmentRef?
targetEnvironmentRef?
sourceDataGeneratingRegimeRef?
targetDataGeneratingRegimeRef?
transportabilityResultRef?
estimateResultRef?
parityVerdict: parityEstablished | degraded | abstain
supportedParityUse
unsupportedParityUse
The record names every changed transport endpoint that matters; population and semantic scheme do not substitute for domain, environment, or data-generating regime. Different rungs, estimands, windows, endpoints, or support components require a bridge/loss, degraded parity, or abstention. G.9 makes the parity conclusion; C.28 supplies the cited causal-support result and does not authorize the benchmark conclusion.
G.9:4.9 — Extensions (pattern‑scoped; non‑core)
For an ordinary parity comparison that uses none of these modes, you can stop after the worked case in §4.3a. Use §4.3b when comparing causal methods or causal-use claims, and the extension blocks below when their mode is active.
The extensions specify only the pins, references, policy IDs, triggers and links needed for each mode. The referenced patterns define what these mean.
GPatternExtension block: G.9:Ext.CrossTraditionParity
GPatternExtension: CrossTraditionParity
- PatternScopeId:
G.9:Ext.CrossTraditionParity - GPatternExtensionId:
CrossTraditionParity - GPatternExtensionKind:
DisciplineSpecific - GoverningPatternId:
G.7 - Uses:
{G.7, F.9, E.18, A.21} - ⊑/⊑⁺:
∅ - RequiredPins/EditionPins/PolicyPins (minimum; conditional on use):
BridgeId/BridgeCardId[](for an actual F.9 sense correspondence; cite a card when used)BridgeMatrixId?CalibrationLedgerId?/BCT.id?RegressionSetId?/SentinelId[]?(when sentinel wiring is used)CL/CL^k/CL^plane(only the channel values consumed or required by the calibration or named assurance account)Φ(CL) policy-id,Φ_plane policy-id,Ψ(CL^k) policy-id?(only when the receiving model or assurance account requires that policy)CrossingBundleId?(required for an actual flow/gate use whose governing rule requires the bundle)
- RSCRTriggerSetIds:
{GCoreTriggerSetId.BridgeCalibrationKit}(preferred; expands inG.Core) - RSCRTriggerKindIds (delta, if any):
∅ - Notes: Apply the channel and receiving-use conditions of G.Core:4.2.3. A kind or plane correspondence keeps its own relation and receiving grounds; it creates no F.9 Bridge. The pins cite the calibration, loss model or crossing bundle consumed.
GPatternExtension block: G.9:Ext.SoSLogGuardNarration
GPatternExtension: SoSLogGuardNarration
- PatternScopeId:
G.9:Ext.SoSLogGuardNarration - GPatternExtensionId:
SoSLogGuardNarration - GPatternExtensionKind:
MethodSpecific - GoverningPatternId:
C.23 - Uses:
{C.23, G.6, G.4} - ⊑/⊑⁺:
∅ - RequiredPins/EditionPins/PolicyPins (minimum; conditional on use):
SoSLogRuleId[]/BranchId[](ids as cited labels; semantics come fromC.23)FailureBehaviorPolicyId/SoSLogBranchIdEvidenceTrace.PathId[]/PathSliceId?AcceptanceClauseId[](when referenced)
- RSCRTriggerKindIds:
{RSCRTriggerKindId.PolicyPinChange, RSCRTriggerKindId.EvidenceSurfaceEdit, RSCRTriggerKindId.MaturityRungChange, RSCRTriggerKindId.TelemetryDelta} - Notes (wiring-only): Explains why a parity run degraded/abstained by citing SoS‑LOG ids and evidence paths; does not redefine guard semantics.
