inquire

name	inquire
description	Infer context insufficiency before execution. Surfaces uncertainties through information-gain prioritized inquiry when AI infers areas of context insufficiency, producing informed execution. Type: (ContextInsufficient, AI, INQUIRE, Prospect) → InformedExecution. Alias: Aitesis(αἴτησις).

Aitesis Protocol

Infer context insufficiency before execution through AI-guided inquiry. Type: (ContextInsufficient, AI, INQUIRE, Prospect) → InformedExecution.

Definition

Aitesis (αἴτησις): A dialogical act of proactively inferring context sufficiency before execution, where AI identifies uncertainties across multiple dimensions (factual, coherence, relevance), collects contextual evidence via codebase exploration, classifies each uncertainty by dimension and verifiability, resolves memory-internal contradictions through evidence, routes cross-domain concerns to other epistemic protocols, and inquires about remaining uncertainties through information-gain prioritized mini-choices for user resolution.

── FLOW ──
Aitesis(X) → Scan(X, dimensions) → Uᵢ →
  [if Uᵢ = ∅] sufficiency_relay(reasoning) → proceed (trivial InformedExecution)
  Ctx(Uᵢ) → (Uᵢ', Uᵣ) →
  classify(Uᵢ', dimension) → [if off-diagonal] Qc → (Uᵣ'_candidates, Uₑ_candidates, Uᵢ'', Uₙ) →
  ReadOnlyVerify(Uᵣ'_candidates) → (Uᵣ' resolved | admissibility-fail → reclassify EmpiricallyObservable) →
  [if Uₑ_candidates ≠ ∅] EmpiricalObservation(Uₑ_candidates) → Uₑ →
  Q(classify_result + Uₑ + Uᵢ'', priority) → A → X' → (loop until informed)
-- Uᵣ' (Step 3 survivors = read_only_resolved): skip Phase 2; ReadOnlyVerify failures rejoin Uₑ_candidates via backward arc (detail in PHASE TRANSITIONS)
-- Uₙ (non-actionable: CrossDomain coherence + detect-only dimensions): shown in classify summary with routing target
-- Uᵢ'' (factual/user-dependent or coherence/MemoryInternal/user-dependent): Phase 2 question candidates

── MORPHISM ──
Prospect
  → scan(prospect, context, dimensions)  -- infer context insufficiency (multi-dimension)
  → collect(uncertainties, codebase)     -- enrich via evidence collection
  → classify(enrichable, dimension)      -- epistemic classification (core act)
  → reclassify(MemoryInternal → Factual)  -- Coherence/MemoryInternal enters Factual resolution path
  → observe(empirically_observable, environment) -- dynamic evidence gathering (factual only)
  → surface(classify_result + observed + remaining, as_inquiry)
  → integrate(answer, prospect)
  → InformedExecution
requires: uncertain(sufficiency(X))      -- runtime checkpoint (Phase 0)
deficit:  ContextInsufficient            -- activation precondition (Layer 1/2)
preserves: task_identity(X)              -- task intent invariant; prospect context mutated (X → X')
invariant: Evidence over Inference over Detection

