aiox-core 5.0.4 → 5.0.5

package/pro/squads/squad-creator-pro/docs/optimize-v4-proposal.md

@@ -1,354 +0,0 @@
-# Proposta: optimize.md v4.0
-
-**Baseado em:** Sessão de Model Tier Qualification 2026-02-11
-**Evidência:** 8 tasks testadas, 3 padrões críticos descobertos
-
----
-
-## Descobertas Empíricas
-
-### 1. Binary Checkpoint Conversion (99.3% match)
-
-**Problema:** Tasks com scoring subjetivo (1-5) geram divergência Opus/Haiku.
-
-**Solução comprovada:** Converter para checkpoints binários.
-
-```yaml
-ANTES (divergência):
-  an-fidelity-score v1.0:
-    scoring: "Avaliar 1-5 cada layer"
-    opus: 82.22%
-    haiku: "varia muito"
-
-DEPOIS (99.3% match):
-  an-fidelity-score v2.0:
-    scoring: "5 checkpoints binários por layer"
-    opus: 82.22%
-    haiku: 81.67%
-    match: 99.3%
-```
-
-**Nova fase proposta:** PHASE 4b: BINARY CHECKPOINT CONVERSION
-
-### 2. Script-First Architecture (0 tokens)
-
-**Problema:** Tasks classificadas como "Hybrid" ainda gastam tokens desnecessários.
-
-**Solução comprovada:** Script pode fazer 90-95% do trabalho.
-
-```yaml
-fidelity-score.sh:
-  lines: 518
-  time: "<1 second"
-  tokens: 0
-  coverage: "95% da task"
-  llm_only_for: "gap recommendations"
-```
-
-**Nova regra:** SE 90%+ é determinístico → SCRIPT-ONLY, não Hybrid.
-
-### 3. Scope Interpretation (Haiku literal)
-
-**Problema:** Haiku interpreta escopo LITERALMENTE.
-
-```yaml
-an-clone-review:
-  task_scope: "Review clone file"
-  opus_interpretation: "wrapper + delegated persona"
-  haiku_interpretation: "wrapper only"
-  result:
-    opus: 93.3% SOLID
-    haiku: 73.3% NEEDS_WORK
-    match: 78.6% (VETO)
-```
-
-**Nova fase proposta:** PHASE 1b: SCOPE CLARIFICATION
-
-### 4. Bidirectional Scoring Bias
-
-**Descoberta crítica:** Haiku pode ser CONSERVADOR ou GENEROSO, dependendo da task.
-
-```yaml
-conservative_example:
-  task: an-clone-review
-  opus: 93.3%
-  haiku: 73.3%  # -20% (mais rigoroso)
-
-generous_example:
-  task: an-validate-clone
-  opus: 73.89%
-  haiku: 88.3%  # +14% (mais generoso)
-```
-
-**Nova fase proposta:** PHASE 7b: BIDIRECTIONAL BIAS TEST
-
-### 5. Gatekeeper Pattern
-
-**Descoberta:** Tasks com decisão pass/fail binária são Haiku-safe mesmo com scores diferentes.
-
-```yaml
-qa-after-creation:
-  opus_score: 8.32
-  haiku_score: 9.9  # +19% diferente
-  opus_decision: APPROVED
-  haiku_decision: APPROVED  # IGUAL!
-  verdict: HAIKU QUALIFIED
-```
-
-**Nova fase proposta:** PHASE 1c: GATEKEEPER DETECTION
-
----
-
-## Novas Fases Propostas
-
-### PHASE 1b: SCOPE CLARIFICATION
-
-```yaml
-scope_clarification:
-  trigger: "After Phase 1 decomposition"
-  duration: "2-3 minutes"
-
-  step_1b_1:
-    action: "Identify all files referenced in task"
-    check:
-      - "Does task reference other files? (delegated persona, configs, etc)"
-      - "Does task use dynamic paths? (e.g., {mind_path})"
-      - "Does task have implicit dependencies?"
