C.19 - Explore-Exploit Live-Pool Governor
Type: C-pattern Status: Stable Normativity: Normative
Plain-name. Explore-exploit governor.
Intent. State and test exploration and exploitation policy over live candidate pools so frontier treatment, graduation, narrowing, and sunset treatment stay explicit and auditable. A C.19 result governs pool treatment only. C.19 does not generate candidates; use C.19:4.4 when the next result requires another operation.
C.19:0 - Use this when
Use C.19 when several candidate lines, family regions or frontier segments remain live and the question is how to treat that pool: widen exploration, keep the frontier, narrow to a subset or sunset a line. Name the governing policy and what would justify changing the treatment. Use C.11 when the remaining question is a local choice among fixed options.
If a proposed pool-policy premise is expressed as learning progress, information gain, novelty, or an articulated former cue, recover its result owner first. Use E.10.LRN only while learning wording hides that result, A.10 when the needed evidence or source basis requires a replayable account, C.17 or C.18 only when the question concerns characterization or possibility-space change, C.11.CRC only when a finite configuration-relative comparison is missing, and C.11 for local option or probe choice. Stop before C.19 unless the remaining question is policy over a live pool.
C.19:0.1 - What goes wrong if missed
- scalarized top-1 picks are mislabeled as “the frontier”, so it becomes unclear whether the result names one lens-ranked winner or the admissible live set
- exploration continues without one named pool, one named governing lens, or one explicit next treatment
- local option choice, pool policy, enactment planning, selector-result declaration, and publication availability collapse into one blurred result
C.19:0.2 - What this buys
- one explicit pool-governance result for exploration, graduation, narrowing, and sunset treatment
- one explicit link from lens or policy state to the next pool-side treatment
- one repeatable way to preserve heterogeneity and frontier discipline without forcing inadmissible totalization
C.19:0.3 - First-minute questions
- Which live pool, frontier segment, or family region is under governance?
- Which lens or policy state is governing it?
- Is the next admissible pool treatment to widen, keep the frontier, narrow to a subset, or sunset a line?
- If the question concerns none of those treatments, which subject pattern applies, and why is it no longer pool policy?
- What justifies active exploration or cheaper retention now, and what change would end that justification? Readiness for exploitation is a separate question.
C.19:0.4 - First output
For loop-engineering practice, use this first output only when the question is pool policy over live candidates such as loops, harnesses, workflows, method families, or framework seeds. A C.19 record may say that the pool should widen, keep its frontier, narrow to an internal subset, or sunset a line under a declared lens. If the question leaves pool policy, retain the established result or qualified open question and use the handoff in C.19:4.4.
When a treatment is supported, state the live pool, governing policy, treatment and the event that would justify changing it. A sufficient ordinary explanation can complete that result. To record it for reuse, use governingLens and one currentTreatment token from the closed set widen | keep_frontier | narrow_to_subset | sunset_line. When a missing fact or commitment can change the treatment, report what is established and identify what would settle the choice. Use nextQuestionPatternLocator from C.19:4.4 when another question has become current.
A claim-bearing account of the pool treatment is an episteme under C.2.1; identify its EntityOfConcern and claim content when that identity matters. An existing adequate account can supply it.
If the account cannot identify the pool, policy, treatment and relevant change trigger, state the missing basis; that pool-policy question remains open.
C.19:1 - Problem frame
C.19 describes policies and lenses for treating an identified live pool. Use the applicable C.18 generation, archive or front account when the treatment depends on it; the existence of a record is not an entry condition.
C.11 concerns local choice among a fixed OptionSet; C.19 concerns treatment of several live candidate lines, family regions or frontier segments. Either question can arise without first completing the other.
Immediate failure indicators for this pattern:
- the pool-policy result cannot name the live candidate pool whose treatment it states
- the governing lens or policy state is missing
- the next pool-side treatment exists only as one vague promise to continue exploration later
Use C.19 while the question concerns treatment of a live pool; otherwise use the applicable exit in C.19:4.4.
C.19:2 - Problem
Ad-hoc exploration mixes ordinal and interval claims, silently scalarizes partial orders, and loses lens or policy provenance, undermining admissibility and reproducibility.
C.19:3 - Forces
- Readiness vs. continuation — exploitation needs its direct qualification, while active exploration and cheaper retention need their own prospective contribution and affordable commitments. None of these judgements follows from the others.
- Heterogeneity vs. focus — fairness quotas by family vs. depth on proven lines.
- Lens expressiveness vs. audit — scalarised choices must not be called ‘the frontier’ and MUST record lens ids.
C.19:4 - Solution
Decide which lines remain worth exploring, which remain worth retaining without new probing, and which no longer warrant a place in the live pool. Use the pool’s intended contribution and horizon, the resources and opportunity window still available, and what keeping each line displaces. State the resulting treatment and the change that would make it worth reconsidering.
Judge readiness for exploitation separately. A line may warrant further research or cheap retention before it supports deployment or transfer. Conversely, neither missing readiness nor past expenditure justifies keeping it indefinitely. Retaining an archive record under C.18 need not keep the line active or fund another experiment.
C.19:4.0 - Develop the continuation basis when needed
Use an adequate policy, analysis or professional subject Method directly. When the prospective contribution or feasible continuation is not yet clear, begin with the candidate artifact, source account, failed attempt or possible receiving use that is available. B.5:4.1.2 supplies the constructive examination; the connection below returns its result to the pool-treatment question.