── TYPES ──
X        = Prospect for action (source-agnostic: task execution, analysis, investigation, or any purposeful action requiring context)
             -- Input type: morphism processes X uniformly; enumeration scopes the definition, not behavioral dispatch
Scan     = Context sufficiency scan: X → Set(Uncertainty)
Uncertainty = { domain: String, description: String, context: Set(Evidence) }
Evidence = { source: String, content: String }                -- collected during Ctx
Priority ∈ {Critical, Significant, Marginal}
Uᵢ       = Identified uncertainties from Scan(X)
Ctx      = Context collection: Uᵢ → (Uᵢ', Uᵣ)
Uᵢ'      = Enriched uncertainties (evidence added, not resolved)
Uᵣ       = Context-resolved uncertainties (resolved during collection)
Q        = Inquiry (Constitution interaction), ordered by information gain
A        = User answer ∈ {Provide(context), Point(location), Dismiss, Unknown(Partial)}
             -- Unknown(Partial) = user declines certainty; Phase 3 auto-promotes via Cite-or-observe tiebreaker (UserTacit → next-preferred
             -- EvidenceSource in ValidSources(v)) and re-enters Phase 1 for reclassification; routing arc formalized in PHASE TRANSITIONS
Ac         = User coherence classification ∈ CoherenceType     -- Phase 1 Qc gate answer type
X'       = Updated prospect (context-enriched)
InformedExecution = X' where remaining = ∅
EarlyExit = X' where user_esc  -- non-convergent early exit: X' as of exit (original prospect + any partial integration), partial transformation trace over resolved uncertainties, remaining declared as unresolved residual (morphism not completed)
-- Layer 1 (epistemic)
Dimension    ∈ {Factual, Coherence, Relevance} ∪ Emergent(Dimension)
               -- open set; external human communication excluded
Observability ∈ {StaticObservation, DynamicObservation, BeliefVerification}
               -- exists(fact, env) sub-modes
-- Layer 2 (tool implementation, Factual and Coherence/MemoryInternal fibers — fibration structure)
Verifiability  ∈ {ReadOnlyVerifiable, EmpiricallyObservable, UserDependent}
EvidenceSource ∈ {UserTacit, Instrumentation, CodeDerivable, CanonicalExternal}
                 ∪ Emergent(EvidenceSource)
               -- open set symmetric with Dimension; Emergent accumulator for novel channels
               -- (TestSuite, AsyncComms, HypomnesisIndex, RuntimeObservability, etc.)
               -- Emergent base promotion under variation-stable observed use
Claim(u)       = { referent: String, scope: String, expected_source_kind: EvidenceSource }
               -- expected_source_kind names the evidence CHANNEL a claim requires (which source-kind), a distinct axis from a claim's semantic category
EvidenceRef(e) = { source: String, source_kind: EvidenceSource, referent: String, scope: String, observed_at: String, content: String }
               -- interpretive extraction over Evidence (base type {source, content}): source_kind/referent/scope are inferred from content, NOT deterministic normalization
               -- this extraction exercises epistemic authority — a mis-extraction surfaces at the Phase 2 classify summary (support_integrity:unverified), not treated as deterministic relay
provenance_coupled(u, e) =
  referent(EvidenceRef(e)) = referent(Claim(u))
  ∧ authorizes(source_kind(EvidenceRef(e)), expected_source_kind(Claim(u)))   -- grantor = evidence's source_kind; claim side = expected_source_kind
  ∧ scope_subsumes(scope(EvidenceRef(e)), scope(Claim(u)))
authorizes : EvidenceSource × EvidenceSource → Bool   -- self-contained (no shared cross-protocol relation; Anamnesis defines its own namespace witness independently)
  authorizes(s, expected) ≡ s = expected
               -- reflexive base: a source-kind authorizes a claim expecting that same kind.
               -- cross-kind authorization (one kind standing in for another) is NOT granted here → defaults to non-authorizing; a richer policy
               --   matrix is a documented future extension, deliberately deferred to stay minimal-structural (deferral, not a silent gap).
scope_subsumes : String × String → Bool   -- path/tag-prefix subsumption: broader scope contains narrower (reused by coverage below)
ValidSources : Verifiability → ℘(EvidenceSource)
  ValidSources(ReadOnlyVerifiable)    = {CodeDerivable, UserTacit, CanonicalExternal} ∪ Emergent(EvidenceSource)
  ValidSources(EmpiricallyObservable) = {Instrumentation, UserTacit}                  ∪ Emergent(EvidenceSource)
  ValidSources(UserDependent)         = {UserTacit}                                    ∪ Emergent(EvidenceSource)
  -- Emergent(EvidenceSource) fallback-admissible when no base element fits the observed channel
  -- cost-ordering tiebreaker (ascending): CodeDerivable < CanonicalExternal < Instrumentation < UserTacit
  -- default selects lowest-cost valid source; override requires cite per Cite-or-observe rule
CoherenceType ∈ {MemoryInternal, CrossDomain}
               -- 2D: Scope(Same/Cross) × Resolution(Evidence/Structure); off-diagonal → Gate
Scope      ∈ {Same, Cross}
Resolution ∈ {Evidence, Structure}
off_diagonal(s, r) = ¬((s = Same ∧ r = Evidence) ∨ (s = Cross ∧ r = Structure))
classify   = Uᵢ' → Σ(d: Dimension). Fiber(d)
             where Fiber(Factual)       = Σ(v: Verifiability). {s: EvidenceSource | s ∈ ValidSources(v)}
                   Fiber(Coherence)     = CoherenceType
                   Fiber(Relevance)     = Unit    -- detect only
                   Fiber(Emergent(_))   = Unit    -- detect only (default; refinable per discovered dimension)
             -- 2-layer model = Grothendieck fibration: Layer 2 exists over Factual fiber;
             --   Factual fiber is itself a dependent sum — pair (v, s) where s is a SINGLE chosen element of ValidSources(v)
             --   (subset type {s | s ∈ ValidSources(v)}, not the power-set element itself)
             --   EvidenceSource choice within ValidSources(v) routes resolution channel
             -- Coherence fiber classifies into CoherenceType, where MemoryInternal instances enter the Factual resolution path
             --   (and inherit EvidenceSource via Factual reclassification)
             -- CrossDomain/Relevance/Emergent → detect + show routing target in classify summary (no EvidenceSource tag)
             -- ReadOnlyVerifiable direct-resolve admissibility = coverage ∧ support_integrity (both required):
             --   coverage         : scope_subsumes(scope(evidence), scope(claim))   -- rebutting axis: is the whole claim covered?
             --   support_integrity: ∃ e ∈ context(u): provenance_coupled(u,e)
             --                      ∧ link(evidence → asserted behavior/current reality) is verified, not silently desynced
             --                      -- undercutting axis: does this source-kind/referent/scope authorize this claim?
             --   currency ⊂ support_integrity (temporal sub-case): freshness is necessary, not sufficient — a current-but-unenforced
             --     artifact (comment/doc asserting behavior with no enforcement channel) is current yet support-unlinked → fails support_integrity
             --   failure of either axis (¬coverage = coverage_gap; ¬support_integrity = support_integrity_unverified) → reclassify EmpiricallyObservable
             --   (rebutting/undercutting framing per Pollock: two kinds of defeater — not asserted exhaustive)
support_integrity(u) ≡ (∃ e ∈ context(u): provenance_coupled(u, e)) ∧ evidence_behavior_linked(u)    -- undercutting axis; formal predicate (was comment-only)
               -- context(u) = evidence accessor over base Uncertainty.context: Set(Evidence) (existing field, not new)
               -- evidence_behavior_linked(u): evidence→behavior link verified (breaks-on-change), not silently desynced; currency ⊂ this (temporal sub-case)
ReadOnlyAdmissible = { u : ReadOnlyVerifiable | coverage(u) ∧ support_integrity(u) }
                   -- refinement over ReadOnlyVerifiable (NOT a new Verifiability constructor): the subset of ReadOnlyVerifiable
                   --   items admissible for Step 3 direct resolution. coverage(u) ≡ ¬coverage_gap(u); support_integrity(u) ≡ ¬support_integrity_unverified(u)
                   --   where support_integrity(u) now requires ∃ e: provenance_coupled(u,e) ∧ link-verified — a refinement INTO support_integrity (not a third conjunct): a value's authority is bound to the source × claim pair it actually supports.
                   --   Failure of either predicate → reclassify EmpiricallyObservable (backward arc T4). Step₃ ReadOnlyVerify takes the ReadOnlyVerifiable-classified candidate set (Uᵣ'_candidates, incl. support_integrity-undetermined items) and enforces this predicate at resolution time; ReadOnlyAdmissible characterizes the resolution survivors (= Uᵣ'), NOT a Step-3 input pre-filter.
ObservationSpec = { setup: Action, execute: Action, observe: Predicate, cleanup: Action }
EmpiricalObservation = (Uᵢ', ObservationSpec) → Uₑ  -- dynamic evidence gathering
Uᵣ'_candidates = { u ∈ Uᵢ' : classify(u) = (Factual, (ReadOnlyVerifiable, s)) ∧ s ≠ UserTacit ∧ s ∉ Emergent(EvidenceSource) }  -- Step 2 output → Step 3 input
               -- includes support_integrity-undetermined items pending resolution-time enforcement; symmetric with Uₑ_candidates (transient set, NOT a MODE STATE partition bucket)
               -- Step 3 partitions this set: survivors → Uᵣ' (read_only_resolved); admissibility failures → backward arc → EmpiricallyObservable
Uᵣ'        = Read-only verified uncertainties    -- Step 3 survivors only (= ReadOnlyAdmissible) → read_only_resolved; resolved (no Phase 2); excludes items routed via UserTacit override per Cite-or-observe rule
Uₑ_candidates = { u ∈ Uᵢ' : classify(u) = (Factual, (EmpiricallyObservable, s)) ∧ s ≠ UserTacit ∧ s ∉ Emergent(EvidenceSource) }
              -- Phase 1 observation checkpoint; excludes Cite-or-observe cite-based UserTacit overrides (those route directly to Uᵢ'')
Uₑ         = Empirically observed uncertainties    -- evidence attached, proceeds to Phase 2
Uᵢ''       = Remaining user-dependent uncertainties
             -- Includes: (a) Factual/UserDependent items
             --           (b) Factual/EmpiricallyObservable with EvidenceSource = UserTacit (Cite-or-observe cited override)
             --           (c) Factual/ReadOnlyVerifiable with EvidenceSource = UserTacit (Cite-or-observe cited override)
             --           (d) reclassified Coherence/MemoryInternal landing in any of (a)-(c) above
             --           (e) any Factual(v) with s ∈ Emergent(EvidenceSource) (channel unvalidated by definition; awaits Phase 2 Qs_emergent_channel confirmation)
             -- Phase 2 question candidates
Uₙ         = Non-actionable detected uncertainties  -- Fiber(Coherence) = CrossDomain or Fiber(d) = Unit; shown in classify summary with routing target
Action     = Tool call sequence (Write, Bash)
EscapeCondition ∈ {EnvironmentMutation, BoundExceeded, RiskElevated}
                    -- maps to Cite-or-observe escape hatches; logged in observation_skips
branching_factor : Uncertainty → ℕ
  branching_factor(u) = |distinct_resolution_paths(u) ∪ distinct_side_effect_branches(u)|
    -- counts mutually-exclusive resolution postures (e.g., in-place fix / redesign / offload / defer)
    -- UNION side-effect branch count (e.g., N downstream mutations per posture); both flavors summed
    -- per-uncertainty count must be citable in Phase 2 classify summary when Divergence-bounding exception is invoked

