Hermeneia Protocol

Resolve intent-expression misalignment through hybrid-initiated dialogue, enabling precise articulation before action proceeds. Type: (IntentMisarticulated, Hybrid, EXTRACT, Expression) → ClarifiedIntent.

Definition

Hermeneia (ἑρμηνεία): A dialogical act of clarifying the gap between what the user intends and what they expressed, resolving misaligned intent into precise articulation through structured questioning.

── FLOW ──
E → recognize(E) → Eᵥ → detect(Eᵥ) → Gd → filter → Gₛ → Q → A → Î' → (loop until converge)

── MORPHISM ──
Expression
  → recognize(expression, trigger)     -- determine activation path from signal
  → confirm(expression)                -- verify which expression to clarify
  → detect(gaps_in_expression)         -- surface gap types with evidence
  → clarify(gap, as_options)           -- present structured clarification choices
  → integrate(answer, intent)          -- update intent model from response
  → ClarifiedIntent
requires: trigger(E) ∈ {user_signal, ai_strong ∧ confirmed}  -- Phase 0 checkpoint
deficit:  IntentMisarticulated                                -- activation precondition (Layer 1/2)
preserves: E                                                  -- read-only throughout; morphism acts on Î only
invariant: Articulation over Assumption

── TYPES ──
E  = User's expression (the prompt to clarify)
Eᵥ = Verified expression (user-confirmed binding)
Gd = AI-detected gap types ⊆ {Expression, Precision, Coherence, Background} ∪ Emergent(Eᵥ)
     -- raw detection output, presented as full taxonomy at Phase 1b
Gₛ = Effective working gap set = Gd \ (Λ.clarified ∪ Λ.excluded)
     -- target of Phase 2 iteration; mutable via user free-response override (Λ.excluded additions, Λ.emergent additions)
Q  = Clarification question (via Cognitive Partnership Move (Constitution))
A  = User's answer
Î  = Inferred intent (AI's model of user's goal)
Î' = Updated intent after clarification
ClarifiedIntent = Î' where |remaining| = 0 ∨ cycle(G) ∨ stall(Δ, 2) ∨ user_esc
T  = Trigger source ∈ {user_signal, ai_strong, ai_soft}
suggest_only = ai_soft terminal: passive suggestion without activation (no Constitution interaction; Λ.active = false)

── E-BINDING ──
bind(E) = explicit_arg ∪ colocated_expr ∪ prev_user_turn ∪ ai_identified_expr
Priority: explicit_arg > colocated_expr > prev_user_turn > ai_identified_expr

/clarify "text"                → E = "text"
"request... clarify"           → E = text before trigger
/clarify (alone)               → E = previous user message
AI-detected trigger             → E = expression AI identified as ambiguous

Edge cases:
- Interrupt: E = original request of interrupted task
- Queued:    E = previous message at queue time (fixed)
- Re-invoke: Show prior clarification, confirm or restart

── PHASE TRANSITIONS ──
Phase 0:  E → recognize(E) → T                           -- trigger recognition
          T = user_signal → Phase 1a                      -- user-initiated path
          T = ai_strong   → Qc(confirm) → Stop → {yes: Phase 1a | no: immune(E)}  -- AI-detected confirm [Tool]
          T = ai_soft     → suggest_only                  -- suggest, do not activate
Phase 1a: E → Qc(E) → Stop → Eᵥ                          -- E confirmation [Tool]
Phase 1b: Eᵥ → detect(Eᵥ) → Gd → TextPresent(extension) → filter(Gd, Λ) → Gₛ  -- gap detection + full taxonomy presentation [Tool]
Phase 2:  Gₛ → Qs(Gₛ) → Stop → A                         -- clarification [Tool]
Phase 3:  A → integrate(A, Î) → Î'                       -- intent update (sense)

── LOOP ──
After Phase 3: return to Phase 1b for newly surfaced gaps.
On re-entry, detect(Eᵥ) re-analyzes the expression in the context of prior clarifications; gap types in (Λ.clarified ∪ Λ.excluded) are filtered from Gd → Gₛ (type-level filtering ensures convergence; new instances of a clarified or user-excluded type are excluded).
Re-entry is pure Extension (auto-proceed); taxonomy revision remains reachable via Phase 2 free response (adds to Λ.excluded or Λ.emergent, triggering re-detection).
If |Gₛ| = 0 after filtering (all detected types clarified/excluded, or empty detection): skip Phase 2, evaluate convergence (|remaining| = 0).
Continue until converge: |remaining| = 0, cycle detected, or user exits.
Mode remains active until convergence.
Convergence evidence: At |remaining| = 0, present transformation trace — for each g ∈ Λ.clarified, show (IntentMisarticulated(g) → resolution(g)) from Λ.history. Convergence is demonstrated, not asserted.