- Follow what the candidate could enable. Inspect how an existing artifact or source example takes an input to a result, then how that result could be used. For a failed attempt, find where the observed trace first departs from the expected operation. Use the subject Method, a worked application or a competent contributor to explain what that difference could change. A research line may instead aim to explain an obstruction, develop a model or make a new question answerable; name that attainable kind of contribution at its research horizon.
- Derive a condition or question from that relation. Follow the proposed effect forward and work back to the inputs and contributions it requires. Inspect the corresponding interface, assumption, observation or arrangement in the available material. Where the relation suggests a revealing contrast, examine a case with and without that premise, or make a small trial that can distinguish them. Keep the established consequence separate from an untested interpretation. If the relation itself cannot yet be explained, use B.5:4.1.2 to construct a smaller case or obtain the specific subject assistance needed; the useful first result may be a better question.
- Make the proposed continuation attainable. Walk the remaining performance from available input to needed output. For an operation that can change feasibility, use its Method to determine the preparation, material, competence and duration it requires. Place those demands in the available time windows and compare overlapping demands on the same people, setup or access with ongoing work and other live lines. Read the allocations or agreements supporting those commitments, or obtain the responsible participant’s answer. This identifies displaced work and unresolved dependencies. An earlier account can supply these facts.
- Compare the commitment with the contribution. Apply C.11.DUA:4.2 and 4.2.1 to active inquiry, continuing on the present basis, lower-burden retention and other available continuations. Include the work of obtaining and adapting the Method, the execution and retention demands recovered above, and consequential waiting or checking. Use qualitative or ordinal comparison when it suffices. An exploratory attempt can earn a bounded allocation through the understanding or possibility it could develop, before a particular discovery is known.
- Return to the treatment decision. Bring the supported contribution, its conditions, the feasible commitment and the comparison back to the pool policy. Use them to justify a treatment or to preserve the dependent judgement still open. When an unknown condition matters, identify the observation, source, agreement or construction that could settle it and whether obtaining it is worthwhile. When no useful continuation can yet be described, identify the missing relation or subject contribution and compare a bounded attempt to obtain it with retention, deferral or ending the live commitment.
C.19:4.1a - Causal data and causal-policy exploration hook
When exploration collects data for a causal claim, learns or evaluates a causal policy, or uses counterfactual replay as a reason to treat a live line, the pool-policy result stays within C.19 and cites C.28 for the causal-support conclusion.
Optional PoolPolicyResult.causalUseSpec?:
PoolPolicyResult.causalUseSpec?:
causalUseQuestionRef?: CausalUseQuestionRef
targetCausalityLadderRung: CausalityLadderRung
causalUseClaimKind: CausalUseClaimKind
causalActionPolicyClass?: CausalActionPolicyClass
causalSupportComponentRefs?: CausalSupportComponentRefs
causalUseEvidenceDesignRef?
offPolicyCausalEvaluationResultRef?
causalUseSupportResultRef?: CausalUseSupportResultRef
supportedUse
unsupportedUse
Omit this tail when the pool treatment makes no causal claim and consumes no causal-support result. Include it when effect, counterfactual replay, causal-policy support, or causal evidence changes the treatment. The C.28 result remains evidence support; it does not authorize ranking, retirement, deployment, or graduation. C.19 makes the pool-treatment decision under its own policy.
Policy fields. EmitterPolicy is a context-local, versioned policy with canonical fields:
{ emitterPolicyId, name?, regimeKey ∈ {UCB, Thompson, BO-EI, GP-UCB, PES, InformationGain, …}, params, explore_share∈[0,1], temperature τ≥0, rebalance_period, wild_bet_quota≥0, graduationConditionRef?, assuranceResultRef?, epsilon_dominance ε, cell_capacity K, insertionPolicyRef, dedupThreshold, deduplicationBasisRef, deduplicationUnit }.
graduationConditionRef cites the direct domain or policy condition for moving a line into exploitation or extending an already supported use. assuranceResultRef is present only when satisfying that condition relies on one exact B.3 result for a named assurance use and bounded scope. Neither field is an assurance level.
The pool policy separately states why active exploration or retention is worthwhile, what continuing commitments it needs, and what would defeat that basis. An exploration horizon may extend beyond the next local decision; it still needs a defensible prospective contribution and obtainable resources.
emitterPolicyId is cited as emitterPolicyRef; the profile is not a U-kind, generation operator, staffing instruction, budget approval, or Work record.
Decision-subject clarification. Attribute any later choice to one declared DecisionSubject at explicit DecisionSubjectGranularity. Record measurement spaces and admissible policies in the semantic-frame epistemes that state them. Use LOG to describe lenses and policies; that description does not enact a choice.
EmitterPolicy use. A C.18 generation or archive account cites this profile only when pool treatment, insertion or deduplication uses it.
Use the ordinary default tokens defined in G.Core and G.5. The rules below explain their pool-policy consequences.
Decision-theory bridge. Use C.11 for theory-side choice among already-available options and for the meaning of ProbeBudget, ValueOfInformation, and ValueOfComputation. A pool-policy record may use those outputs as inputs to its treatment judgement. A probe’s lack of value for the next local choice does not by itself settle the value of longer-horizon exploration; that prospective contribution and its opportunity cost belong to the pool policy. A worthwhile research line need not become a prerequisite for the present decision.