── PHASE TRANSITIONS ──
Phase 0: X → Scan(X, dimensions) → Uᵢ?                        -- context sufficiency checkpoint (silent)
       [Uᵢ = ∅] sufficiency_relay(reasoning) → proceed          -- zero-signal: present the sufficiency finding as relay text; trivial InformedExecution (remaining = ∅), Aitesis not activated
Phase 1: Uᵢ → Step₁ Ctx(Uᵢ) → (Uᵢ', Uᵣ) →                    -- Step 1: context collection [Tool]
         Step₂ classify(Uᵢ', dimension) → (Uᵣ'_candidates, Uₑ_candidates, Uᵢ'', Uₙ) → -- Step 2: epistemic classification (core act); Uₙ = non-actionable
         [if off-diagonal(scope, resolution)] Qc(scope_assessment, resolution_assessment) → Stop → Ac  -- Coherence 2D Constitution interaction [Tool]
         -- evaluation order: Qc resolves before Uₑ_candidates computation; reclassified MemoryInternal/EmpiricallyObservable enters Uₑ_candidates
         Step₃ ReadOnlyVerify(Uᵣ'_candidates) →     -- Step 3: read-only verification (CodeDerivable + CanonicalExternal); enforces admissibility (coverage ∧ support_integrity) at resolution time over the candidate set (incl. support_integrity-undetermined items) — survivors = ReadOnlyAdmissible = Uᵣ' (resolve directly, read_only_resolved); failures take the backward arc below [Tool]
           [if support_integrity_unverified(u) ∨ coverage_gap(u)] reclassify(u, EmpiricallyObservable) → goto Step₂  -- backward arc (T4): support-integrity/coverage failure re-enters classification (staleness = temporal sub-case of support_integrity_unverified)
         [if Uₑ_candidates ≠ ∅] Step₄ EmpiricalObservation(Uₑ_candidates) → Uₑ  -- Step 4: dynamic evidence gathering [Tool]
Phase 2: Qs(classify_result + Uₑ + Uᵢ''[cluster], framing) → Stop → A          -- uncertainty surfacing [Tool]; cluster = one coherent cluster (size ≤ 4)
Phase 3: A → integrate(A, X) → X'                               -- prospect update (track: mutates Λ.X)
         [if A = Unknown(Partial)] auto_promote(uncertainty, next_source(ValidSources(v))) → goto Phase 1  -- backward arc (T2): user declines certainty → re-enter classification with next-preferred EvidenceSource

── LOOP ──
After Phase 3: re-scan X' for remaining or newly emerged uncertainties.
New uncertainties accumulate into uncertainties (cumulative, never replace).
If Uᵢ' remains: return to Phase 1 (collect context for new uncertainties).
If remaining = ∅: proceed with execution.
User can exit at Phase 2 (early_exit).
Continue until: informed(X') OR user ESC.
Convergence evidence: At remaining = ∅, present transformation trace — for each u ∈ (Λ.context_resolved ∪ Λ.read_only_resolved ∪ Λ.empirically_observed ∪ Λ.user_responded), show (ContextInsufficient(u) → resolution(u)). Convergence is demonstrated, not asserted.
On user ESC (EarlyExit, not InformedExecution): present the same partial transformation trace restricted to uncertainties already resolved, then declare `remaining` as explicit unresolved residual.

── CONVERGENCE ──
actionable(Λ) = uncertainties \ non_factual_detected       -- Fiber(Factual) + Fiber(Coherence)=MemoryInternal uncertainties
informed(X') = remaining = ∅                                -- non_factual_detected does not block convergence
progress(Λ) = 1 if |actionable(Λ)| = 0 else 1 - |remaining| / |actionable(Λ)|   -- |actionable| = 0 (zero-signal or all-nonactionable trivial convergence) is fully converged, not undefined; denominator excludes non-actionable (CrossDomain + detect-only dimensions)
narrowing(Q, A) = |remaining(after)| < |remaining(before)| ∨ context(remaining(after)) ⊃ context(remaining(before))
early_exit = user_declares_sufficient

── TOOL GROUNDING ──
-- Realization: Constitution → TextPresent+Stop; Extension → TextPresent+Proceed
Phase 0 Scan    (sense)       → Internal analysis (no external tool)
Phase 0 sufficiency_relay (extension) → TextPresent+Proceed (Uᵢ = ∅: present the sufficiency finding with reasoning; proceed with X unchanged, trivial InformedExecution)
Phase 1 Ctx     (observe)     → Read, Grep (stored knowledge extraction: codebase, memory, references); WebSearch, WebFetch (conditional: CanonicalExternal channel — RFCs, vendor API docs, standards; `source: "web:{url}"` tag + staleness guard via codebase version cross-check); Bash (conditional: VersionControlHistory channel — read-only commit-log queries via subprocess (content pickaxe, message search, temporal range); `source: "history:{ref}"` tag; collection-only — ref-type staleness classification handled per Phase 1 Step 1 staleness rule)
Phase 1 Classify (observe)    → Internal analysis (multi-dimension assessment); Read, Grep (stored knowledge cross-reference analysis)
Phase 1 Qc      (constitution)        → present (conditional: Coherence 2D off-diagonal Constitution interaction; fires only when scope ≠ resolution assessment; user classifies coherence type as MemoryInternal or CrossDomain)
Phase 2 Qs_emergent_channel (constitution) → present (specialization of Phase 2 Qs: channel unvalidated by definition; regardless of parent Verifiability, the classify summary records the observed channel description and awaits user confirmation before proceeding; confirmation rides the parent A coproduct — Point(location) designates/validates the authoritative channel, Provide(context) supersedes it, Dismiss declines it (proceed-with-assumption), Unknown(Partial) leaves the item unresolved — no answer auto-resolves the item through the unconfirmed channel; the answer is recorded in Λ.channel_validations — a channel already Point-validated this session skips this gate only (prior in-session user decision); each later item on that channel still takes the claim-specific Phase 1 evidence pass against the validated channel, per the Point(location) semantics (record location, resolve via next Phase 1 iteration) — never blanket-resolved as user-responded)
Phase 2 Qs_staleness (constitution) → present (specialization of Phase 2 Qs: when staleness cannot be verified; require BOTH `staleness:unverified` tag — the temporal sub-case of the general `support_integrity:unverified` tag — AND classify summary surfacing — no silent escalation path; publishing authority claim warrants user awareness)
Phase 1 Observe (transform)   → Write, Bash, Read (dynamic evidence gathering, Factual only); cleanup via Bash
Phase 2 Qs      (constitution)        → present (mandatory: classify result + uncertainty surfacing; user provides context judgment on insufficiency; Esc key → loop termination at LOOP level, not an Answer)
Phase 3         (track)       → Internal state update
converge     (extension)       → TextPresent+Proceed (convergence evidence trace; proceed with informed execution)
esc          (extension)       → TextPresent+Proceed (partial transformation trace + unresolved residual declaration; terminate as EarlyExit, not InformedExecution)

