| name | reduced-space-test |
| description | Scoped empirical validation utility. Decomposes a target↔surrogate equivalence claim into verifiable facets, bounds a user-synchronized test space, captures evidence inside it, and carries the uncovered complement forward. Use when an inference-uncertain proposition (does it behave / perform / transfer / hold value) needs evidence in a smaller stand-in space, and you want the claim scoped to the conditions actually tested rather than asserted absolutely. |
Reduced-Space Test: Scoped Empirical Validation
Validate an inference-uncertain proposition inside a constraint-bounded stand-in space synchronized with the user, obtain a scoped resolution ("within these conditions, whether it holds, fails, or remains inconclusive"), and carry the uncovered complement forward to a follow-up protocol. This skill does not run the experiment substrate, open branches, or create PRs. It orchestrates existing protocols around one disciplined empirical move.
This is an orchestration utility, not a runtime executor and not a new epistemic protocol. Reduced-Space Test introduces no new interaction deficit. It realizes a known composite — decompose the target↔surrogate equivalence claim into verifiable facets, then /bound a synchronized test space (+ residual) ∘ /inquire for evidence inside it → scoped resolution + carried complement. It is sibling to /triage and /forge: a thin composition over existing protocols, kept out of the protocol graph because it owns no deficit of its own.
The core recognition act is decomposing the equivalence claim into verifiable facets — not "creating a reduced space." A stand-in space is only as good as the facets on which it is claimed equivalent to the target; the value lives in making those facets explicit and observable.
Core Contract
/reduced-space-test owns scoped empirical validation:
InferenceUncertainClaim
-> ScopedClaimFrame (core: decompose target↔surrogate equivalence into verifiable facets)
-> BoundedTestSpace (/bound: user-synchronized in-scope space + residual complement)
-> EmpiricalEvidence (/inquire: observe inside the bounded space, evidence over inference)
-> ScopedResolution | CoverageShortfall (scoped outcome — holds, fails, or inconclusive — within the claim's defined conditions; or, on under-coverage, a CoverageShortfall: slice-scoped resolution or no resolution, with the remainder re-bounded or carried)
-> Residual (uncovered complement -> follow-up protocol)
The orchestration produces two first-class outputs: a scoped resolution (the proposition's outcome — holds, fails, or inconclusive — within the tested conditions) and a carried residual (the complement the test did not cover). Neither stands without the other — a scoped resolution that hides its residual overclaims.
Types
| Type | Meaning |
|---|
InferenceUncertainClaim | A proposition about behavior, performance, transfer, or value that inference alone cannot settle and that the user wants grounded in evidence. |
EquivalenceClaim | The asserted target-environment ↔ surrogate-space equivalence the test rests on. The test is only valid on the facets where this equivalence is itself examined. |
VerifiableFacet | One decomposed, observable dimension of the equivalence claim — a place where target and surrogate can be compared and a gap measured. |
ScopedClaimFrame | The Phase 1 output: the decomposed set of VerifiableFacets the test will and will not speak to — critical facets, the surrogate↔target difference inventory, influence-path hypotheses, and the chosen gap-measurement approach. Surfaced for user recognition before it constrains the boundary; it is what keeps the later evidence sentence honest. |
BoundedTestSpace | The in-scope validation space defined with the user through /bound, paired with its residual complement. The space's definition is what constitutes the verifiable claim. |
Residual | The complement the bounded space does not cover. A first-class output, carried forward, never dropped. |
EmpiricalEvidence | Observation captured inside the bounded test space through /inquire — evidence with cited basis, scoped to the conditions actually exercised. |
ScopedResolution | The scoped outcome within the bounded space on the tested facets: the proposition holds, fails, or remains inconclusive — stated as an updated failure probability within the defined conditions (lower on confirmation, higher on disconfirmation), never a bare "it works". A disconfirming or inconclusive result is a first-class resolution, not a loop failure. |
CoverageShortfall | The typed outcome when the bounded space turns out narrower than the claim it must license (under-coverage): no ScopedResolution is issued for the full claim; the honest exit re-scopes the resolution to the slice actually covered (resolution over the slice + the uncovered remainder carried as Residual) or re-bounds the space (→ Phase 2). A bounded exit, never an implicit retry. |
Composition
Reduced-Space Test orchestrates existing protocols; most per-step work is delegated to them, while this skill owns the Phase 1 facet decomposition outright, plus the sequencing and the scoping discipline across the protocols it composes.
