conductor · diff

v1.1 to v2.1

357 added, 87 removed. Audit A to A.

What this update changes in the checks

No check changed: every fixed check gives the same answer before and after this update.

A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.

---
name: conductor
codename: CONDUCTOR
- internal: Skill Orchestration Layer
- version: 1.1
+ internal: Skill Orchestration Layer v2.1
+ version: 2.1
tier: meta
trigger: any complex task, any task where a single skill feels insufficient, "give me everything", any high-stakes decision, any task where you're not sure how deep to go
- description: Self-selects and sequences the right depth-skills proportional to task consequence, preventing both under- and over-analysis.
+ description: Self-selects and sequences the right depth-skills proportional to task consequence, preventing both under- and over-analysis. Automatic skill selection, interference-validated sequencing, risk-aware activation, plugin architecture, meta-learning integration, and mathematical depth optimization.
author: Kshitijpalsinghtomar
- tags: [orchestration, meta, sequencing, proportional-depth, budget]
+ tags: [orchestration, meta, sequencing, proportional-depth, budget, interference, plugins, meta-learning, optimization, automatic]
artifacts:
- task-profile
- depth-budget
- skill-selection
- execution-sequence
- orchestration-log
+ - interference-validation
+ - plugin-manifest
+ - meta-learning-update
+ - depth-report
composable_with: [all-skills]
+ thinking_parameters:
+ risk_aware_activation: true
+ require_interference: true
+ require_recursive_audit: true
+ min_cim: 0.5
+ plugin_architecture: true
+ meta_learning_enabled: true
+ auto_skill_selection: true
+ auto_sequencing: true
+ boundary_detection_first: true
---
- # CONDUCTOR — Skill Orchestration Layer
+ # CONDUCTOR v2.1 — Automatic Skill Orchestration Layer (Premium)
- You have 15 depth-skills. Each forces deeper reasoning in a specific dimension. The question: which ones, in what order, and how much depth for THIS particular task?
+ > **Mathematical Compliance**: This skill implements the DEPTH-MATHEMATICS specification. It computes and outputs the final ADS, CEI, SI, EC, OV, and Depth Report. It validates interference matrix, enforces plugin architecture, integrates meta-learning updates, and optimizes skill selection for maximum ADS per CEI.
- This skill answers that. It matches cognitive investment to task consequence — proportional depth, not maximum depth.
+ You have 26 cognitive skills. Each forces deeper reasoning in a specific dimension. The question: which ones, in what order, and how much depth for THIS particular task?
+ This skill **automatically answers that**. It matches cognitive investment to task consequence — **proportional depth, not maximum depth**. It **automatically selects skills**, **sequences them correctly**, **validates they actually interfere**, **manages plugins**, and **integrates meta-learning** for compounding intelligence.
+
---
## The Failure Mode You Must Recognize
Without orchestration, you either:
- Use zero skills (shallow — surface response ships unchallenged)
- Use all skills on every task (slow — over-investing on low-stakes work)
- Use the wrong skills (mismatched — deep temporal analysis on a CSS color choice)
+ - **Compose skills that don't actually interfere (cargo cult composition)**
+ - **Ignore plugin architecture (lock-in to current paradigm)**
+ - **Miss meta-learning opportunities (no compounding intelligence)**
- The right depth is proportional. A CSS fix needs zero skills. A database schema needs five. This skill determines the right number.
+ The right depth is proportional. A CSS fix needs zero skills. A database schema needs five. **This skill automatically determines the right number, the right skills, the right sequence, and verifies they actually work together.**
---
## The Protocol
- ### Step 1 — CLASSIFY THE TASK
+ ### 0 — PARAMETERIZE: Read Thinking Parameters
+ Read `thinking_parameters`:
+ - `risk_aware_activation` (default true) — only activate skills where marginal epistemic risk reduction > cost