Ordinary default references (if policy is unspecified):
- Dominance: consume
DefaultId.DominanceRegimefromG.CoreandG.5; in ordinary Q-front use this means{Q components}withConstraintFit=passas eligibility gate. - Tie-breakers: the current policy may use a Novelty coordinate,
DeltaDiversity_P/ΔDiversity_P, Surprise, or Illumination only when it names that tie-breaker. Unused optional tie-breakers need no result.- For Novelty, cite each bearer’s coordinate-result episteme: a complete C.16 measurement result for a measured value, or a C.2.1 ascription when the declared rule permits that reading. Before comparing bearers, confirm compatible Novelty Characteristic and Scale editions, corpus/reference set and inclusion rule, similarity Method and encoder/model editions, ClaimScope, and the window, uncertainty and support required by the construction and comparison.
- For Surprise, cite the coordinate result and its generative-model and training-basis editions, Scale, ClaimScope, and the window, uncertainty and support required by its construction and comparison.
- For
DeltaDiversity_P, cite the retained set, candidate, calculation or measurement rule and Scale editions, descriptor or distance basis, resulting marginal reading, and the window and support required by the use. - Illumination remains telemetry over
Diversity_Punless the named policy explicitly promotes it. A promoted use still cites the report and its measurement basis. The words Novelty, Surprise, and diversity alone are not executable policy inputs.
- Archive:
K=1,ε=0, deduplication inCharacteristicSpace. - Policy family: one uncertainty-aware explore policy family with one declared regime key and explicit change triggers;
UCB-class with moderate temperature andexplore_share ≈ 0.3–0.5is one didactic starter profile, not the semantic default family. - Provenance (minimum for the current use): record
DescriptorMapRef.editionandDistanceDefRef.editionwhen their definitions are used. A consumed DHC coordinate carries the active fields of C.21’sDHCReplayBasis, includingDHCMethodRef.edition. CiteemitterPolicyRefandinsertionPolicyReffor the policies used. An applied deduplication threshold keeps scalardedupThreshold,deduplicationBasisRef, anddeduplicationUnit, whether supplied by a cited policy or stated here. RetaintimeWindowandseedswhen they affect replay.
Use-value and declared-Q boundary. C.16.Q is the pattern for the selector-context meaning of use-value and its Objective form. When use-value participates in the current Q, declare QS.UseValue as an objective head in that exact Q and cite the current Q/comparator basis. When it does not participate in the current Q, keep the use-value criterion explicitly outside Q as a declared side condition or tie-breaker. A pool-policy record may use either declared position but cannot silently promote use-value into Q or construct the Q model.
Scalarization lenses (policy‑level). A lens J_ℓ declares: (a) hard eligibility conditions (e.g., ConstraintFit=pass), (b) soft aggregation (weights or curves), (c) trust policy (how any applicable assurance result and any declared CL discount enter).
Conformance. A pool-policy record MUST name the lens used to pick from a frontier; scalarized rankings MUST NOT be presented as “the frontier”; the lens id MUST be recorded in the provenance of each selection.
Promotion rules (policy).
- Tie-breaks. Use only the constituted and compatible results named by the current policy. Promotion of Surprise or Illumination into the dominance set MUST be declared by lens or policy id and captured in provenance.
- Graduation. A candidate line or pool member moves from Explore to Exploit only when eligibility holds and the direct condition cited by
graduationConditionRefis satisfied. When that condition relies on assurance,assuranceResultRefcites the exact B.3 result whose named use and bounded scope support the judgement. An optional profile may supply evidence; neither the profile nor a label graduates the line. - Continue, retain, narrow or sunset. At
rebalance_period, judge the line’s prospective exploration or stepping-stone contribution against the remaining opportunity, obtainable resources, retention burden and displaced lines. Keep active exploration only while that commitment is warranted; cheaper retention may remain worthwhile without new probing. Narrow, pivot or sunset when the applicable continuation basis no longer warrants the current treatment. An unsatisfied graduation condition blocks the use it governs, not continuation by itself. A defeated continuation basis can justify retirement even if much has already been spent. The optional profile remains evidence, not the treated object. Pool-policy operation. Determine the treatment of the identified live pool. Obtain a needed generation, archive, front or implementation result through the corresponding operation in §4.4.
Pool-policy pass (per rebalance_period).
- Identify the live pool. Read any C.18 archive/front account on which the treatment depends and retain its replay boundary; front or archive recomputation remains a C.18 question.
- Recover the governing lens and the continuation basis; use §4.0 when that basis still needs construction. Record the policy values used, such as
explore_share, emitter-profile preference,wild_bet_quota, or an admitted heterogeneity constraint. - Apply the conditions for the proposed treatment. For exploitation or a wider supported use, apply eligibility and
graduationConditionRef, citing the bounded B.3 result when assurance is needed. For exploration or retention, assess the continuation basis and the whole pool’s competing commitments. Return the supportedcurrentTreatmentfromwiden | keep_frontier | narrow_to_subset | sunset_lineand state whether a retained line warrants active probing or only lower-burden retention. If a needed premise is unresolved, retain the supported account and identify the dependent treatment judgement that remains open. - If that judgement requires fresh candidates, a changed emitter mix or temperature, archive insertion, or front recomputation, set
nextQuestionPatternLocator = C.18and pass only the desired emitter profile, quota or constraint, and the generation/archive/front reason. Use §4.4 to obtain the needed generation, archive or front result. - If carrying out the treatment requires dated implementation, planning, staffing, or budget use, pass the policy record to the A.15 family; the policy record itself grants none of them.
- State the supported conclusion or the dependent question left open. When a reusable account of a settled treatment is needed, use
PoolPolicyResultwithlivePool,governingLens,currentTreatment,changeTrigger, and the inputs required by the next subject pattern. A qualified open question is not recorded as a settled treatment.