(conditional front) /elicit | /induce Phase 0
-> decompose equivalence claim into facets [owned: ScopedClaimFrame] Phase 1
-> /bound [BoundaryUndefined -> DefinedBoundary, + residual] Phase 2
-> /inquire [ContextInsufficient -> InformedExecution, Observe] Phase 3
-> residual carry-forward (/inquire | /elicit) Phase 4
/bound natively emits a boundary plus its residual, so the complement is produced by the bound step itself rather than bolted on afterward. /inquire natively captures observation evidence under a scope-covers-claim discipline, so the empirical step reuses that grounding rather than re-deriving it.
Phase 0: Intake + Name the Pattern (conditional)
Read the InferenceUncertainClaim: the proposition the user cannot settle by reasoning alone, and why evidence in a stand-in space is wanted.
Front with a crystallization step only when the test intent is genuinely under-formed:
- If the test intent is aporetic — the user senses an unknown but cannot yet state the proposition — front with
/elicit (Euporia) to surface it.
- If the test pattern recurs but is unnamed — the same stand-in-validation shape appears across cases without a handle — front with
/induce (Periagoge) to crystallize it.
When the proposition is already stated and the pattern is familiar, proceed directly to Phase 1. The front step is a conditional affordance, not a mandatory gate.
Phase 1: Decompose the Equivalence Claim into Facets
This is the core act. The test rests on an EquivalenceClaim — that the surrogate space stands in for the target on the dimensions that matter. Make that claim observable:
- Surface critical facets: the dimensions on which the proposition's truth in the target depends (behavior, load, distribution, configuration, value).
- Inventory surrogate↔target differences: where the stand-in space diverges from the target, named rather than assumed away.
- Hypothesize influence paths: how each difference could change the outcome, so the test targets the differences that matter.
- Choose a gap-measurement approach: the means by which the surrogate-to-target gap is observed (for example randomization across conditions, shadowing live input, or a staged canary), selected for the facets in play.
The output is a ScopedClaimFrame: the facets that the test will and will not speak to. This frame is what keeps the later evidence sentence honest.
Surface the frame for recognition before it constrains the boundary. The facet set is an AI-formed hypothesis, not a settled determination — present it to the user as a structured, recognizable set (the critical facets, the surrogate↔target differences, and what is deliberately out of frame), with an explicit Emergent probe: is any decision-relevant facet missing, and are these the dimensions that actually matter? The user confirms, extends, or reweights the frame. Because the facet decomposition is the act that shapes where the test's attention goes — and that frame licenses every later claim — letting the AI fix it silently would inject attention bias at the root; surfacing it for recognition keeps the choice of which facets count with the user.
Phase 2: Bound the Test Space (compose /bound)
Compose /bound (Horismos: BoundaryUndefined -> DefinedBoundary) to define the BoundedTestSpace with the user and to emit its Residual complement in the same move.
Constraint sync is a Constitution interaction. The user defines the reduced space, and that definition constitutes the verifiable claim — testing "does the API respond" and testing "does the API sustain its rated throughput" are different reduced spaces that license different claims. The boundary the user draws is therefore not a mechanical narrowing; it determines what the eventual resolution is allowed to assert. Surface the facet frame from Phase 1 so the user draws the boundary in recognition of what each cut includes and excludes.
/bound returns the in-scope space and the residual together; carry both forward.
Phase 3: Capture Empirical Evidence (compose /inquire)
Compose /inquire (Aitesis: ContextInsufficient -> InformedExecution) to observe inside the BoundedTestSpace and settle the scoped uncertainty (confirm, disconfirm, or find it inconclusive).
- Capture observation evidence in the bounded space — evidence over inference, with cited basis.
- Hold the scope-covers-claim discipline: the space actually exercised must cover the claim the evidence will license. Evidence drawn from a narrower slice than the claim it is asked to support is under-covering and does not yield a
ScopedResolution.
The output is EmpiricalEvidence plus a ScopedResolution recording the outcome — confirming, disconfirming, or inconclusive — for the bounded space. Under-coverage (evidence drawn from a slice narrower than the claim) is the one case that yields no ScopedResolution for the full claim — its typed exit is a CoverageShortfall: re-scope the resolution to the slice actually covered (resolution over the slice + the uncovered remainder carried as Residual) or re-bound the space (→ Phase 2). A disconfirming result, by contrast, is a resolution, not its absence.