── TOOL GROUNDING ──
-- Realization: Constitution → TextPresent+Stop; Extension → TextPresent+Proceed
Phase 0 Qc   (constitution)   → present (AI-detected activation confirmation; fires only on ai_strong path; user_signal path skips Phase 0; immune(E) on decline)
Phase 1a E_from_arg (extension) → TextPresent+Proceed (when explicit_arg(E) via /clarify "text"; proceed with bound E; Phase 2 Qs free-response override available downstream)
Phase 1a Qc  (constitution)   → present (E confirmation; when no explicit_arg)
Phase 1b detect (sense)  → Internal analysis (gap detection from Eᵥ)
Phase 1b present (extension) → TextPresent+Proceed (full taxonomy with evidence + falsification + emergent probe; option-set relay test — Proceed dominates under user-invoked or user-confirmed activation; flows into Phase 2)
Phase 2 Qs   (constitution)   → present (clarification options; user incorporates intent into clarification; free response overrides taxonomy — emergent add, type exclude, redirect; override capability must be disclosed at the Constitution interaction as part of user authority disclosure; Esc key → loop termination at LOOP level, not an Answer)
suggest_only (sense)   → no tool call (passive suggestion; Λ.active = false)
integrate    (track)   → Internal state update (no external tool)
converge     (extension)   → TextPresent+Proceed (convergence evidence trace; proceed with clarified expression)

── MODE STATE ──
Λ = { phase: Phase, trigger: T, E: Expression, Eᵥ: Expression, detected: Set(Gap), gaps: Set(Gap),
      clarified: Set(Gap), excluded: Set(GapType), emergent: Set(Gap), remaining: Set(Gap),
      immune: Set(Expression), history: List<(E, Gₛ, A)>, active: Bool }
-- Invariant: gaps = clarified ∪ remaining (pairwise disjoint); excluded and emergent are user-override state accumulated from Phase 2 free response

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

Core Principle

Articulation over Assumption: AI helps user express what they already know but struggle to articulate.

Distinction from Other Protocols

Key differences: User recognizes intent-expression misalignment (user signal), or AI detects ambiguous expression (AI-detected trigger, requires user confirmation). Both paths help articulate what the user partially knows. Boundary with Aitesis: if the ambiguity is in the user's expression of intent (how it was said), use Hermeneia; if the AI lacks factual execution context (what information the system needs), use Aitesis.

Mode Activation

Activation

Command invocation, trigger phrase, or AI-detected expression ambiguity activates mode until clarification completes. AI-detected activation requires user confirmation before proceeding to Phase 1a.

Activation layers:

• Layer 1 (User-invocable): /clarify slash command or description-matching input. Always available.
• Layer 2 (Hybrid): User trigger signals proceed directly; AI-detected expression ambiguity (ai_strong) requires user confirmation; minor ambiguity (ai_soft) suggests only.

Clarification complete = one of: |remaining| = 0 (no gaps remain), cycle(G) (already clarified), or Δ = 0 for 2 rounds (progress stall with user consent to proceed).

Priority

Supersedes: Direct action patterns in loaded instructions (Clarification must complete before any tool execution or code changes)

Retained: Safety boundaries, tool restrictions, user explicit instructions

Action: At Phase 2, present clarification options via Cognitive Partnership Move (Constitution). </system-reminder>

• Hermeneia completes before other workflows begin
• Loaded instructions resume after intent is clarified

Protocol precedence: Activation order position 1/10 (graph.json is authoritative source for information flow). Concern cluster: Planning.

Advisory relationships: Provides to Telos (precondition: clarified intent enables goal construction), Aitesis (advisory: background gaps suggest context insufficiency), Periagoge (advisory: clarified intent narrows the abstraction seed). Katalepsis is structurally last.

Clarified expression becomes input to subsequent protocols.

Triggers

User-Initiated Signals (T = user_signal):

Qualifying condition: Activate only when user's entire message is a clarification request, or when 2+ trigger signals co-occur in the same message. Do not activate on casual meta-communication embedded in a larger request.