Named lenses (policy‑level) The following lens profiles are illustrative heuristics, not norms. Practitioners MAY reuse or modify them.
- Frontier‑sweeper — maintain attention on the full front; promote only when the direct graduation condition holds.
- Barbell — enforce
explore_share ≥ θwith awild_bet_quota; otherwise exploit top‑trust region. - Spike‑first — pick highest Use‑Value subject to
ConstraintFit=passand a small Cost‑to‑Probe cap. - Safety‑first — minimize SafetyRisk subject to
Use‑Value ≥ θandConstraintFit=pass. - Platform‑option — maximize Option‑Value under probe cost bounds.
- Pilot-then-scale — optimize Use-Value on the declared pilot scope. Set
currentTreatment = widenonly whenassuranceResultRefcites the exact B.3 assurance result whose supported scope includes the proposed wider pool, andchangeTriggernames the satisfied assurance condition and that newly supported scope; otherwise keep the pilot scope. - Heterogeneity-first (illustrative profile). Use only when the applicable policy already admits a heterogeneity constraint or sampler policy. The applicable policy may declare a
FamilyCoverageorMinInterFamilyDistancegate, a family or subfamily quota, or a diversity-promoting sampler. Use that policy’s thresholds, quotas and sampler rules, and record only the policy values and IDs used.
Conformance (lens recording). A pool-policy record that uses a lens MUST record its lens id. Include emitterPolicyRef when the current treatment uses that profile.
C.19:4.1 - Explicit pool-policy result
Canonical record vocabulary. A serialized PoolPolicyResult uses the field governingLens and exactly one currentTreatment token from widen | keep_frontier | narrow_to_subset | sunset_line. Reader prose may say widen, keep the frontier, narrow to a subset, or sunset a line, but those phrases are labels, not alternate serialized values. Do not use lens as a second field name.
When the policy and available evidence support a treatment, state the treatment and its grounds. A record preserves that result for later inspection or reuse. An inquiry can also establish why the choice of treatment remains open (§4.3).
That result should state:
- the live pool, frontier, or family scope under governance;
- the governing lens id or policy state;
currentTreatment, chosen fromwiden | keep_frontier | narrow_to_subset | sunset_line;- the event or threshold that would justify changing that treatment next.
A result may therefore state, for example:
livePool = frontier_FgoverningLens = barbell_policy_v2currentTreatment = keep_frontierchangeTrigger = graduation_condition_v3 is satisfied for one retained line
or, for one narrower family region:
livePool = family_region_betagoverningLens = heterogeneity_firstcurrentTreatment = narrow_to_subsetchangeTrigger = quota satisfaction plus a compatible cited C.17 Novelty coordinate result clearing novelty_floor_policy_v2
Both examples identify the pool, policy, treatment and change trigger; their different scopes require different continuation grounds.
C.19:4.2 - Closure rule over the live pool
A supported pool-treatment conclusion identifies the pool and its next treatment. If a needed premise is unresolved, the inquiry may finish with the supported account and the dependent judgement left open, as in §4.3.
- Close as
widenwhen wider exploration is warranted under the pool’s contribution, horizon and resource conditions. If widening also asserts a wider supported use, satisfy that use’s graduation and assurance conditions. - Close as
keep_frontierwhen keeping the live lines is warranted under the current policy. A line retained for later reconsideration need not receive a new probe. - Close as
narrow_to_subsetwhen the continuation basis warrants a smaller internal live set, without implying that one scalar winner has already been chosen. - Close as
sunset_linewhen a line’s prospective contribution, feasible continuation or retention no longer warrants its burden under the pool policy. Failure to graduate is not enough. Its archived result may still be retained under C.18 when that separate retention remains useful.
When comparing two policies, hold the candidate lines, evidence and pool question fixed and apply each policy. Identify any treatment the change permits or excludes under those conditions. Applying the same policy to newly generated candidates or newly observed performance can also change the treatment. State which input changed so the next team can tell whether to revise its policy, update the pool or use the new evidence.
When the question has stopped being pool policy, carry the established result or qualified open question into the applicable handoff in C.19:4.4; the next pattern is recorded outside currentTreatment.
One internal retained subset here is still one pool-treatment result. It is not yet a declared Shortlist or RankedShortlist, and it has no ShortlistId merely by being retained. When a downstream use needs declaration or audience availability, use C.19:4.4.
If the account cannot identify which pool remains live, which lens and policy apply, or which event would justify changing the treatment, it has not established a settled pool-policy result. Preserve the supported part and the unresolved question.