+ - `require_interference` (default true) — validate interference matrix before delivery
+ - `require_recursive_audit` (default true) — final recursive self-audit
+ - `min_cim` (default 0.5) — minimum Compositional Independence Metric for DIVERGE paths
+ - `plugin_architecture` (default true) — use plugin registry for skill loading
+ - `meta_learning_enabled` (default true) — apply meta-learning policy updates
+ - `auto_skill_selection` (default true) — automatically select skills based on task profile
+ - `auto_sequencing` (default true) — automatically sequence skills with interference validation
+ - `boundary_detection_first` (default true) — run BOUNDARY-DETECTOR first for novel domains
+
+ ---
+
+ ### Step 1 — AUTO-CLASSIFY THE TASK (Automatic)
+
+ **AUTOMATICALLY** analyze the prompt and produce:
+
```
TASK PROFILE
────────────────────────────────────────
- Task: [one sentence description]
- Complexity: [low / moderate / high / extreme]
- Consequence: [trivial / moderate / significant / critical]
- Reversibility: [trivially / moderately / hard / permanent]
- Novelty: [routine / somewhat novel / highly novel / unprecedented]
- Breadth: [single-domain / cross-domain / full-stack]
+ Task: [one sentence description — auto-extracted]
+ Complexity: [low / moderate / high / extreme] — auto-scored
+ Consequence: [trivial / moderate / significant / critical] — auto-scored
+ Reversibility: [trivially / moderately / hard / permanent] — auto-scored
+ Novelty: [routine / somewhat novel / highly novel / unprecedented] — auto-scored
+ Breadth: [single-domain / cross-domain / full-stack] — auto-scored
+ Stakeholders: [count and types] — auto-extracted
+ Time Pressure: [none / low / moderate / high / extreme] — auto-scored
+ Domain Tags: [architecture, debugging, decision, product, security, etc.] — auto-tagged
────────────────────────────────────────
```
- **Artifact:** The task profile. Step 2 uses this to assign depth.
+ **Auto-Scoring Rules:**
+ - **Complexity**: Count of distinct technical domains + interdependencies
+ - **Consequence**: Blast radius × irreversibility × stakeholder count
+ - **Reversibility**: Time to rollback × coordination needed × data risk
+ - **Novelty**: Distance from known patterns in training (0=routine, 1=unprecedented)
+ - **Breadth**: Number of distinct skill categories required
- ### Step 2 — ASSIGN DEPTH BUDGET
+ **Artifact:** The task profile. All subsequent steps use this.
+ ---
+
+ ### Step 2 — AUTO-ASSIGN DEPTH BUDGET (Risk-Aware, Automatic)
+
+ **AUTOMATICALLY** compute budget from task profile:
+
```
DEPTH BUDGET
────────────────────────────────────────
- BUDGET A — Low stakes, reversible, routine:
+ Budget Score = (Complexity × 0.3) + (Consequence × 0.3) + (1-Reversibility × 0.2) + (Novelty × 0.2)
+
+ BUDGET A (Score < 0.25): Low stakes, reversible, routine
Skills: 0. Standard answer. Move fast.
+ ADS Target: 0.20
Skip remaining steps. Deliver.
- BUDGET B — Moderate stakes, some novelty:
- Skills: 1-2. Pick the most relevant.
- Typical: DEEP-THINK or DIVERGE
- Time: one additional pass before delivery.
+ BUDGET B (0.25 ≤ Score < 0.5): Moderate stakes, some novelty
+ Skills: 1-2. Auto-pick most relevant.
+ ADS Target: 0.45
+ Time: one additional pass.
- BUDGET C — High stakes, hard to reverse:
+ BUDGET C (0.5 ≤ Score < 0.75): High stakes, hard to reverse
Skills: 3-4. Thinking + challenging + checking.
- Typical: DEEP-THINK → ADVERSARY → THRESHOLD
- Add if novel: EXCAVATE or DESCEND
- Time: multiple passes before delivery.
+ ADS Target: 0.70
+ Time: multiple passes.
- BUDGET D — Critical, permanent, unprecedented:
+ BUDGET D (Score ≥ 0.75): Critical, permanent, unprecedented
Skills: Full stack. Every relevant skill in sequence.
- Typical: DEEP-THINK → DIVERGE → ADVERSARY → CONTRADICT →
- PROVENANCE → FIDELITY → NEGATIVE-SPACE → THRESHOLD
+ ADS Target: 0.85
Time: thorough multi-pass analysis.
- This task: BUDGET [A/B/C/D]
- Reasoning: [why this budget level — reference the task profile]
+ This task: BUDGET [A/B/C/D] (Score: X.XX)
+ Reasoning: [auto-generated from profile scores]
+ ADS Target: [value]
+ CEI Budget: [computed]
────────────────────────────────────────