GPatternExtension block: G.9:Ext.DHCParityPins
GPatternExtension: DHCParityPins
- PatternScopeId:
G.9:Ext.DHCParityPins - GPatternExtensionId:
DHCParityPins - GPatternExtensionKind:
MethodSpecific - GoverningPatternId:
C.21 - Uses:
{C.21} - ⊑/⊑⁺:
∅ - Required replay values and pins (minimum; conditional on DHC parity):
DisciplineRefIntendedUseClaimScopeRefComparisonBasisCharacteristicRef.editionScaleRef.editionUnitRef.edition?DHCMethodRef.editionMethodRefMethodDescriptionRef.edition?MeasurementModelRef.edition?CalibrationBasisRef?TimeOrPopulationBasisDHCDefinitionSetRef.edition?TargetSliceRef?DistanceDefRef.edition?
- RSCRTriggerKindIds:
{RSCRTriggerKindId.EditionPinChange, RSCRTriggerKindId.PolicyPinChange, RSCRTriggerKindId.EvidenceSurfaceEdit} - Notes (wiring-only): Carry exactly the active fields of the C.21 replay basis.
TargetSliceRefappears only when the parity computation consumes that A.2.6 selection and states its relation toClaimScopeRef. Compatible same-semantics readings use the admitted C.16 comparison basis directly; actual distinct-local-sense use also cites the obtaining F.9 relation, direction, admitted use, and loss. C.21 defines the DHC semantics.
GPatternExtension block: G.9:Ext.QDArchiveParity
GPatternExtension: QDArchiveParity
- PatternScopeId:
G.9:Ext.QDArchiveParity - GPatternExtensionId:
QDArchiveParity - GPatternExtensionKind:
MethodSpecific - GoverningPatternId:
C.18 - Uses:
{C.18, C.19, G.5} - ⊑/⊑⁺:
∅ - RequiredPins/EditionPins/PolicyPins (minimum; conditional on use):
DescriptorMapRef.editionDistanceDefRef.editionCharacteristicSpaceRefwith separatecharacteristicSpaceDescriptionRef: U.EpistemeRef(when discretisation/topology is referenced; resolve the same-space pair under A.19:5.1.1.1)EmitterPolicyRefInsertionPolicyRef
- RSCRTriggerKindIds:
{RSCRTriggerKindId.EditionPinChange, RSCRTriggerKindId.PolicyPinChange, RSCRTriggerKindId.TelemetryDelta} - Notes: Pin the QD definitions used by the comparison; their semantics remain in
C.18,C.19andG.5.
GPatternExtension block: G.9:Ext.OEEParity
GPatternExtension: OEEParity
- PatternScopeId:
G.9:Ext.OEEParity - GPatternExtensionId:
OEEParity - GPatternExtensionKind:
MethodSpecific - GoverningPatternId:
C.19 - Uses:
{C.19, G.5} - ⊑/⊑⁺:
∅ - RequiredPins/EditionPins/PolicyPins (minimum; conditional on use):
TransferRulesRef.editionEnvironmentValidityRegionIdExplorationBudgetPolicyId?EvidenceTrace.PathSliceId?(for transfer‑keyed events)
- RSCRTriggerKindIds:
{RSCRTriggerKindId.EditionPinChange, RSCRTriggerKindId.PolicyPinChange, RSCRTriggerKindId.TelemetryDelta} - Notes (wiring-only): For open-ended parity, cite the policies and editions above and connect their telemetry to refresh. The extension adds no core rule.