C.19:4.3 - Minimal pool-policy record
The smallest useful C.19 record usually states:
livePool = ...governingLens = ...currentTreatment = widen | keep_frontier | narrow_to_subset | sunset_linechangeTrigger = ...nextQuestionPatternLocator? = ...only when the question is no longer pool policy- one or more native direct-owner reference fields only when an already constituted result or claim supports the treatment; retain the field name, kind, identity, claim episteme when applicable, and subject-pattern locator supplied by that owner—for example the A.2.2
capabilityHolderRefandcapabilityStatementRef; an information-gain or articulated-endpoint use keeps the ref name and kind defined by its own direct owner—rather than replacing them with one C.19 signal or cue a10RelianceRef? = ...when the needed evidence or source basis requires a replayable A.10 account; retain its relied-on claim, bounded pool-treatment use, evidence-provenance path, applicable window andRelianceDispositioncompetenceModelRef? = ...only when it cites one exact model episteme used by the pool policy; that model is neither the capability, the owner-defined result, nor proof that the treatment may rely on eithergoalSpaceExpansionPolicyRef? = ...only when one independently declared archive or curriculum expansion policy governs goal- or task-space growthassuranceResultRef? = ...when graduation, scaling, or widening relies on one exact B.3 assurance result and its bounded supported scopewhyNotLocalChoice = ...when the result might otherwise be mistaken forC.11
A record may therefore read:
livePool = frontier_F
governingLens = barbell_policy_v2
currentTreatment = keep_frontier
changeTrigger = graduation_condition_v3 is satisfied for one retained line
whyNotLocalChoice = several family regions remain live
When currentTreatment = narrow_to_subset, livePool still names one internal retained subset or one live pool subset. Retaining the subset does not declare a public Shortlist or RankedShortlist or assign a ShortlistId. If selector-facing result declaration is now required, the C.19 record leaves currentTreatment as the last pool treatment and fills nextQuestionPatternLocator = G.5, with the reason that the next question concerns result declaration rather than pool policy.
Goal and task space growth is one pool-policy doctrine over the archive or curriculum side. When autotelic or capability-discovery pressure is active, cite goalSpaceExpansionPolicyRef only for an independently declared policy.
Retain every supporting information-acquisition, capability, novelty, objective, articulated-endpoint, or other result or claim through its native direct-owner reference and kind. Add a10RelianceRef when the needed evidence or source basis requires a replayable A.10 account. Use competenceModelRef only for one exact model episteme, never as an alternative name for the capability, result, or reliance account.
These inputs may support widen, keep_frontier, narrow_to_subset, or sunset_line; none becomes a generic signal or cue, default Q, dominance coordinate, probe choice, or selector-facing shortlist by entering the record.
When the available evidence leaves the treatment open, report the pool, applicable policy, remaining treatments and the premise that would decide among them. Reserve PoolPolicyResult and its currentTreatment field for a treatment that can already be justified.
C.19:4.3a - Worked closure slice
Four short contrasts keep the closure law practical.
Several family regions remain live. When the point is to keep several lines active under one declared lens, the pool-policy result must not imply that one local choice has already been made:
livePool = frontier_F
governingLens = frontier_sweeper_v3
currentTreatment = keep_frontier
changeTrigger = one retained line satisfies graduation_condition_v3
whyNotLocalChoice = three family regions remain live
A capability claim supports pool treatment.
An A.2.2 statement about holder diagnostic_agent_v4 supports its qualified ability in task region alpha; it does not establish transfer into region beta. In this constructed case, a declared curriculum-expansion policy warrants keeping both regions live while beta’s prospective contribution and retention burden remain acceptable. That policy does not establish transfer. In this case, the curriculum review needs a replayable account of the capability statement’s support and bounded use, so the record cites its A.10 reliance account:
livePool = diagnosis_task_regions_{alpha,beta}
governingLens = curriculum_expansion_policy_v3
currentTreatment = keep_frontier
changeTrigger = beta obtains qualified transfer support, its opportunity window closes, or its continuation burden changes
capabilityHolderRef = diagnostic_agent_v4
capabilityStatementRef = capability_statement_CS-44
a10RelianceRef = A10_CS-44_keep-frontier_W8
goalSpaceExpansionPolicyRef = curriculum_expansion_policy_v3
whyNotLocalChoice = both regions remain live; no individual task or probe is selected
capabilityHolderRef identifies the holder System; capabilityStatementRef identifies the C.2.1 episteme asserting that System’s bounded ability; and a10RelianceRef qualifies only the stated bounded reliance. None is renamed as a signal or cue, entered into the declared dominance set, or emitted as a ChoiceResult.
The same missing beta qualification permits different continuation decisions.
Hold the alpha-only capability evidence fixed in these three hypothetical policy conditions. None supports deployment in beta.
| Continuation basis | Pool-policy result | What that result supports |
|---|---|---|
| An informative beta simulation fits the available qualified setup and this month’s allocation; its prospective contribution warrants the work it displaces. | keep_frontier, with active exploration of beta. | Keep both regions live. Use C.11 for the particular probe choice and the relevant planning and authority rules for actual work. The simulation has not yet been performed. |
| No new beta probe now warrants its cost, but keeping the existing result and lineage costs little and preserves a named later reuse. | keep_frontier, with beta retained for reconsideration and no new beta probe. | Preserve the stepping stone under the existing retention basis. Reconsider when the reuse opportunity or maintenance burden changes; the retention decision creates no study assignment. |
| Further beta work would displace a better-supported line’s use of the only available setup; its prospective contribution does not warrant that displacement, and no useful lower-burden live commitment remains. | sunset_line for beta. | End beta’s live-pool commitment. Keep an archived result only if its separate C.18 retention remains useful. Reopen if a feasible route or changed contribution warrants the cost. |
A compatible Novelty floor or quota may affect these judgements when the policy justifies it for this continuation use. Its failure is not a universal retirement rule, and the unsatisfied transfer condition is unchanged across all three cases.
For the third condition, the record can state:
livePool = family_region_beta
governingLens = curriculum_expansion_policy_v3
currentTreatment = sunset_line
changeTrigger = a feasible beta route or changed prospective contribution warrants its continuation cost
whyNotLocalChoice = other regions still remain live under the same pool policy
Obtaining an initially missing beta basis. Start with the alpha and beta configuration artifacts, their code and input definitions, a beta trial log, a saved input capture and one shared test window. Read the beta output and the receiving analysis: the output has elapsed time and scores, while the receiving analysis joins scores to equipment events by event identifier. The failed trial reports unmatched rows.