── MODE STATE ──
Λ = { phase: Phase, X: Prospect, uncertainties: Set(Uncertainty),
      dimensions_detected: Set(Dimension),                           -- π₁ image of classify_results
      classify_results: Map(Uncertainty, Σ(d: Dimension). Fiber(d)), -- fibration-typed classification
      context_resolved: Set(Uncertainty),  -- Uᵣ from TYPES
      read_only_resolved: Set(Uncertainty), -- Uᵣ' from TYPES
      empirically_observed: Set(Uncertainty), -- Uₑ from TYPES
      non_factual_detected: Set(Uncertainty), -- Uₙ from TYPES; Fiber(Coherence) = CrossDomain or Fiber(d) = Unit, classify summary routing
      user_responded: Set(Uncertainty),
      remaining: Set(Uncertainty), dismissed: Set(Uncertainty),
      history: List<(Uncertainty, A)>, observation_history: List<(ObservationSpec, Result, Evidence)>,
      observation_skips: List<(Uncertainty, EscapeCondition, String)>,  -- audit trail for Cite-or-observe escape hatches
      source_choice_overrides: List<(Uncertainty, EvidenceSource, String)>,  -- audit trail for Cite-or-observe cite-based UserTacit overrides
      channel_validations: List<(Uncertainty, EvidenceSource, A)>,  -- Qs_emergent_channel answers recorded at Phase 3; audit trail feeds variation-stable observed use for (cross-session) base promotion; a channel Point-validated this session does not re-enter the gate this session
      active: Bool,
      cause_tag: String }
-- Invariant: uncertainties = context_resolved ∪ read_only_resolved ∪ empirically_observed ∪ non_factual_detected ∪ user_responded ∪ remaining ∪ dismissed (pairwise disjoint)
-- Note: observation_skips and source_choice_overrides are audit logs orthogonal to the partition —
--       observation_skips: logged when EmpiricallyObservable is reclassified to UserDependent via Cite-or-observe escape conditions
--       source_choice_overrides: logged when UserTacit is selected over cheaper EvidenceSource with cited dominance basis (Cite-or-observe dominance); audit trail supports variation-stable observed use for cost-ordering

── COMPOSITION ──
*: product — (D₁ × D₂) → (R₁ × R₂). registered dependency edges preserved. Dimension resolution emergent via session context.

Core Principle

Evidence over Inference over Detection: Aitesis operates on an epistemic hierarchy with two boundaries. The lower boundary (Inference > Detection): infer context insufficiency from requirements rather than detecting via fixed taxonomy — the protocol dynamically identifies what context is missing, not mechanically checking against a preset list. The upper boundary (Evidence > Inference): gather evidence through direct environmental observation rather than substituting inference from reasoning alone — when a fact is observable, observe it. Corollary: admissibility of found evidence on the direct-resolve path has two axes — coverage (does the evidence span the whole claim?) and support_integrity (does the evidence actually track the behavior it asserts, or is it silently desynced?). Support-integrity includes provenance coupling: the evidence's referent, source-kind, and scope must authorize the claim being resolved. Partial coverage is inference for the uncovered portion; current-but-unenforced evidence is inference dressed as evidence for the asserted behavior. Both must be verified before treating evidence as resolution (rebutting and undercutting defeaters, per Pollock — two kinds, not an exhaustive taxonomy).

Within this hierarchy, the AI first collects contextual evidence via codebase exploration to enrich question quality, then classifies each uncertainty by dimension and verifiability — classification is the protocol's core epistemic act, not a routing sub-step. For factual uncertainties, the AI resolves read-only verifiable facts directly and empirically observes dynamically accessible ones with direct evidence before asking. For coherence, the AI classifies by scope and resolution method — memory-internal contradictions are resolved through factual reclassification, while cross-domain contradictions are routed to downstream protocols via deficit-matched routing. For relevance, the AI detects and shows routing targets in the classify summary. The purpose is multi-dimensional context sufficiency sensing — asking better questions for what requires human judgment, self-resolving what can be observed, resolving memory-internal contradictions through evidence, and routing cross-domain concerns to other epistemic protocols.

Write is authorized for observation instrument setup (temporary test artifacts with mandatory cleanup). Cite-or-observe is the structural expression of the upper boundary — the adversarial guard against stopping at Inference when Evidence is achievable.

Mode Activation

Activation

Two activation layers: Layer 1 (User-invocable) — /inquire slash command or description-matching input, always available; Layer 2 (AI-guided) — context insufficiency inferred before execution via in-protocol heuristics. Inference is silent (Phase 0), except the zero-unknown sufficiency relay which presents its reasoning without yielding the turn; surfacing always requires user interaction via Cognitive Partnership Move (Constitution) (Phase 2).

Context insufficient = the prospect contains requirements not available in the current context and not trivially inferrable. Context insufficiency spans multiple dimensions: missing facts, incoherent facts, and facts not relevant to the prospect's goals. Sufficiency encompasses both executability (can the action proceed?) and analysis confidence (is the context adequate for reliable judgment?).

Gate predicate:

uncertain(sufficiency(X)) ≡ ∃ requirement(r, X) : ¬available(r, context) ∧ ¬trivially_inferrable(r)

Priority

When Aitesis is active:

Supersedes: Direct execution patterns in loaded instructions (Context must be verified before any execution begins)

Retained: Safety boundaries, tool restrictions, user explicit instructions

Action: At Phase 2, present highest information-gain uncertainty candidate with classify results via Cognitive Partnership Move (Constitution).

• Aitesis completes before execution proceeds
• Loaded instructions resume after context is resolved or dismissed

Trigger Signals

Heuristic signals for context insufficiency inference (not hard gates): Novel domain — knowledge area not previously addressed in session; Implicit requirements — task carries unstated assumptions; Ambiguous scope — multiple valid interpretations exist and AI cannot determine intended approach from available context; Environmental dependency — relies on external state (configs, APIs, versions); Prior-decision implication — prospect references or rests on a decision made in an earlier session (rationale, commitment, convention) not present in current conversation context.

Default-scan trigger (resolves the bootstrapping asymmetry of the Prior-decision implication signal): When the prospect scope touches architecture decisions, API or protocol design, persisted state schemas, or user-facing behavior commitments, the Prior-decision signal fires by default — Phase 1 Ctx begins with a bounded scan over persistent memory (MEMORY.md; project-local prior-decision logs where the project maintains them) regardless of explicit user mention. Without this default, detection would require the AI to already suspect prior involvement, which is the condition the signal is meant to surface. The default-scan is scope-triggered (not blanket), bounded to Phase 1 Ctx (no gate), and subject to the memory staleness rule in Phase 1 Step 1 (verify against current state before treating as resolved).

Cross-session enrichment: Prior session indices from the hypomnesis store (prior-session recall indices), when present, may seed the Phase 0 uncertainty scan; the constitutive judgment remains with the user.

Revision threshold: When accumulated observation_skips entries across 3+ sessions cluster around a specific EscapeCondition with consistent rationale, the Verifiability classification boundary warrants revision — the escape is systematic, not exceptional. When accumulated Emergent dimension detections across 3+ sessions reveal a recurring non-factual uncertainty pattern, the Layer 1 dimension set warrants a new fiber in the fibration structure — promoted fibers default to Unit (detect-only) unless the pattern exhibits internal classification structure requiring a structured fiber type.

Skip: execution context is fully specified in current message; user explicitly says "just do it" or "proceed"; same (domain, description) pair was dismissed in current session (session immunity); Phase 1 context collection resolves all identified uncertainties; read-only / exploratory task — no prospect to verify.

Mode Deactivation

All uncertainties resolved (context, read-only, observed, or user) → proceed with updated prospect. All remaining uncertainties dismissed → proceed with original prospect + defaults. User Esc key → EarlyExit (not InformedExecution): present partial transformation trace + declare remaining as unresolved residual, then return to normal operation.

Uncertainty Identification

Uncertainties are identified dynamically per task — no fixed taxonomy. Each uncertainty is characterized by domain (the knowledge area where context is missing — e.g., "deployment config", "API versioning", "user auth model"), description (what specifically is missing or uncertain), and context (evidence collected during Phase 1 that enriches question quality).

Priority

Priority reflects information gain — how much resolving this uncertainty would narrow the remaining uncertainty space.

Level	Criterion	Action
Critical	Resolution maximally narrows remaining uncertainty space	Must resolve before execution
Significant	Resolution narrows uncertainty but alternatives partially compensate	Surface to user for context
Marginal	Reasonable default exists; resolution provides incremental improvement	Surface with pre-selected Dismiss option

Priority is relational, not intrinsic: the same uncertainty may be Critical in one context and Marginal in another, depending on what other uncertainties exist and what context is already available.