G.9:5 — Interfaces (minimal I/O; conceptual)
| Interface | Consumes | Produces |
|---|---|---|
G.9‑1 Plan_Parity | exactly one subject branch—one EntityOfConcernRef or exact targetRefs[] under their existing kinds and editions—plus GroundingHolonRef, ReferencePlane, ClaimScope, EvaluationWindow, BaselineSet, BaselineBindingRef, FreshnessWindows, Budgeting?, EpsilonDominance?, CNSpecRef.edition, CGSpecRef.edition, ComparatorSpecRef.edition, mode-specific measurement or normalization editions when used, SCPRef.edition?, MinimalEvidenceRef.edition?, UNM_id?, NormalizationMethodId[]?, NormalizationMethodInstanceId[]?, ParityPinSet, EvidenceGraphId, PathId[], PathSliceId?, PlannedFillingRows[]? | one immutable ParityPlan WorkPlan edition and its ParityPlanRef |
G.9‑2 Run_Parity | ParityPlanRef and comparison inputs; G.5‑3 Select only for a selector-facing result, using the plan’s TaskSignatureRef and conditional matching TaskMapRef | parity/comparison result refs and, when selected, G.5 outcome refs; DRR and SCR pins with PathId[] and, when needed, PathSliceId |
G.9‑3 Publish_ParityReport | ParityPlanRef, parity-run trace refs, and active pins | ParityReport carrying the same plan ref and baseline binding (UTS publication record; emits canonical RSCR ids) |
G.9‑4 Expose_ParityTelemetry | Telemetry deltas (archive changes, coverage/regret signals, etc.) | Telemetry events carrying PathSliceId?, policy‑ids, and edition pins for refresh wiring |
Publication records are conceptual here; serialisations belong in shipping and interop publication forms (see G.10 and interop annexes), not in G.9.
G.9:6 — Conformance Checklist (CC‑G9)
CC‑G9‑CoreRef (normative).
G.9 conforms only if it satisfies the effective set of CC‑GCORE‑* declared in G.9:4.0 GCoreLinkageManifest (including trigger typing, Default Governing Definition Index links, and P2W split).
-
CC‑G9.1 — Exact comparison boundary, equal windows (and budgets), and pinned spec editions (local). A
ParityPlanRef = <ParityPlanId, planEdition>SHALL resolve one immutable plan edition. That ParityPlan SHALL choose exactly one subject branch: oneEntityOfConcernRef, ortargetRefs[]under their existing kinds and editions. It SHALL also nameGroundingHolonRef,ReferencePlane,ClaimScope,EvaluationWindow, baseline set and binding, and evidence refs, and SHALL declare a singleFreshnessWindowsshared across baselines.BaselineSetsupplies the target refs only when the plan explicitly identifies the same refs in both places; otherwiseBaselineBindingRefrelates the separate baseline to the named subject. IfBudgetingis used and pinned, it SHALL be shared across baselines as well.ParityPinSetSHALL include the editions required by the referenced specification, comparator, and any measurement or normalization method in use (at minimumCNSpecRef.edition,CGSpecRef.edition,ComparatorSpecRef.edition). If the parity run depends on planned slot fillings, its exactParityPlanWorkPlan SHALL carry the relevant declaration-local A.15.3 rows inPlannedFillingRows[](nil-elision when not applicable). Each row resolves only inside that WorkPlan and has no independent reference, kind, or edition. -
CC‑G9.2 — Mode‑specific definition pins are declared via Extensions (local; conditional). When parity depends on mode‑specific definition records beyond the pinned governing spec refs (e.g., DHC/QD/OEE), the ParityPlan/Report SHALL include the corresponding
GPatternExtensionblocks and satisfy theirRequiredPins/EditionPins/PolicyPins(typically carried insideParityPinSet, and echoed via pins deltas in audit):- DHC parity →
G.9:Ext.DHCParityPins - QD archive parity →
G.9:Ext.QDArchiveParity - OEE parity →
G.9:Ext.OEEParity
- DHC parity →
-
CC‑G9.3 — CSLC-admissible orders and arithmetic (delegation point + local constraint). Delegated to
CC‑GCORE‑SET‑1(and the relevant G.5PortfolioMode/ selected-set semantics). Additionally: any numeric comparison or aggregation invoked by parity SHALL be CSLC-admissible and cite the corresponding CG‑Spec entry; non-admissible operations (e.g., ordinal means / mixed‑scale weighted sums) SHALL be refused or abstained with path‑cited trace (citation only; arithmetic admissibility comes fromCG‑Spec/MM‑CHR). -
CC‑G9.4 — Normalization discipline (local citation only). If Characteristics differ by unit, scale, or space, the ParityPlan SHALL cite the CSLC-admissible comparability mapping by id (
UNM_id?,NormalizationMethodId[]?,NormalizationMethodInstanceId[]?) and compare only after that mapping is applied (“normalize, then compare”). If such mapping ids are used, the ParityReport SHALL echo the same ids (directly or via explicit pins deltas) so the run is reproducible and auditable without unrecorded information. The harness SHALL NOT define a local mapping. -
CC‑G9.5 — Dominance/PortfolioMode interpretation & telemetry separation (local).