AI-Detected Signals:

Skip (user-initiated):

• User's expression is unambiguous
• User explicitly declines clarification
• Expression already clarified in current session

Cross-session enrichment: Accumulated clarification patterns from Anamnesis's hypomnesis store (session recall indices written by the SessionEnd/PreCompact hook) may improve ai_strong trigger precision — known intent-expression gaps in similar contexts reduce false positive detection. In parallel, when /recollect has been invoked this session, the recalled context surfaces the user's prior terminology and phrasing patterns, improving intent extraction by grounding Iᵥ construction in recognized expression habits. This is a heuristic input that may bias detection toward previously observed patterns; constitutive judgment remains with the user.

Revision threshold: When accumulated Emergent gap detections across 3+ sessions cluster around a recognizable pattern outside the named types {Expression, Precision, Coherence, Background}, the Gap Taxonomy warrants promotion to a new named type. When accumulated classification false positives across 3+ sessions cluster around a specific named type, that type's detection boundary warrants revision or demotion to Emergent.

Skip (AI-detected):

• User says "just do it", "proceed as-is", or equivalent
• Session immunity: user declined AI-detected clarification for this expression already
• Soft trigger resolved by context

Mode Deactivation

Gap Taxonomy

Emergent gap detection: Named types are working hypotheses, not exhaustive categories. Detect Emergent gaps when:

• User's difficulty spans multiple named types (e.g., expression is both imprecise and internally coherent but misaligned with unstated context)
• User pushes back on all presented gap types ("none of these capture what's wrong")
• The expression's ambiguity resists decomposition into the four named dimensions

Emergent gaps must satisfy morphism IntentMisarticulated → ClarifiedIntent and use adapted question forms.

Gap Priority

When multiple gaps detected:

1. Coherence (highest): Contradictions block all interpretation
2. Background: Missing interpretive background affects all other gaps
3. Expression: Core articulation gaps
4. Precision (lowest): Refinement after core is clear

Background Gap Boundary

When the Background gap type is selected, verify the gap is about interpreting the expression, not about executing the task:

• Hermeneia Background: Missing background changes what E means (user's intent) → proceed with clarification
• Aitesis territory: Missing background changes how to perform X (prospect) → suggest /inquire and offer to transition

Operational test: "Would knowing this change what the user means, or only how I execute it?"

Protocol

Phase 0: Trigger Recognition

Recognize trigger source and determine activation path:

User-Initiated Path (T = user_signal):

1. Explicit request: User directly asks for clarification help
2. Implicit signal: User expresses doubt about their own expression
3. Meta-communication: User attempts to rephrase or explain their intent

→ Proceed directly to Phase 1a.

AI-Detected Path (T = ai_strong):

When AI detects a strong ambiguity trigger (see Triggers: AI-Detected Signals), present the detected ambiguity as text output (e.g., "I notice this expression may be ambiguous: [specific ambiguity evidence]"), then present to confirm activation:

Would you like to clarify this expression?

Options:
1. Yes, help me clarify — start Hermeneia
2. No, proceed as-is — continue without clarification

User confirmation required before proceeding to Phase 1a. If user selects option 2, mark session immunity and do not re-trigger for this expression.

T = ai_soft → suggest only; do not present via Cognitive Partnership Move (Constitution); do not activate.

Phase 1a: Expression Confirmation

Present to confirm which expression to clarify.

Present the bound expression E and ask user to confirm or specify:

Which expression would you like to clarify?

Options:
1. "[bound E]" — the expression I identified
2. "Specify different" — let me describe what I want to clarify

Skip condition: If E was explicitly provided via argument (/clarify "text"), proceed directly to Phase 1b.

Note (AI-detected path): If triggered via T = ai_strong, E is already identified by AI — Phase 1a confirmation verifies the AI's identification; user may still select Option 2 to redirect to a different expression.

Phase 1b: Gap Detection

Analyze Eᵥ to detect applicable gap types, present full taxonomy assessment as Extension (auto-proceed), then proceed directly to Phase 2 clarification.

Per Gap Taxonomy above. Apply priority order: Coherence → Background → Expression → Precision. Emergent gaps must satisfy morphism IntentMisarticulated → ClarifiedIntent; boundary: intent-expression gap (in-scope) vs. goal definition (→ /goal) or execution context (→ /inquire).