When multiple uncertainties are identified, surface in priority order (Critical → Significant → Marginal). Only one uncertainty surfaced per Phase 2 cycle.

Protocol

Phase 0: Context Sufficiency Gate (Silent)

Analyze prospect requirements against available context across multiple dimensions. This phase is silent — no user interaction.

1. Scan prospect X for required context: domain knowledge, environmental state, configuration details, user preferences, constraints
2. Check availability: For each requirement, assess whether it is available in conversation, files, or environment
3. Dimension assessment: Identify which dimensions are potentially insufficient — factual (missing information), coherence (conflicting information), relevance (information not relevant to goal)
4. If all requirements satisfied: present the sufficiency finding with reasoning before proceeding, then Aitesis is not activated
5. If uncertainties identified: record Uᵢ with domain, description — proceed to Phase 1

Scan scope: Current prospect context, conversation history, observable environment. Does NOT modify files or call external services.

Phase 1: Context Collection + Classification + Empirical Observation

Collect contextual evidence, classify each uncertainty by dimension and verifiability, and empirically observe accessible uncertainties with direct evidence.

Step 1 — Context collection: For each uncertainty in Uᵢ:

• Call Read/Grep to search for relevant information in codebase, configs, documentation
• If definitive answer found: mark as context-resolved (Uᵣ), integrate into execution context
• If partial evidence found: enrich uncertainty with collected evidence (Uᵢ'), retain for classification
• If conflicting evidence found: enrich uncertainty with conflicting findings (Uᵢ'), retain for classification
• If no evidence found: retain in Uᵢ' with empty context
• Cross-session state consultation is in-scope: When a prospect implicates prior decisions, rationale, or committed state from earlier sessions, Read/Grep over persistent memory files (MEMORY.md; project-local prior-decision logs where present) and version-control history scans (commit-content pickaxe, message search, temporal range queries — when the substrate provides version control) are part of Phase 1 Ctx — not a separate protocol. Consulting past state at the current decision moment is context collection. Evidence from memory is tagged source: "memory:{path}"; evidence from version-control history is tagged source: "history:{ref}" (commit hash / tag / branch ref) for traceability. Support-integrity guard (general axis): evidence is eligible for ReadOnly direct resolution only when it is support-linked — its referent, source-kind, and scope authorize the exact claim being resolved, and its claim is coupled to (enforced by) actual behavior, not merely present. A current-but-unenforced codebase artifact (a comment or doc that asserts behavior with no enforcement channel) is fresh yet fails support_integrity, so it is routed to EmpiricallyObservable rather than auto-resolved: currency is necessary but not sufficient. Staleness is the temporal sub-case of this axis: memory- and version-control-sourced evidence is temporally decoupled from the current environment — the referenced decision may have been superseded, the referenced file may have been renamed, the convention may have been revised — which is one way the evidence→behavior link goes unverified. History ref currency by ref-type: full commit hashes (40-char SHA) are immutable content addresses — the commit's content snapshot is stable; mutable refs (branches, tags, partial-hash refs) may be amended/rebased/force-pushed and require currency verification. Even immutable commit-hash evidence may be interpretively superseded by later commits (the cited change may no longer reflect current convention). Such temporally-decoupled evidence is therefore NOT eligible for the Factual/ReadOnly direct-resolve path: classify memory- or history-sourced evidence as Factual/EmpiricallyObservable (verify against current codebase/environment state before treating as resolved) or escalate to user confirmation via the classify summary with the explicit staleness:unverified tag (the temporal sub-case of the general support_integrity:unverified tag). This is discovery of what the AI already committed to, not pattern discovery across sessions. Routing distinction: this channel covers AI-driven scope expansion at runtime (objective historical investigation triggered by Prior-decision implication signals); subjective recall of prior session content (user-utterance-bound, empty_intention trigger) routes through /recollect (Anamnesis) per the protocol's input-side authority shape, not through this channel.

Step 2 — Epistemic classification (core act): For each remaining uncertainty in Uᵢ':