-
-  step_1b_2:
-    action: "Add explicit scope definition"
-    template: |
-      ## SCOPE DEFINITION
-
-      **Primary file:** {main_file}
-      **Include in scope:**
-      - [ ] Delegated persona files (if wrapper)
-      - [ ] Referenced configs
-      - [ ] Source directories
-
-      **Explicit scope:**
-      ```
-      WHEN analyzing this clone/squad/task:
-      - INCLUDE: {list of files}
-      - EXCLUDE: {list of files}
-      - IF wrapper delegates to persona → READ BOTH
-      ```
-
-  veto: |
-    Tasks with ambiguous scope MUST NOT proceed to Haiku testing.
-    Clarify scope FIRST, then test.
-```
-
-### PHASE 4b: BINARY CHECKPOINT CONVERSION
-
-```yaml
-binary_checkpoint_conversion:
-  trigger: "After Phase 4 (code stubs), if task has subjective scoring"
-  duration: "10-15 minutes"
-
-  detect_subjective_scoring:
-    patterns:
-      - "Score 1-5"
-      - "Avaliar qualidade"
-      - "Rate from"
-      - "Evaluate"
-      - "Quality score"
-
-  conversion_template:
-    for_each_subjective_score:
-      original: "Score layer X from 1-5"
-      converted: |
-        ## Layer X (5 checkpoints)
-
-        | # | Checkpoint | O que procurar | Passa se... |
-        |---|------------|----------------|-------------|
-        | 1 | {specific_check} | {where_to_look} | {binary_condition} |
-        | 2 | {specific_check} | {where_to_look} | {binary_condition} |
-        | 3 | {specific_check} | {where_to_look} | {binary_condition} |
-        | 4 | {specific_check} | {where_to_look} | {binary_condition} |
-        | 5 | {specific_check} | {where_to_look} | {binary_condition} |
-
-        **Score = count(passed) → 0-5**
-
-  add_scoring_calibration:
-    template: |
-      ## SCORING CALIBRATION (CRITICAL)
-
-      ```yaml
-      scoring_philosophy:
-        principle: "SCORE O QUE EXISTE, não o que falta"
-        bias_correction: "Haiku tende a sub-pontuar. Compensar sendo generoso."
-        evidence_rule: "Se existe evidência, conta ponto. Gaps vão para recommendations."
-
-        checkpoint_rules:
-          - "PASS = evidence EXISTS (not perfect)"
-          - "FAIL = evidence MISSING (not weak)"
-          - "Partial = PASS (generous interpretation)"
-      ```
-
-  quality_gate:
-    - "All subjective scores converted to binary checkpoints?"
-    - "Scoring calibration section added?"
-    - "Each checkpoint has specific lookup location?"
-    - "Binary condition is yes/no, not judgment?"
-```
-
-### PHASE 1c: GATEKEEPER DETECTION
-
-```yaml
-gatekeeper_detection:
-  trigger: "During Phase 1 analysis"
-  duration: "1-2 minutes"
-
-  definition: |
-    A GATEKEEPER task has ONE primary output: PASS or FAIL.
-    Score details are secondary. The decision is what matters.
-
-  detection_criteria:
-    is_gatekeeper:
-      - "Output is PASS/FAIL or APPROVED/REJECTED?"
-      - "Score is used to DERIVE decision, not as final output?"
-      - "Users care about decision, not exact score?"
-
-  if_gatekeeper:
-    action: "Add gatekeeper marker to task metadata"
-    haiku_qualification: |
-      For GATEKEEPER tasks, Haiku qualifies if:
-      - Final DECISION matches Opus (not score)
-      - Score is within 30% (allows variance)
-
-      Example:
-        opus: 8.32 → APPROVED
-        haiku: 9.9 → APPROVED
-        decision_match: true → QUALIFIED
-```
-
-### PHASE 7b: BIDIRECTIONAL BIAS TEST
-
-```yaml
-bidirectional_bias_test:
-  trigger: "After Phase 7 (Empirical Validation)"
-  duration: "5-10 minutes"
-  condition: "Haiku passed with >95% match"
-
-  rationale: |
-    Haiku can be EITHER conservative OR generous.
-    One test is not enough. Test with different inputs.