ParityPlanandParityReportSHALL either pin the applicable dominance regime and portfolio mode through explicit references and policy ids, or cite their corresponding defaults inG.Core.DefaultGoverningDefinitionIndex. Any non-default promotion behaviour must be bound to a policy and recorded through its policy-id pin.IlluminationSummary, coverage, and regret SHALL be treated as telemetry (report-only by default); any promotion into dominance is an explicitly pinned CAL policy and MUST be recorded in the audit pins and SCR.5a. CC‑G9.5a — Adaptation parity disclosure (local; conditional). When the parity claim concerns bounded specialization, the ParityPlan and ParityReport SHALL pin the declared task family or target scope cut, the work-measure threshold target, adaptation budget, prior exposure, and the performer and assistance relevant to the claimed capability. Retain the basis for any transfer, retention, later exploitation efficiency, downside or corridor-entry claim that affects the comparison.
-
CC‑G9.6 — Epsilon‑front thinning (local; conditional). If ε‑front thinning is used,
EpsilonDominance (ε≥0)SHALL be explicit in the plan/report and pinned (param/id) such that the same ε is reproducible. -
CC‑G9.7 — Crossing visibility (delegation point). Delegated to
CC‑GCORE‑CROSS‑1andCC‑GCORE‑PEN‑1. This item remains as a stable delegation point for Bridge and reference-plane crossing visibility plus R-channel penalty placement discipline. -
CC‑G9.8 — Report replay and evidence trace completeness (local). A ParityReport SHALL carry the
ParityPlanRefandBaselineBindingRefused for the run, echoTaskSignatureRefand any requiredTaskMapRefwhen the selector branch was used, and include an EvidenceTrace withEvidenceGraphIdand the relevantPathId[](andPathSliceId?when needed), covering inclusions, refusals, abstentions, and degradations. If the historical plan edition or binding cannot be resolved, return that unresolved input instead of substituting a current edition. -
CC‑G9.9 — Telemetry hooks are emitted with pins (local). When parity emits telemetry for refresh, emitted telemetry SHALL carry the active edition pins and policy‑ids needed to re‑run parity (including the active subset of
ParityPinSetrelevant to the emitted event). In particular, telemetry items SHOULD citePathSliceIdwhen available, and SHALL include the policy id governing the telemetry interpretation. Mode‑specific definition pins SHALL be included as declared by the activeExtensionsblocks (e.g.,G.9:Ext.QDArchiveParity,G.9:Ext.OEEParity, includingEnvironmentValidityRegionIdwhen OEE parity is in scope). -
CC‑G9.10 — RSCR parity tests are published (local). Parity publication SHALL include RSCR parity tests (via
F.15harness refs) that cover negative/refusal paths relevant to this plan (missing pins, edition drift, missing bridge calibration refs, etc.). -
CC‑G9.11 — GateCrossing visibility (delegation point). Delegated to
CC‑GCORE‑CROSS‑1and the applicable GateCrossing/CrossingBundle harness checks (E.18,A.21,F.9, and relevant Part G bridge or crossing wiring). -
CC‑G9.12 — Tech‑register lexical discipline (local). Tech prose and heads SHALL follow E.10: do not introduce drift‑prone primitives (e.g., “metric” as a Tech primitive); reference the source pattern’s canonical terms and pinned refs.
-
CC‑G9.13 — MOO disclosure for parity (local).