• Dimension assessment (Layer 1): Is this factual, coherence, or relevance?
  • Factual: a fact is missing from context and required for execution
  • Coherence: collected facts are mutually inconsistent
  • Relevance: collected facts are not relevant to the execution goal
• Verifiability assessment (Layer 2, Factual dimension — Observability sub-modes guide classification):
  • ReadOnlyVerifiable: fact exists in environment (StaticObservation or BeliefVerification) and is observable with current tools, AND both admissibility axes pass — coverage (evidence scope ⊇ claim scope) AND support_integrity (the evidence's referent/source-kind/scope authorizes the claim, and the evidence→behavior link is verified, not silently desynced) → resolve directly via extended context lookup
    • support_integrity operational test: first normalize the claim and evidence through Claim(u) and EvidenceRef(e), then verify provenance_coupled(u,e) and ask: "If the behavior this evidence asserts changed right now, would the evidence break or fail — or would it silently drift (no enforcement mechanism surfaces the divergence)?" Breaks-on-change → support-linked (admissible); silently drifts → support_integrity fails. Currency (freshness) is the temporal sub-case: a current artifact still fails this test when nothing couples it to the behavior it claims.
    • When coverage fails (evidence scope ⊊ claim scope): split — covered portion proceeds to Step 3 (ReadOnly resolution), uncovered portion is classified separately and enters the appropriate verifiability path
    • When support_integrity fails determinably at Step 2 (the evidence is current/present but visibly not coupled to the asserted behavior): reclassify directly within this step as EmpiricallyObservable and observe the behavior rather than trusting the artifact — this codebase support-integrity failure self-resolves through observation, requiring no classify-summary escalation. If bounded, non-destructive observation is blocked (an escape condition), it routes onward to user inquiry per the standard EmpiricallyObservable path, escalating via the classify summary with the support_integrity:unverified tag (the general tag of which staleness:unverified is the temporal sub-case). When support_integrity cannot be determined from the collected evidence at Step 2, classify tentatively as ReadOnlyVerifiable; discovery of the failure at resolution time is handled by the Step 3 backward arc (PHASE TRANSITIONS), not at classification time
  • EmpiricallyObservable: fact requires DynamicObservation — does not exist statically but is observable through non-destructive execution, reversible, and bounded (< 30s) → empirical observation
  • UserDependent: neither read-only verifiable nor empirically observable → Phase 2 directly
• EvidenceSource selection (Layer 2, Factual dimension — Fiber(Factual) = Σ(v: Verifiability). {s | s ∈ ValidSources(v)}):
  • For each Factual(v) uncertainty, select a single EvidenceSource from ValidSources(v) (function body in TYPES).
  • Default: lowest-cost valid source per cost-ordering in TYPES (ascending: CodeDerivable < CanonicalExternal < Instrumentation < UserTacit).
  • External-dependency preference (cost-ordering adjustment): when the uncertainty carries an environmental-dependency signal (external API version, vendor behavior, RFC/standard semantics), prefer CanonicalExternal over CodeDerivable unless an authority-override basis is cited (e.g., internal fork supersedes upstream, vendor behavior already reverse-engineered in-repo). Log override to Λ.source_choice_overrides.
  • Override — Cite-or-observe rule (extended): choosing UserTacit over a cheaper source requires cited dominance basis and is logged to Λ.source_choice_overrides:
    • UserTacit over Instrumentation (EmpiricallyObservable): cite (a) operational context absent from instrumentation, (b) temporally-scoped knowledge not reproducible statically, or (c) setup > 30s bound
    • UserTacit over CodeDerivable/CanonicalExternal (ReadOnlyVerifiable): cite (a) tacit domain knowledge not captured in source/external doc, (b) interpretive judgment over canonical text, or (c) authority-override rationale
    • Without citation, default to the cheaper source
  • Routing consequence (determines Step 3/4/Phase 2 channel):
    • CodeDerivable → Step 3 (Read-only verification; codebase Read/Grep)
    • CanonicalExternal → Step 3 with WebFetch/WebSearch (published external docs; source: "web:{url}" tag + determinism verification + staleness guard — see Web context below)
    • Instrumentation → Step 4 (Empirical observation via Bash lifecycle)
    • UserTacit → Phase 2 directly (user-dependent inquiry; includes reclassified Coherence/MemoryInternal items)
    • Emergent(source) → always Phase 2 (Constitution per TOOL GROUNDING Phase 2 Qs_emergent_channel): record observed channel description in classify summary, await user confirmation that this channel is appropriate; accumulate toward variation-stable observed use for base promotion. Within the session, the validation is recorded in Λ.channel_validations; a channel already Point-validated this session skips the re-gate only (prior in-session user decision); each later item on that channel still routes through the claim-specific Phase 1 evidence pass against the validated channel, per the Point(location) semantics (record location, resolve via next Phase 1 iteration) — base promotion itself stays cross-session. Confirmation rides the standard Phase 2 answer coproduct — Point(location) designates/validates the authoritative channel, Provide(context) supersedes it, Dismiss declines it (proceed-with-assumption), Unknown(Partial) leaves the item unresolved — no separate answer type, and no answer auto-resolves the item through the unconfirmed channel. Parent Verifiability tier is NOT used to bypass Phase 2 — the channel is unvalidated by definition.
• Coherence classification (Layer 2, 2D model: Scope × Resolution):
  • Pre-filter: cross-scope + rule-resolvable (existing scope hierarchy, established precedence) → coexistence (exit Coherence; not a contradiction)
  • Same scope + evidence-resolvable → MemoryInternal → factual reclassification (ReadOnlyVerifiable / EmpiricallyObservable / UserDependent) → follows Factual resolution path (Step 3, Step 4, or Phase 2)
    • EvidenceSource inheritance procedure: reclassified MemoryInternal items enter EvidenceSource selection identically to directly-classified Factual(v) items — the same cost-ordering default, external-dependency preference, and Cite-or-observe rule override requirements apply; source_choice_overrides logging applies identically
  • Cross scope + structure-requiring → CrossDomain → deficit-matched routing: MappingUncertain→/ground, BoundaryUndefined→/bound, FrameworkAbsent→/frame, GapUnnoticed→/gap, AbstractAporia→/elicit, DirectionUnrecognizable→/preview
  • Off-diagonal (Scope ≠ Resolution): present both assessments with evidence via conditional gate; user classifies as MemoryInternal or CrossDomain
    • (Same, Structure): same-scope contradiction where factual verification is insufficient — resolution requires understanding structural relationships within the scope
    • (Cross, Evidence): cross-scope contradiction where evidence comparison can determine which scope's claim is current — despite scope difference, factual verification suffices
  • MemoryInternal → actionable (proceeds to resolution); CrossDomain → record as Uₙ (non_factual_detected) with deficit-matched routing target
• Other non-actionable dimensions: Relevance and Emergent → detect and record as Uₙ (non_factual_detected); shown with routing target in classify summary, not Phase 2 question
  • Relevance → deficit-matched: GapUnnoticed→/gap, BoundaryUndefined→/bound, AbstractAporia→/elicit
  • Emergent(_) → match observed deficit condition against candidate protocol deficit conditions
• Store all results in Λ.classify_results

Step 3 — Read-only verification: For the ReadOnlyVerifiable-classified candidate set (Uᵣ'_candidates, including items whose support_integrity was undetermined at Step 2). Step 3 enforces both admissibility axes during resolution — items confirmed to pass coverage ∧ support_integrity resolve directly (these survivors constitute ReadOnlyAdmissible, marked Uᵣ' / read_only_resolved); items discovered to fail either axis at resolution time are reclassified to EmpiricallyObservable via the backward arc and re-enter Step 2 (support_integrity failures already determinable at Step 2 were kicked there and never reach Step 3):

• Targeted context lookup via Read/Grep — classification narrows search scope to specific files/locations that Step 1's broad sweep did not cover (e.g., spec files, config schemas identified by classify)
• Static cross-checks over proxy evidence resolve by collecting the proxy's enforced source, normalizing it through EvidenceRef(e), then re-entering classify on that source; this is the existing collect + classify + Step 3 path, not a distinct observation channel
• Scope and provenance re-verification: if targeted lookup reveals evidence scope ⊊ claim scope not detected at Step 2 (subtle coverage gap), or a referent/source-kind mismatch between evidence and claim, apply the same split or backward-arc reclassification — covered coupled portion resolved, uncoupled portion reclassified separately
• Resolved: mark as Uᵣ' (read_only_resolved), skip Phase 2

Step 4 — Empirical observation: For EmpiricallyObservable uncertainties:

• Pre-observation escape check: If Cite-or-observe escape conditions apply BEFORE execution (environmental constraint prevents bounded, non-destructive observation), reclassify as UserDependent and log (uncertainty, condition, rationale) to Λ.observation_skips — the skip rationale must cite the triggering escape condition and the empirical basis that established it. Skip remaining Step 4 execution for this uncertainty.
• Execution (if no pre-observation escape): Construct ObservationSpec: { setup, execute, observe, cleanup }; execute lifecycle (setup → execute → observe → cleanup → record)
• Outcome — differentiating evidence found: Attach evidence to Uₑ → proceeds to Phase 2 with positive evidence
• Outcome — no differentiating signal: Observation executed but no state change distinguishes candidate answers. Attach the null-signal finding as negative evidence to Uₑ → proceeds to Phase 2 with observation record. This is NOT an escape or reclassification; the observation output itself is the evidence. Phase 2 surfaces the observed null-signal to the user for judgment, not AI-internal categorical dismissal.

If all uncertainties context-resolved or read-only-resolved (no observed or user-dependent remaining): proceed with execution (no user interruption). If observed or user-dependent uncertainties remain: proceed to Phase 2.

Web context (CanonicalExternal channel, conditional): When uncertainty carries an environmental dependency signal (external API versions, library maintenance status, breaking changes, RFC/standard semantics, vendor documentation) and the information is not available in the codebase, extend context collection to WebSearch/WebFetch. Web evidence is tagged with source: "web:{url}" for traceability.

Determinism verification (Extension precondition): CanonicalExternal is classified as Extension only when the source is deterministic for the claim scope. Verify one of: (a) pinned version or dated snapshot (specific RFC with publication date, vendor doc with version pin, W3C spec with date stamp), (b) tag-pinned URL (/v1.2/, ?version=X), or (c) cached copy with recorded fetch timestamp. When the source is undated or versionless and the claim depends on temporal context (API behavior, deprecated features, vendor defaults), the fetch is NON-deterministic → classify as Constitution and escalate via Phase 2 classify summary before treating as evidence.

Staleness guard (analogous to memory evidence): documentation may be temporally decoupled from the library version actually in use — verify against codebase import/pin version before treating as resolved. When staleness cannot be verified, the guard requires BOTH: (1) staleness:unverified tag (the temporal sub-case of the general support_integrity:unverified tag) attached to the evidence record, AND (2) surfacing to the user in Phase 2 classify summary regardless of whether EmpiricallyObservable reclassification subsequently resolves the uncertainty — no silent escalation path (no "or" fallback). The CanonicalExternal source carries publishing authority (standards body, vendor) distinct from internal CodeDerivable evidence — cite both when cross-validating, and always cite the authority source when its temporal alignment cannot be independently verified (see TOOL GROUNDING Phase 2 Qs_staleness Constitution entry).