-
-  step_7b_1:
-    action: "Run Haiku with DIFFERENT target"
-    inputs:
-      - "Use a SMALLER target (should score higher)"
-      - "Use a LARGER target (should score lower)"
-      - "Use a EDGE CASE target"
-
-  step_7b_2:
-    action: "Compare all results"
-    analysis:
-      - "Is Haiku consistently conservative?"
-      - "Is Haiku consistently generous?"
-      - "Is Haiku inconsistent (varies)?"
-
-  verdicts:
-    consistent_conservative: "ADD Scoring Calibration (bias up)"
-    consistent_generous: "ADD Scoring Calibration (bias down)"
-    inconsistent: "VETO - Task needs more structure"
-
-  quality_gate:
-    qualified_only_if: "Haiku behavior is CONSISTENT across 3+ tests"
-```
-
----
-
-## Script-First Priority Rule
-
-```yaml
-script_first_rule:
-  principle: "SE 90%+ é determinístico → SCRIPT-ONLY, não Hybrid"
-
-  decision_tree:
-    if_deterministic_pct >= 90:
-      action: "Create SCRIPT-ONLY executor"
-      llm_role: "Optional: gap analysis / recommendations"
-      cost: "~$0 per execution"
-
-    elif_deterministic_pct >= 60:
-      action: "Create HYBRID executor"
-      llm_role: "Interpretation of script results"
-      cost: "~20% of full Agent"
-
-    else:
-      action: "Keep as AGENT"
-      cost: "100%"
-
-  examples:
-    an-fidelity-score:
-      deterministic: 95%
-      decision: "SCRIPT-ONLY"
-      script: "fidelity-score.sh"
-      llm_for: "gap recommendations"
-
-    an-clone-review:
-      deterministic: 90%
-      decision: "SCRIPT-ONLY"
-      script: "clone-review.sh"
-      llm_for: "priority actions"
-
-    an-validate-clone:
-      deterministic: 70%
-      decision: "HYBRID"
-      script: "fidelity-score.sh + hackability checks"
-      llm_for: "hackability tests, authenticity markers"
-```
-
----
-
-## Updated Phase Flow
-
-```yaml
-optimize_v4_phases:
-  scan_mode:
-    - PHASE 0: Target Identification
-    - PHASE 1: Determinism Analysis
-    - PHASE 1b: Scope Clarification (NEW)
-    - PHASE 1c: Gatekeeper Detection (NEW)
-    - PHASE 2: ROI Calculation
-    - PHASE 3: Report Generation
-
-  implement_mode:
-    - All scan phases +
-    - PHASE 4: Implementation Suggestions
-    - PHASE 4b: Binary Checkpoint Conversion (NEW)
-
-  hybrid_mode:
-    - All implement phases +
-    - PHASE 5: Script-First Executor (UPDATED)
-    - PHASE 6: GAP ZERO Enforcement
-    - PHASE 7: Empirical Validation
-    - PHASE 7b: Bidirectional Bias Test (NEW)
-
-  post_mode:
-    - PHASE 8: Post-Refactoring Economy Analysis
-```
-
----
-
-## Changelog v4.0 Proposal
-
-| Change | Rationale | Evidence |
-|--------|-----------|----------|
-| PHASE 1b: Scope Clarification | Haiku interprets scope literally | an-clone-review 78.6% match |
-| PHASE 1c: Gatekeeper Detection | Pass/fail tasks are Haiku-safe | qa-after-creation qualified |
-| PHASE 4b: Binary Checkpoint Conversion | 99.3% match with binary | an-fidelity-score v2.0 |
-| PHASE 7b: Bidirectional Bias Test | Haiku can be conservative OR generous | an-clone-review vs an-validate-clone |
-| Script-First Priority | 95% deterministic = script-only | fidelity-score.sh, clone-review.sh |
-
----
-
-## Related Files
-
-- `test-cases/BATCH-PROGRESS.md` - Session evidence
-- `scripts/fidelity-score.sh` - Script-only executor example
-- `scripts/clone-review.sh` - Script-only executor example
-- `tasks/an-fidelity-score.md` - Binary checkpoint conversion example
-- `config/model-routing.yaml` - Tier decisions with evidence