Run_Parity/Publish_ParityReportSHALL record the ParityHarness identity (UTS ids) and the active pins required to interpret the outcome (editions + policy‑ids), so the reader can recover the harness and pins from the parity record itself. -
CC-G9-CLP-1 - Causal method rung parity. For a report comparing causal methods, the comparison SHALL begin with
CausalRungParityScreen. When full parity remains plausible, the report SHALL declare target causality-ladder rung, causal-use claim kind,causalEstimandRef, interventional-action basis, causal support-component refs, exact transport endpoints and transportability result when needed, estimate result when needed, bridge and loss where rungs differ, andcausalUseSupportResultRefwhen relevant C.28 support is consumed. Where parity cannot be established, the report SHALL state degraded parity or abstention.
G.9:7 — Anti‑patterns and remedies
- AP‑1 Hidden edition drift. Remedy: require edition pins in
ParityPinSet; treat changes as RSCR‑relevant via canonical trigger kinds. - AP‑2 Baseline set is informal prose. Remedy: require
BaselineBindingRefand EvidenceTrace pins. - AP‑3 Comparator semantics are “whatever the code did”. Remedy:
ComparatorSpecRef.edition(and any normalization/comparability refs) must be cited and pinned. - AP‑4 Cross-sense or reference-plane reuse without its obtaining relation and visible pins. Remedy: recover the exact F.17 cells and cite the obtaining F.9 relation when local meanings differ; cite the exact reference-plane crossing basis and visibility records when planes differ (delegated to G.Core).
- AP‑5 Parity report becomes a hidden scoring sheet. Remedy: preserve CSLC-admissible outcome shape and keep telemetry as telemetry unless explicitly policy‑promoted by the governing policy pattern.
- AP‑6 “Metric” as a primitive in Tech. Remedy: name the exact
CharacteristicRef,ScaleRef,UnitRefwhen applicable,DHCMethodRef,MethodRef, and the C.16 reading content or its result episteme; addDistanceDefRefonly when used. “Metric” may appear only in Plain with a pointer to those canonical objects. - AP‑7 Hidden DHC replay drift. Remedy: carry every active field of the C.21 replay basis and refuse parity reuse when a required field is unresolved or differs across the compared readings. Register refresh tests only for a named receiver that consumes those changes.
G.9:8 — Archetypal grounding
Show-A — Multi-tradition parity for decision systems.
ParityPlan pins a rolling evidence window and comparator refs; ParityReport publishes a selected-set outcome plus the evidence trace. Family labels such as preference-learning comparators, causal decision pipelines, offline-RL evaluation pipelines, and robust BO-style selectors remain illustrative until a G.2 SoTA pack or named current source pins the exact family being compared. The parity report must preserve the selected set or partial order rather than collapse everything into a single scalar.
Show-B — QD parity (for example, CMA-MAE arXiv:2205.10752 and DQD arXiv:2106.03894).
ParityPlan pins descriptor/distance definitions and archive insertion policy editions. ParityReport identifies the compared archive states, records the lawful comparison result and telemetry deltas, and keeps illumination summaries separate from dominance unless explicitly promoted.
Show-C — Open-ended parity (POET arXiv:1901.01753 as an environment–method co-evolution lineage).
ParityPlan pins transfer rule editions and exploration policy refs. ParityReport publishes the comparison and any declared selected-set outcome with transfer-keyed traces (PathSlice). A policy change identifies the dependent uses for G.11 to reconsider.
Show-D — Causal method rung parity.
A team compares an observational predictor, an intervention optimizer, and a counterfactual policy strategy under one “best causal method” headline. The team first applies CausalRungParityScreen. If differences in rungs, support components, estimands, endpoints or outcome windows make the proposed comparison untenable, it reports degraded parity or abstention without opening the full record.