Following the producer code shows that its projection after sorting omits the identifier. This yields a question: can the identifier be preserved through that operation so the receiving analysis can use the scores? The capture still contains the identifiers, and the replay Method works on that saved capture. A small replay can examine the mapping before using the shared setup.
Ask the receiving team which comparison would consume the joined data and read the corresponding protocol. That use supplies the reason to investigate the mapping; the diagnostic value of beta’s scores remains to be established by the comparison.
Recover the remaining replay work from its Method and compare its demands with the actual allocations. If replay fits available analysis time while the shared setup is committed to alpha, compare the limited beta inquiry with cheap retention; a full beta trial remains conditional on obtaining a suitable window. If no receiving comparison is yet established, retain that specific question and use the bounded examination in §4.0 when it is worth its burden.
The result may be a supported use for further comparison or a reason to end that attempt. These obtained facts and remaining conditions supply the pool-treatment comparison above; a changed allocation or receiving operation reopens the affected judgement.
The pool has already been narrowed and the next question is selector-facing result declaration. When one internal retained subset is already explicit and the next question is to declare it for downstream use, close the pool-policy question by naming the applicable pattern instead of presenting that subset as though it were already one selector result:
livePool = retained_subset_{option_B, option_C}
governingLens = selection_preparation_policy_v2
currentTreatment = narrow_to_subset
changeTrigger = eligibility, comparison basis, or retention burden changes for option_B or option_C
nextQuestionPatternLocator = G.5 because selector-facing result declaration is now current
whyNotLocalChoice = the next use needs declaration of the retained set, not a choice among its members
C.19:4.3b - Cultural and style live pools
For a cultural or style live pool, state how the variants remain live under one policy. The pool might contain style, tradition, Method-family, Work-family, canon, scene or technique variants; use the local kind. The same minimal record can preserve that account.
PoolPolicyResult:
livePool:
governingLens:
currentTreatment:
changeTrigger:
termBridgeRefs?:
nextQuestionPatternLocator?:
When source labels need disambiguation, use F.17 to identify their local senses. Use F.18 when a durable new name is needed. Source-local labels can retain their cited meanings.
Include termBridgeRefs only for an independently established F.9 relation between identified sense cells. That reference identifies the relation; a pool-treatment claim relying on it must also establish the named use, direction, correspondence rule and tolerated loss, with A.10/B.3 reliance when required. An existing adequate claim episteme can carry this account.
If the question becomes the cultural-evolution case, use C.36. Result declaration, publication availability and currentness use §4.4.
C.19:4.4 - Exit from pool treatment
When fresh candidates, a changed emitter mix or temperature, archive insertion or front recomputation are needed, use C.18 with the desired policy values and the pool-policy reason. C.18 keeps the generation, archive and front claims distinct. If obtaining the variants requires construction or change of usable material, use the generation Method, with C.40 when that development is the missing contribution; a generation record does not perform it.
When one evaluation suffices to answer whether one bearer or version should improve, use the applicable object evaluation, with E.22 when its framing is needed. When the object version will be improved through repeated passes under a declared evaluation, use E.23. For either exit, pass the exact bearer or version, objective or criterion, evidence, and pool-policy reason that made improvement current. A C.19 treatment is neither a generation operation nor an improvement result.
An internal subset retained by narrow_to_subset is still the live pool named by one C.19 policy record. It is not a public Shortlist, RankedShortlist, or ShortlistId-bearing selector artefact, and no public selector artefact is emitted by this pool-policy use. Front and Archive retain their C.18 meanings; a scalarized pick does not rename either one.
When the retained set must be declared for downstream comparison, registry use, or another selector-facing use, pass G.5 the exact declared source set, lens or policy id, eligibility conditions, dominance set, tie-breakers, promotion policy, and provenance pins. Use G.5 to declare the selected-set result and any stable public shortlist identity required by a named use. The C.19 account retains the preceding treatment or qualified open question and the reason for declaring the result. If the question also concerns audience availability, use E.17 for a source-backed publication face and return to source and E.24.PUB for the publication occurrence and availability.
When the question is which option to choose, pass the fixed option set and comparison basis to C.11. For planning or performed work, use the applicable A.15 pattern. Use C.24 after the action or option is fixed and tool or service calls must be planned. When the question concerns edition, source, descriptor, policy or evidence currentness, use G.11. Retain the established pool-policy result or the qualified open question in each handoff.
For the handoff, identify the pool question and preserve the evidence the receiver can use. For an established treatment, give the treatment and its change trigger. For an open choice, give the possible treatments and the fact or agreement that would decide between them. Include an archive or front reference when the account relies on one. The receiving pattern governs declaration, publication, choice, work or refresh.
C.19:5 - Archetypal Grounding
C.19:5.1 - System grounding
A product-search or architecture-search team can keep several family regions alive even after one line looks best locally. keep_frontier under frontier_sweeper_v3 is warranted while those regions’ prospective contribution and commitments fit the pool policy. Graduation is reconsidered when graduation_condition_v3 is satisfied; continued exploration or retention is reconsidered when its own basis changes. The team need not choose between premature exploitation and indefinite funded search.
C.19:5.2 - Episteme grounding
A SoTA pack often compares traditions that stay non-dominated for different reasons: one clears current evidence quality, one keeps broader transfer value, one preserves family coverage. The C.19 result is then often keep_frontier or narrow_to_subset, rather than a scalarized winner.