Scope restriction:

• Context collection: Read-only investigation (Read, Grep, WebSearch). — core preserved
• Read-only verification: Extended context lookup for verifiable facts (Read, Grep). — resolves directly
• Empirical observation: Non-destructive observation via Bash execution, with optional Write for observation instrument setup (temp test artifacts). Observation results are evidence for Phase 2, not resolution — evidence gathering, not replacement. Design constraints: Minimal (create the smallest possible observation instrument), Reversible (all observation artifacts created in temp locations and cleaned up after observation), Sandboxed (observation must not modify existing project files), Transparent (log observation lifecycle in Λ.observation_history), Bounded (30-second timeout → fall back to user inquiry), Risk-aware (elevated-risk observation → reclassify as UserDependent).

Phase 2: Uncertainty Surfacing

Present the highest-priority remaining uncertainty with classify results via Cognitive Partnership Move (Constitution).

Classification transparency (Always show): Phase 2 always includes classify results for remaining uncertainties. This is informational — no approval required. Users can override classification by stating objection. When only one uncertainty remains, inline the classification with the uncertainty description rather than showing a separate summary block.

Surfacing format:

Present the classification results, uncertainty description, and evidence as text output:

• Classification summary (each remaining uncertainty shows how the AI plans to resolve it, with the basis cited; each line is one case the user may see):
  • a factual question whose answer can be looked up in the codebase, where the source is actually coupled to the behavior it describes (basis: evidence summary + a check that the source would break/fail if that behavior changed)
  • a factual question whose answer comes from a published external document (basis: doc url + codebase version cross-check)
  • a factual question that can be checked by running a test (basis: evidence summary)
  • a factual question that could be tested but is faster to ask the user (basis: dominance reason — operational context, temporal scope, or setup beyond a short time bound)
  • a factual question that would have been tested but the test is blocked by an environmental escape condition, falling back to asking the user
  • a factual question with partial evidence — the evidence covers one part of the claim, the uncovered part is classified separately
  • a factual question whose codebase evidence is present and current but not provenance-coupled to the claim or not coupled to the behavior it asserts — routed to a test/observation rather than trusted as-is (and escalated with a support_integrity:unverified note only if that observation is blocked)
  • a factual question via a newly observed evidence channel; your confirmation requested before treating this channel as resolved
  • a consistency question within the same scope, treated as a factual question (its evidence path is selected by the same procedure as a directly-classified factual item)
  • a consistency question spanning multiple scopes — routed to a downstream protocol
  • a relevance question — routed to the protocol matching its deficit (/gap, /bound, or /elicit)
  • Any of these classifications to revise?
• [Specific uncertainty description — highest priority]
• Evidence: [Evidence collected during context collection and observation, if any]

Then present:

How would you like to resolve this uncertainty?

Options:
1. **[Provide X]** — [what this context enables]
2. **[Point me to...]** — tell me where to find this information
3. **Dismiss** — proceed with [stated default/assumption]
4. **Unknown / Partial** — I don't know, or I only have partial context; the AI will try the next-preferred way of finding the answer and re-classify

Option 4 is the typed Unknown(Partial) constructor (TYPES A): Phase 3 auto-promotes to the next-preferred EvidenceSource in ValidSources(v) and re-enters Phase 1 classification via the backward arc (PHASE TRANSITIONS). This is a type-preserving materialization — not gate mutation — since the TYPES coproduct already admits Unknown(Partial) as a constructor.

Design principles:

• Classification transparent: Show classify results (dimension + verifiability) for all uncertainties — "visible by default, ask only on exception"
• Context collection transparent: Show what evidence was collected and what remains uncertain
• Current uncertainty framed: Surface the uncertainty currently in play as framing — the kind of context being resolved this cycle — rather than a completion count across all uncertainties
• Actionable options: Each option leads to a concrete next step
• Dismiss with default: Always state what assumption will be used if dismissed

Selection criterion: Choose the uncertainty whose resolution would maximally narrow the remaining uncertainty space (information gain). When priority is equal, prefer the uncertainty with richer collected context (more evidence to present).

Phase 3: Prospect Update

After user response:

1. Provide(context): Integrate user-provided context into prospect X'
2. Point(location): Record location, resolve via next Phase 1 iteration
3. Dismiss: Mark uncertainty as dismissed, note default assumption used

After integration:

• Re-scan X' for remaining or newly emerged uncertainties
• If uncertainties remain: return to Phase 1 (collect context for new uncertainties first)
• If all resolved/dismissed: proceed with execution
• Log (Uncertainty, A) to history

Intensity

Level	When	Format
Light	Marginal priority uncertainties only	Constitution interaction with Dismiss as default option
Medium	Significant priority uncertainties, context collection partially resolved	Structured Constitution interaction framing the current uncertainty
Heavy	Critical priority, multiple unresolved uncertainties	Detailed evidence + collection results + classify results + resolution paths

UX Safeguards

Rule	Structure	Effect
Gate specificity	activate(Aitesis) only if ∃ requirement(r) : ¬available(r) ∧ ¬trivially_inferrable(r)	Prevents false activation on clear tasks
Context collection first	Phase 1 before Phase 2	Enriches question quality before asking
Uncertainty cap	One coherent cluster per Phase 2 cycle (size ≤ 4, default), priority order, per-item non-overlapping information-gain leverage; cluster size > 1 requires Clustering basis: [decision frame] — per-item gain rationale: [item₁: reason₁; item₂: reason₂; ...] cite (auditable parallel to Cite-or-observe); yields to Divergence-bounding when frame-first triggers fire — items whose answers depend on other cluster items form a compound question (forbidden — collapses decision space into 2ⁿ implicit states)	Prevents question overload while permitting shared-frame clusters
Session immunity	Dismissed (domain, description) → skip for session	Respects user's dismissal
Current-uncertainty framing	Phase 2 surfaces the uncertainty currently in play (the kind of context being resolved this cycle) — a framing readout, not an [N resolved / M] completion count	User recognizes which context is being resolved without parsing a progress tally; granular progress stays in session
Narrowing signal	Signal when narrowing(Q, A) shows diminishing returns	User can exit when remaining uncertainties are marginal
Early exit	User can declare sufficient at any Phase 2	Full control over inquiry depth
Cross-protocol fatigue	Syneidesis triggered → suppress Aitesis for same task scope	Prevents protocol stacking (asymmetric: Aitesis context uncertainties ≠ Syneidesis decision gaps, so reverse suppression not needed)
Classify transparency	Always show classify results (dimension + verifiability) in Phase 2 surfacing format	User sees AI's reasoning and resolution path per uncertainty
Routing transparency	Uₙ items show → /protocol in Phase 2 classify summary	User sees routing destination at detection time
Observation transparency	Log observation lifecycle in Λ.observation_history	User can audit what was tested
Skip transparency	Log escape hatch usage in Λ.observation_skips with condition + rationale	User can audit why observation was skipped
Observation cleanup	All test artifacts removed after observation	No residual files
Observation timeout	30s limit → fall back to user inquiry	Prevents hanging
Observation risk gate	Elevated-risk observation → reclassify as UserDependent	Safety preserved
Free-response escape	Phase 2 Constitution interaction accepts Unknown(Partial) as typed answer constructor (TYPES A) — Phase 3 auto-promotes uncertainty to next-preferred EvidenceSource in ValidSources(v) and re-enters Phase 1 classification via backward arc (PHASE TRANSITIONS)	User need not simulate certainty for unknowns; routing degrades gracefully through typed, formally-tracked arc
Cost-ordering tiebreaker	Default ascending cost-ordering defined in TYPES (CodeDerivable < CanonicalExternal < Instrumentation < UserTacit). External-dependency preference: when uncertainty carries environmental-dependency signal, prefer CanonicalExternal over CodeDerivable unless authority-override basis cited. Override logged to Λ.source_choice_overrides	Minimizes user-interrupt cost; preserves external authority when relevant; Cite-or-observe basis required for UserTacit preference
EvidenceSource transparency	Factual uncertainty classify summary shows EvidenceSource: {selected} with basis citation when override applied; Λ.source_choice_overrides preserves cite audit across sessions for variation-stable observed use	Cite-or-observe rule audit visibility