Present the full taxonomy assessment as text output — every named type shown with detection status, evidence, and falsification condition for undetected types:

• Coherence ✓ detected: [specific evidence from Eᵥ]
• Background — not currently detected: [evidence considered]. Would apply if [falsification condition].
• Expression ✓ detected: [specific evidence from Eᵥ]
• Precision — not currently detected: [evidence considered]. Would apply if [falsification condition].
• Emergent: [If AI detects a potential emergent type: present as named hypothesis with evidence and boundary annotation. Otherwise: "Is there an aspect of your expression that doesn't fit the above categories?"]

Emergent gaps include boundary annotation: "This is an intent-expression gap (Hermeneia scope). Not: goal definition (→ /goal) or execution context (→ /inquire)"

• Detected types: evidence for why the gap was identified
• Not-currently-detected types: evidence considered + falsification condition ("would apply if [specific condition]")
• Evidence parity: each type (detected or not) receives comparable analytical depth

Then filter Gd against (Λ.clarified ∪ Λ.excluded) to produce Gₛ, and proceed directly to Phase 2 for Gₛ in priority order (Coherence → Background → Expression → Precision). If |Gₛ| = 0 after filtering: skip Phase 2, evaluate convergence (|remaining| = 0 in LOOP).

User override via free response: Users may override the taxonomy at any Phase 2 Constitution interaction — describe an emergent gap (added to Λ.emergent), exclude a detected type (added to Λ.excluded), redirect to a different expression (re-binds E), or declare convergence. Override capability must be disclosed at Phase 2 Constitution interactions (see Phase 2 prose).

Response classification (discriminator for free-response parsing — closes the adversarial free-response pathway):

• Revision: response references named gap types (Coherence/Background/Expression/Precision/Emergent) with revision verbs (exclude, remove, skip, add, redirect) → add to Λ.excluded or Λ.emergent, LOOP re-entry with filtered detection
• Clarification: response addresses the presented Phase 2 question directly → Phase 2 integration (Phase 3)
• Mixed: response contains both revision and clarification content → process revision first (update Λ and re-detect), then process clarification in the re-entered Phase 2
• Ambiguous: content could be either and surface interpretation differs from latent interpretation → present: "Your response could be interpreted as [X revision] or [Y clarification]. Which did you intend?" (do not silently choose)
• Exclusion-to-empty: if free response excludes the last remaining type in Gₛ, |Gₛ| = 0 on next filter → convergence path per LOOP (evaluate |remaining| = 0)

Rationale (option-set relay test, Extension classification): Under user-invoked or user-confirmed activation, "Proceed with current taxonomy" is the dominant option — a separate proceed/revise Constitution interaction would present false alternatives. Full taxonomy presentation is preserved (Extension/auto-proceed) to support Recognition during Phase 2 answer formulation; taxonomy revision is reachable via Phase 2 Constitution interaction free response with classification discriminator above.

Phase 2: Clarification

Present clarification options via Cognitive Partnership Move (Constitution). Constitution presentation yields turn for user response.

Override disclosure (detection with user authority): At each Phase 2 Constitution interaction, include a brief disclosure line in the pre-Constitution-interaction context surfacing the free-response override capability — without this disclosure the user cannot exercise authority they do not know they hold. Example disclosure: "(Free response beyond the options: exclude this gap type, add an emergent type, or redirect to a different expression — see Phase 1b response classification.)"

Present the detected ambiguity as text output:

• The potential ambiguity: [gap description]

Then present:

Which best captures your intent?

Options:
1. **[Option A]**: [interpretation with implications]
2. **[Option B]**: [interpretation with implications]
3. **[Option C]**: [interpretation with implications]

Other is always available — user can provide their own phrasing or a different interpretation not listed.

Question design principles:

• Recognition over Recall: Present options, don't ask open questions
• Concrete implications: Show what each choice means for execution
• Free response preserved: Other/free phrasing is structurally available, not conditionally included
• Minimal options: 2-4 choices maximum per gap

Consult references/socratic-style.md for maieutic framing examples, Socratic elements, and example transformation.