Phase 4: Scope the Claim + Carry the Residual Forward
Restrict the resolution to the conditions actually tested, and route the complement onward.
- Scope the claim: state the resolution as an updated failure probability within the defined conditions — reduced where the evidence confirmed, raised where it disconfirmed, left open where it was inconclusive. An absolute assertion in either direction ("it behaves identically" / "it is broken") exceeds what any stand-in space can support and is set aside in favor of the scoped form.
- Carry the residual forward: route the
Residual to a follow-up protocol — /inquire to gather more before a next decision, or /elicit to crystallize what the uncovered region still leaves open. The carry-forward is explicit output, recorded so a later session re-enters the uncovered region without re-deriving it.
At convergence, present the pairing: the scoped resolution with its defined conditions, and the carried residual with its routing. When the run ends in a CoverageShortfall, the pairing is the slice-scoped resolution (or an explicit no-resolution where no slice was covered) together with the uncovered claim-remainder as residual — the under-coverage path converges through the same pairing, never as a silent retry. The two together are the evidence that the test was honest about its reach.
Rules
- Facet decomposition is the core act (orchestration identity): The test's value is the explicit decomposition of the target↔surrogate equivalence into verifiable facets, not the construction of a stand-in space. A reduced space presented without its facet frame is an untested equivalence assumption wearing the appearance of evidence. The decomposed frame is surfaced for user recognition — with an Emergent probe for missing facets — before it constrains the boundary, so which facets count is never a silent AI determination.
- Constraint sync is Constitution (authority boundary): The user defines the bounded space, and that definition constitutes the verifiable claim. The skill surfaces the facet frame and the boundary's includes/excludes; the user draws the cut.
- Scoped claim only (evidence discipline): Resolution sentences assert an updated failure probability within the defined conditions — lower on confirmation, higher on disconfirmation, open on inconclusive. Absolute claims in either direction exceed what a stand-in space supports and are reframed to the scoped form. A disconfirming or inconclusive result is a first-class resolution, never suppressed to force a confirmation.
- Residual carries forward (completeness): The uncovered complement is a first-class output routed to a follow-up protocol. A scoped resolution that omits its residual overclaims its reach.
- Scope covers claim (coverage): The bounded test space must cover the claim it licenses. Evidence from a slice narrower than its claim is under-covering: it yields no scoped resolution for the full claim and exits as a
CoverageShortfall — re-scope to the covered slice (resolution over the slice + remainder as residual) or re-bound — never an implicit retry.
- Compose, do not reinvent (cost discipline): Reduced-Space Test orchestrates
/bound and /inquire (with a conditional /elicit or /induce front). It introduces no new interaction deficit and adds no protocol-graph node.
Boundary Note
/reduced-space-test sequences existing protocols and holds the scoping discipline across them. It reads the user's proposition and emits a scoped resolution plus a carried residual; it does not run the experiment substrate, provision environments, or enforce gate passage that belongs to a harness. Domain-specific facet templates (for data, software, or infrastructure stand-in spaces) are a deferred extension — the present contract is domain-free and decomposes facets per claim rather than from a fixed template.
Anti-patterns
- Absolute-claim leak: asserting "it behaves identically" or "it is the same" from stand-in evidence. The claim must be scoped to the tested conditions; absolutes are not defensible from a reduced space.
- Surrogate without difference inventory: bounding a test space without naming where it diverges from the target. The unnamed differences are exactly where the scoped resolution silently fails.
- Residual drop: presenting the scoped resolution while discarding the complement. The uncovered region is the part most likely to surprise a later session.
- Space-creation framing: treating "build the reduced space" as the work and skipping the facet decomposition. The equivalence claim, left implicit, becomes an assumption rather than a tested facet.
- New-protocol creep: growing the utility into a standalone protocol with its own deficit and graph node. The composite is sufficient; a new mechanism would duplicate
/bound and /inquire.
- Constraint sync by relay: drawing the reduced-space boundary on the user's behalf. The boundary constitutes the claim, so it is the user's to set.
- Silent facet selection: forming the facet frame and bounding the space without surfacing the frame for user recognition. Which facets count is the attention-shaping act; chosen silently, it injects the AI's bias into the very frame that licenses the claim.
- Confirmation-only terminal: treating the test as complete only when it confirms, with no honest terminal for a disconfirming or inconclusive result. A validation that cannot fail validates nothing; the disconfirming outcome is first-class.
Operational checklist (per cycle)