When full parity remains plausible, G.9 requires CausalMethodRungParityRecord. For each method, the record states targetCausalUseClaimKind, target CausalityLadderRung, causalEstimandRef, interventional-action basis, the support components consumed, relevant C.28 support result, follow-up window, outcome measure, changed transport endpoints, and estimate-result basis.
When fields differ, the report names the applicable declaredCausalityLadderBridgeOrLossRef and any available transportability or estimation references. It reports degraded parity or abstention where parity cannot be established. The admissible output may be a selected set by comparable rung, not one scalar winner.
G.9:9 — Cited Records (what this pattern publishes)
Exports (UTS‑publishable, edition‑pinned):
ParityPlanand itsParityPlanRef(oneU.WorkPlanepisteme and immutable edition reference; any planned-filling rows remain declaration-local content)ParityReport(UTS publication record carrying the plan and baseline-binding refs; work-result or audit-facing publication record only when the neighboring source relation is live)- DRR and SCR refs by id and, when applicable,
PortfolioPackRef?and selector-output refs by id, for downstream consumption. - Telemetry pins and events by id, for refresh wiring (
G.11) and RSCR harnesses (F.15).
G.9:10 — Relations
C.27 temporal-claim relation.
- C.27 may flag: dynamic parity when a benchmark compares rate-change, rhythm change, recovery speed, intervention effect, effort budget, or dynamic outcome.
- This pattern keeps: baseline, freshness, comparator edition, effort/budget parity, bridge discipline, parity plan, parity report, and reproducible benchmark publication.
- Non-admissible use: faster improvement is not benchmark superiority, and
dyn2BenchmarkParityBlock?is a benchmark input declaration, not a benchmark harness. - Exit: when live, recover
dynOrderCompared, baseline window, adaptation or intervention window, effort or budget parity reference, rate or rate-change measure,G9ParityPlanRef, and optionalG9ParityReportRef; G.5 is relevant only if a selector-facing result declaration consumes such a benchmark result.
C.29 mathematical-lens use relation.
- C.29 may flag: parity or benchmark input whose comparator, distance, descriptor geometry, embedding, normalization, surrogate model, learned representation, parity measure, model-family label, or model-selection basis depends on a mathematical lens that changes the parity claim and is missing, under-specified, or overread.
- This pattern keeps: baseline set, freshness, comparator edition, normalization ids, bridge discipline, parity plan, parity report, and reproducible benchmark publication.
- The
C.29mathematical-lens account is a cited input to parity.G.9governs the benchmark report and conclusion, andG.5governs any selector-facing output declaration. Parity-measure admissibility comes from the citedCG‑Spec/MM‑CHR. - C.29 application: for an under-lensed or overread parity input, cite the applicable
C.29output for the stated use:NoMathLensUseNeeded,MathLensUse.LensCandidateNote,MathLensUse.OneLine,MathLensUse.MiniCard,MathLensUse.FullCard, orNeighborGoverningPatternNote. Use the least costly adequate output that changes the next admissible parity use; full-card work is only required when the live parity or benchmark claim needs it.
Builds on: G.Core, G.5, G.6, G.4, F.15, E.17, E.18, A.21, F.17, E.5.2, E.10.
Publishes to: UTS (plan/report ids), G.11 (refresh wiring), G.10 (shipping publication form; parity records are cited records).
Uses: G.0, A.19, A.2.6 for exact U.ClaimScope, F.9, and C.28 when parity compares causal methods or causal-use claims.
Uses (optional, via Extensions): G.7, C.18 and C.19 (QD/OEE wiring), C.23 (SoS‑LOG narration and failure‑policy pins).
G.9:11 — Working reading checks
- If two baselines are being compared under different freshness windows, comparator editions, or silent normalization rules, this pattern has not yet been satisfied.
- If parity cannot tell the reader what was held constant, what remained telemetry, and what crossings or penalties were active, the report is not yet usable.
- If a scalar winner is being claimed where only a selected set or partial order is CSLC-admissible, parity is overclaiming and should publish the CSLC-admissible outcome shape instead.