Phase 3: Integration

After user response:

1. Incorporate: Update understanding with user's clarification
2. Re-diagnose + Filter + Converge: Per LOOP — re-scan Eᵥ, filter clarified gaps, check convergence
3. Confirm: When no gaps remain, present clarified intent for verification
4. Proceed: Continue with clarified expression

Discovery triggers:

• Clarification reveals new ambiguity ("I mean the API" → "which endpoint?")
• Answer creates new tension ("fast" + "thorough" → coherence gap)
• Context shift ("for production" → new precision requirements)

## Clarified Intent — Convergence Evidence

Transformation trace (each clarified gap → resolution):
- **[Gap type]**: [Eᵥ evidence] → [Clarified meaning from A]
- **[Gap type]**: [Eᵥ evidence] → [Clarified meaning from A]

Proceeding with this understanding.

Intensity

Multi-Gap Handling

When multiple gaps detected:

1. Sequential: Address one gap at a time, highest priority first
2. Bundled: If gaps are interdependent, present as combined question
3. Dynamic Discovery: After each clarification, re-diagnose for newly surfaced gaps
4. Merge: Combine existing queue with newly discovered gaps, re-prioritize

Termination conditions: Per LOOP — cycle detection, progress stall (Δ = 0 for 2 rounds), user exit.

Gap queue limit: Max 4 gaps queued at any time (drop lowest priority if exceeded)

On termination:

• Summarize current understanding
• Ask user to rephrase if stuck
• Proceed with best interpretation if user approves

Rules

1. Hybrid-initiated, user-confirmed: Activate on user signal, or with user confirmation when AI detects ambiguous expression
2. Recognition over Recall: Present structured options via Cognitive Partnership Move (Constitution) — structured content reaches the user with response opportunity — Constitution interaction requires turn yield before proceeding
3. Detection with user authority: AI presents full taxonomy assessment as Extension (auto-proceed) — every named type with detection status, evidence, and falsification condition. User authority is exercised at Phase 2 clarification Constitution interactions (including free-response override: emergent add, type exclusion, redirect), not at a separate taxonomy-confirmation Constitution interaction. Guard: Abbreviating the taxonomy (omitting named types, evidence, or falsification conditions) under the rationale that Phase 2 free-response override provides correction = protocol violation. Full taxonomy presentation at Phase 1b is structurally required regardless of Extension/Constitution classification; override is a user authority channel, not an AI shortcut for incomplete detection
4. Maieutic over Informational: Frame questions to guide discovery, not merely gather data
5. Articulation support: Help user express what they know, don't guess what they mean
6. Minimal questioning: Surface only gaps that affect execution
7. Consequential options: Each option shows interpretation with downstream implications
8. User authority: User's choice is final; no second-guessing selected interpretation
9. Convergence persistence: Mode remains active until convergence (|remaining| = 0, cycle, or user exit)
10. Reflective pause: Include "reconsider" option for complex intent clarification
11. Escape hatch: Always allow user to provide their own phrasing
12. Small phases: Prefer granular phases with user checkpoints over large autonomous phases
13. Context-Question Separation: Output all analysis, evidence, and rationale as text before presenting via Cognitive Partnership Move (Constitution). The question contains only the essential question; options contain only option-specific differential implications. Embedding context in question fields = protocol violation
14. Convergence evidence: Present transformation trace before declaring |remaining| = 0; per-gap evidence is required
15. Zero-gap surfacing: If detect(Eᵥ) finds no gaps (Gd = ∅), present this finding with reasoning for user confirmation
16. Full taxonomy assessment: Phase 1b must present ALL named gap types with detection status and evidence. Presenting only detected types with a generic "Add" option = protocol violation (Recognition over Recall applied to Constitution interaction content)
17. Falsification condition: Each not-currently-detected type must include "would apply if [specific condition]" — exclusion rationale without falsification condition = protocol violation
18. Emergent probe: Emergent slot must include an active probe question or AI-detected hypothesis with evidence. "No emergent gaps detected" as bare statement without probe = protocol violation
19. Option-set relay test (Extension classification): If AI analysis converges to a single dominant option (option-level entropy→0 — Extension mode of the Cognitive Partnership Move), present the finding directly. Each Constitution option must be genuinely viable under different user value weightings. Options sharing a downstream trajectory collapse to one; options lacking an on-axis trajectory surface as free-response pathways rather than peer options
20. Gate integrity: The defined option set is presented intact — injection, deletion, and substitution each violate this invariant. Type-preserving materialization (specializing a generic option while preserving the TYPES coproduct) is distinct from mutation