Rules

1. AI-guided, user-resolved (Cognitive Partnership Move): AI infers context insufficiency; resolution requires user choice via Cognitive Partnership Move (Constitution) at Phase 2 — present options with differential futures, yield turn, parse typed answer.
2. Recognition over Recall: Present structured options with anticipatable post-selection states; Constitution interaction requires turn yield before proceeding.
3. Evidence over Inference over Detection (Aitesis Core Principle): Three-level hierarchy. Lower boundary (Inference > Detection): infer the highest-gain question rather than detecting via fixed checklist. Upper boundary (Evidence > Inference): when a factual uncertainty is empirically observable, observe directly rather than infer from reasoning alone — Cite-or-observe is the structural guard. Direct-resolve admissibility has two axes: coverage (verify scope ⊇ claim) and support_integrity (verify provenance_coupled(source × claim) and the evidence→behavior link, not silently desynced). Partial coverage is inference for the uncovered portion; present evidence with the wrong referent/source-kind/scope is inference dressed as evidence for a different claim; current-but-unenforced evidence is inference dressed as evidence for the behavior it asserts. Verify both before treating as resolved, and reclassify a failing portion separately (coverage gap → split; support-unlinked or provenance-uncoupled → EmpiricallyObservable). Inquiry above Evidence in cost: facts discoverable by AI through Evidence (codebase, memory, observation, canonical external) are resolved at Phase 1; Phase 2 surfaces only judgment-requiring uncertainties.
4. Open scan: No fixed uncertainty taxonomy — identify uncertainties dynamically per prospect requirements; every surfaced uncertainty cites specific observable evidence or collection results, not speculation.
5. Context collection precedes classification precedes inquiry: Phase 1 Step 1 (codebase + memory + history collection) → Step 2 (dimension + verifiability + EvidenceSource classification) → Step 3 (read-only verification) → Step 4 (empirical observation) before Phase 2 inquiry. Memory and version-control evidence is staleness-guarded — tagged source: "memory:{path}" or source: "history:{ref}" and NOT eligible for Factual/ReadOnly direct-resolve; verify against current state or escalate via classify summary with the staleness:unverified flag (the temporal sub-case of the general support_integrity:unverified tag). Procedural detail in Phase 1 Step 1-2.
6. Convergence persistence: Mode active until all identified uncertainties are resolved (context, read-only, observed, user-responded) or dismissed; convergence demonstrated via per-uncertainty transformation trace before declaring remaining = ∅.
7. Always show classification: Phase 2 surfacing always includes classify results (dimension + verifiability + EvidenceSource); free response can override classification — visible by default, ask only on exception.
8. Context-Question Separation: Analysis, evidence, rationale as text output before the gate; the gate contains the essential question and option-specific differential implications only.
9. Option-set relay test (Extension classification): Single dominant option (entropy → 0) presented as relay. Each Constitution option genuinely viable under different user value weightings; shared-trajectory options collapse to one; off-axis prompts surface as free-response pathways rather than peer options.
10. One coherent cluster: Surface one coherent cluster per Phase 2 cycle (size ≤ 4 default); cluster admissible when items share decision frame, have non-overlapping information-gain rationale, and are independently answerable — compound questions (where one item's answer depends on another) collapse the decision space into 2ⁿ implicit states, degrading Recognition. Clustering-basis cite required when cluster size > 1. Yields to Divergence-bounding (frame-first) when divergence triggers fire — frame first, then cluster within selected frame. Procedural detail in UX Safeguards (Uncertainty cap entry).
11. Divergence-bounding (frame-first): Compute per-uncertainty branching_factor(u) during Phase 1 Ctx. When |Uᵢ| × max branching_factor(u) ≥ 16 OR Σᵤ branching_factor(u) ≥ 16, present a decision frame option set before per-branch detail; supersedes One coherent cluster when both apply. Exception (full enumeration without frame-first) requires BOTH |Uᵢ| ≤ 3 AND each branching_factor(u) ≤ 2 AND cited per-uncertainty counts (silent self-assessment "uniformly low" forbidden). branching_factor definition in TYPES.
12. Cite-or-observe (Evidence-Inquiry boundary structural guard): When a Factual uncertainty has a cheaper EvidenceSource available than UserTacit, resolve via cheaper source OR cite explicit dominance basis (logged to Λ.source_choice_overrides). Reclassification of EmpiricallyObservable to UserDependent is legitimate only under empirically verifiable escape conditions evaluated BEFORE observation (env mutation / 30s bound / elevated risk; logged to Λ.observation_skips). Observation outputs always proceed to Phase 2 via Uₑ — positive (differentiating evidence) or negative (null-signal); pre-observation categorical classification is not a substitute. On the ReadOnly direct-resolve path, support_integrity is the structural test for the same upper boundary: evidence that is present (and current) but not coupled to the source × claim pair or behavior it asserts (support_integrity_unverified) is inference, not evidence, and reclassifies to EmpiricallyObservable rather than auto-resolving; admissibility on this path requires confirming provenance coupling and evidence→behavior coupling, beyond freshness. Verifiability-tier dominance basis detail in Phase 1 Step 2.
13. No pre-filter rationalization (Cite-or-observe dual): Pre-filter (Coherence coexistence exit) applies only when an explicit, named scope hierarchy rule or documented precedence ordering resolves the apparent contradiction without epistemic protocol intervention. Classifying genuine cross-domain structural contradiction as "rule-resolvable" to avoid routing = pre-filter misuse.
14. Gate integrity (Safeguard tier): The defined option set is presented intact — option injection/deletion/substitution each violate this invariant. Type-preserving materialization (specializing a generic option while preserving the TYPES coproduct) is distinct from mutation.
15. Plain emit discipline: User-facing emit (Phase 2 surfacing prose, convergence traces, gate options, and any text shown to the user) uses everyday language to reduce the user's cognitive load — every emit token should carry decision-relevant meaning, not project-internal overhead. SKILL.md formal-block vocabulary — variable names with subscripts, Greek-rooted terms in narrative, formal type labels inline, and code-style backtick tokens — stays in the formal block. What the user reads is the action, observation, or question in their idiom.
16. Round-local salience bundling: Each user-facing round bundles the current judgment, its nearest evidence, and the differential implication that matters for the next move. Keep adjacent material together so the user can recognize the decision without context-switching; defer background, distant context, and unrelated findings to pre-gate text, convergence traces, or later cycles.
17. Formal blocks are runtime-normative: This protocol's formal blocks — those defined in its Definition code block above — are LLM-facing and constitutive of protocol identity: they type the prose and carry the operational contract executed at runtime. A reduced or single-shot realization carries every one of them through as runtime contract, since each block is the type that constitutes the protocol — preserving the blocks keeps the protocol intact. How its symbols render to the user is a separate emit-layer concern (see Plain emit discipline).