C.19:5.3 - Collective and contextual grounding
A regional or stakeholder-diverse pool may have to sunset one line while keeping others alive to preserve coverage, fairness quotas, or contextual fit. Use C.19 to state and test that pool-treatment decision only while the question is still about the live set. Once the result must become one local choice, one enactment plan, or one declared selected set, apply the pattern that defines and tests that result.
C.19:6 - Bias-Annotation
Keep partial orders unless the applicable policy justifies scalarization. Apply only operations supported by each Scale; an ordinal value supplies no interval or ratio arithmetic. Preserve lens and policy identity for selections, independently of notation or tool.
C.19:7 - Conformance Checklist
-
C19-1 When a C.18 generation or archive record relies on a named C.19
EmitterPolicy, it SHALL cite that profile inemitterPolicyRef?. If the active insertion policy is not inherited, record it ininsertionPolicyRef?. If the deduplication threshold is not inherited, record scalardedupThreshold?together with itsdeduplicationBasisRef?anddeduplicationUnit?; never encode that scalar as a reference. These fields are not required without that policy dependence. -
C19-2 The characteristic set and indicators used for dominance MUST be declared and eligibility conditions applied first. If use-value participates in current
Q, the record cites the C.16.QQS.UseValueobjective head in that Q; otherwise it states that the criterion remains outside Q. (References to C.18 generator operators are descriptive only; LOG exports no Γ.) -
C19-3 If a lens is used, its id MUST be recorded; do not label scalarized top-1 as “frontier”.
-
C19-4 Promotion of
SurpriseorIlluminationinto dominance MUST be explicit in policy. -
C19-5 A pool-policy record creates no
SystemRoleAssignmentStateRelation, system-role assignment, permission, plan, budget or Work occurrence. For each such claim needed when implementation follows, establish its context, scope and independent basis under the applicable subject rule. Precise role, assignment or assignment-state claims need their respective grounds; ordinary implementation wording does not require asserting all of them. -
C19-6 When a pool-treatment lens uses dominance or tie-breakers, it MUST document or cite the applicable
Eligibility → Dominance → Tie-breakerscomparison. Eligibility is for the affected use. Continuation and retention without that comparison follow §4’s contribution, commitment, and opportunity-cost basis; exploitation readiness is separate. For every tie-breaker used, cite a constituted result with the compatible basis required above; unused optional tie-breakers need no result. Any promotion of Surprise or Illumination into the dominance set MUST be named by lens or policy id and recorded in provenance. -
C19-7 (pattern-change boundary). A project-local choice or revision of an
EmitterPolicy,DescriptorMap,DistanceDef, sampler, quota, orδ_familythreshold stays under C.19 and the decision or result that consumes it; it does not invoke E.15 merely because a profile changed. When the definition is changed in an existing FPF pattern edition, use E.15 to compare the exact predecessor and candidate, classify the effect, and repair dependent consumers. Compare materially plausible candidate definitions when the question requires it. Use C.18 for their generation/archive account, C.40 for missing construction and C.19 only when their live-pool policy is the question. No default heterogeneity quota or sampler is defined here. Keep the policy and card ids in the existing decision, change, or SCR result that needs them; create no separate authoring trace. -
C19-8 When a heterogeneity-first profile is used, provenance MUST name each admitted heterogeneity constraint and its governing policy id. If a family or subfamily quota applies, record the exact quota vector and family-definition id; if sampling applies, record the sampler class, seed when relevant, and sampler-policy id.
-
C19-9 A
PoolPolicyResultMUST identifylivePool,governingLens,changeTrigger, and exactly onecurrentTreatmenttoken fromwiden | keep_frontier | narrow_to_subset | sunset_line;lensand space-separated treatment spellings are not alternate record fields or values. -
C19-10 If the question under repair is local option choice, an enactment-facing plan, selector-facing result declaration, or publication availability,
C.19MUST name the applicable pattern rather than restate it:C.11, the applicable A.15 pattern (C.24only for tool or service call planning after the action is fixed),G.5,E.17, orE.24.PUB. -
C19-11 If goal- or task-space expansion, autotelic pressure, or capability-discovery support is used, the record MUST cite
goalSpaceExpansionPolicyRefonly when one independently declared policy governs the treatment; retain each supporting result or claim under its native direct-owner reference, kind, identity, claim episteme when applicable, and subject-pattern locator; adda10RelianceRefwhen the needed evidence or source basis requires a replayable A.10 account; and usecompetenceModelRefonly for one exact model episteme. None of these inputs becomes a generic signal or cue, default dominance coordinate, probe choice, or selector result merely by supporting the treatment; any dominance promotion still requires the explicit lens or policy rule and provenance required above. -
C19-12 If exploration collects data for a causal claim, learns or evaluates a causal policy, or treats counterfactual replay as support,
PoolPolicyResult.causalUseSpec?MUST carry the target rung, claim kind, available support-component refs, supported use, unsupported use, and the C.28 support-result ref when one is consumed. -
C19-13 A pool-policy record for live loop-engineering candidates—for example, loops, agent harnesses, workflows, or DPF seeds—names the pool, governing lens, current treatment, and change trigger. Fresh generation, archive work, or front recomputation uses
C.18as the pool-policy pass specifies. Any other next-result question uses the transfer inC.19:4.4; C.19 does not absorb improvement, declaration or publication, choice, Work, or refresh. -
C19-15 Graduation, scaling, or widening an already supported use MUST cite its direct
graduationConditionRef. If that judgement relies on assurance,assuranceResultRef?cites the exact B.3 result andchangeTriggernames the satisfied condition and bounded supported scope. Widening exploration, continuing a line, retaining it without new probing, or sunsetting it MUST follow the stated continuation basis, including its contribution, feasible commitments and opportunity cost. Failure to graduate alone supplies neither a retirement decision nor a reason for indefinite continuation. A policy threshold or label does not create an assurance result.
C.19:8 - Common Anti-Patterns and How to Avoid Them
-
Using deployment readiness to settle research continuation. Judge active exploration and cheaper retention by their prospective contribution and remaining cost. Keep the qualification for deployment or transfer; past expenditure is not a reason to keep funding the line.
-
Treating one scalarized top-1 as the frontier. Avoid by naming the governing lens and keeping the live frontier distinct from any lens-ranked pick.
-
Treating an unresolved policy question as a settled treatment. Return the supported
currentTreatmentwhen its conditions hold. Otherwise identify the unresolved premise and dependent judgement; if another question has become current, name its subject pattern. -
Letting
SurpriseorIlluminationquietly become dominance criteria. Avoid by promoting them only through one declared lens or policy id and recording that promotion in provenance. -
Absorbing neighboring questions. Avoid by using the exact handoff values in
C.19:4.4instead of adding a neighboring result’s fields or claims toPoolPolicyResult.
C.19:9 - Consequences
- the result states whether the pool is being widened, kept live, narrowed, or sunset; if the question leaves pool policy, the record names the next subject pattern separately
- active exploration, cheap retention and exploitation can receive different justified decisions; heterogeneity need not collapse into one scalar winner
- the cost is stricter provenance and the need to name lenses, policies, and change triggers explicitly
C.19:10 - Rationale
Pool governance decides how to treat several still-live lines over the relevant horizon. A local choice among fixed options can select an action without deciding which other lines to continue or retain; a pool-policy result makes those commitments and their change conditions explicit.
When an applicable policy independently admits coverage or heterogeneity pressure, keep that pressure explicit until one declared reason justifies retirement or use of a different subject pattern. Sunset a line only when the current pool-policy result states the policy reason for retiring that line; name the next subject pattern only when the next question no longer concerns pool policy under C.19.
C.19:11 - SoTA-Echoing
The cited editions inform different policy pressures; none supplies a default algorithm, quota or sampler. If a changed source premise affects a pool-policy claim, reopen that dependent use through G.11 while retaining unaffected results.
| Source or source family | Adopted FPF move | Rejected overread | Practitioner implication |
|---|---|---|---|
Russo et al., A Tutorial on Thompson Sampling, arXiv:1707.02038v3 (2020-07-14). | Treat explore/exploit balancing as an explicit sequential policy pressure. | Thompson sampling or any bandit algorithm becomes the default C.19 method. | A pool-policy result names the policy or lens and the change trigger; local option choice still requires C.11. |
Frazier, A Tutorial on Bayesian Optimization, arXiv:1807.02811v1 (2018-07-08), and Yu et al., Efficient and Principled Scientific Discovery through Bayesian Optimization: A Tutorial, arXiv:2604.01328v3 (2026-04-07). | Keep acquisition, cost, uncertainty, and experiment-selection pressure visible when pool policy is used for expensive probing. | Bayesian optimization vocabulary locally redefines FPF choice, work, or evidence kinds. | Use BO-style source pressure to require policy ids, evidence/cost boundaries, and stop or change triggers, while comparison and enactment stay with their subject patterns. |
Qin et al., A survey on Quality-Diversity optimization: Approaches, applications, and challenges, Swarm and Evolutionary Computation 100:102240 (2026), DOI 10.1016/j.swevo.2025.102240, https://www.sciencedirect.com/science/article/pii/S2210650225003979. | Preserve live front, archive, coverage, and diversity pressure without collapsing them to one scalarized winner. | QD taxonomy redefines the archive and front relations stated through C.18, redefines the G.5 selected-set result, or authorizes a default family quota, DPP sampler, or max-min rule. | Keep keep_frontier, narrow_to_subset, and sunset_line distinct; use C.18 for archive and front meaning, G.5 for result declaration, and a profile admitted by the applicable policy for any heterogeneity rule. |
C.19:12 - Relations
C.27 temporal-claim relation.
- C.27 may flag: a temporal claim that changes exploration, exploitation, narrowing, widening, convergence speed, or search cadence in a way that changes admissible use.
- This pattern keeps: pool-policy result and explore and exploit governance, including
keep_frontier,narrow_to_subset, andsunset_line. - Non-admissible use: faster narrowing is not automatically a positive result; it may collapse exploration health, diversity, archive coverage, or frontier discovery.
- Exit: use C.19 for the pool-policy result; use C.27 only for the temporal-claim adequacy question when speed or change affects admissible use.
Builds on: C.18, C.16, A.19.CPM, A.19.SelectorMechanism, and B.3.
Coordinates with:
E.10.LRNonly for unresolved learning-family wording;A.10for a replayable account of the needed evidence or source basis;C.22.PFRfor actual Problem identity;C.18for generation, Archive, Front, and possibility-space change;C.32.P2S,C.32, andC.35for architecture-alternative carry-through and candidate admission;C.28for causal-use support;C.17andG.9for evaluation and parity inputs;C.11.CRConly for a missing finite configuration-relative comparison;C.11for local option or probe choice;- the other next-result patterns and transfer values named in
C.19:4.4.