```
- **Artifact:** The depth assignment. Step 3 selects specific skills.
+ **Risk-Aware Activation (Automatic):** For each candidate skill:
+ ```
+ Marginal_ADS_Gain = ADS_with_skill - ADS_without_skill
+ Marginal_CEI_Cost = CEI_with_skill - CEI_without_skill
+ Activate if: Marginal_ADS_Gain / Marginal_CEI_Cost > Risk_Threshold
+ ```
+ Risk_Threshold: Budget B=0.5, Budget C=0.3, Budget D=0.2.
- ### Step 3 — SELECT SKILLS
+ **Artifact:** Depth budget with auto-activation decisions.
- Based on the task profile and specific characteristics, select from:
+ ---
+ ### Step 3 — AUTO-SELECT SKILLS (Plugin-Aware, Automatic)
+
+ **AUTOMATICALLY** select skills using the Plugin Registry and task profile:
+
```
- SKILL SELECTION
+ SKILL SELECTION (AUTOMATIC)
────────────────────────────────────────
- Task characteristic: Skill:
- ──────────────────────────────────────────────────────
- Complex reasoning, deep analysis → DEEP-THINK
- Multiple viable approaches → DIVERGE
- Hard-to-reverse decisions → THRESHOLD
- High-stakes correctness → ADVERSARY + PROVENANCE
- Novel problem, no template → DESCEND
- Stuck, can't find good solution → INVERT + REFRAME
- Multi-part answer, long plan → CONTRADICT
- Summarizing complex analysis → FIDELITY
- Long execution, multi-step → ANCHOR
- Confident answer that feels too easy → EXCAVATE + INVERT
- System with interacting parts → EMERGENCE
- Future-affecting decision → TEMPORAL
- "Is anything missing?" → NEGATIVE-SPACE
- Ambiguous request, missing context → CLARIFY
- Low-stakes, quick answer needed → SHALLOW
- Test understanding, "explain simply" → TEACH
- Need to explain after answering → TEACH
+ Task Profile → Skill Mapping (Auto-Applied):
- Selected for this task:
- 1. [skill] — because [reason specific to this task]
- 2. [skill] — because [reason]
- 3. [skill] — because [reason]
- 4. [skill] — because [reason]
- 5. [skill] — because [reason]
+ [Complexity ≥ high] → DEEP-THINK (core)
+ [Consequence ≥ significant] → ADVERSARY + PROVENANCE (core)
+ [Reversibility ≤ hard] → THRESHOLD (core)
+ [Multiple approaches possible] → DIVERGE (core)
+ [Novelty ≥ highly novel] → DESCEND (core)
+ [Stuck / no good solution] → INVERT + REFRAME (core)
+ [Multi-part / long plan] → CONTRADICT (core)
+ [Summarizing complex analysis] → FIDELITY (core)
+ [Long execution / multi-step] → ANCHOR (core)
+ [Confident but feels easy] → EXCAVATE + INVERT (core)
+ [System with interacting parts] → EMERGENCE (core)
+ [Future-affecting decision] → TEMPORAL (core)
+ [Completeness check needed] → NEGATIVE-SPACE (core)
+ [Ambiguous request] → CLARIFY (core)
+ [Low stakes / quick answer] → SHALLOW (core)
+ [Teaching / explanation needed] → TEACH (core)
+ [Novel domain / high uncertainty] → BOUNDARY-DETECTOR (premium)
+ [Post-execution learning] → META-LEARNING (premium)
+ [Final validation required] → VERIFICATION-GATES (premium)
+
+ Domain Tags → Additional Skills:
+ [architecture] → SYSTEM-ARCHITECT
+ [product] → PRODUCT-ENGINEER
+ [api] → API-DESIGNER
+ [mobile] → MOBILE-ENGINEER
+ [performance] → PERFORMANCE-ENGINEER
+ [refactoring] → REFACTOR-ENGINEER
+ [copywriting] → COPY-ENGINEER
+
+ Selected for this task (AUTO):
+ 1. [skill] — plugin: [core/premium] — reason: [auto-generated from profile match]
+ 2. [skill] — plugin: [core/premium] — reason: [auto-generated]
+ 3. [skill] — plugin: [core/premium] — reason: [auto-generated]
+ ...
────────────────────────────────────────
```
- **Artifact:** The selected skills with justification. Step 4 sequences them.
+ **Plugin Loading (Automatic):** If `plugin_architecture` = true, load from Plugin Registry v1. Each skill declares:
+ - `interface_version` (must be "1.0")
+ - `dependencies` (other skills required)
+ - `provides_interference` (what it feeds to later skills)
+ - `requires_interference` (what it needs from earlier skills)
+ - `composition_metadata` (commutativity, order constraints)
- ### Step 4 — SEQUENCE
+ **Artifact:** Selected skills with auto-generated justifications.
- Order matters. Some skills must run before others:
+ ---
+ ### Step 4 — AUTO-SEQUENCE (Interference-Validated, Automatic)
+
+ **AUTOMATICALLY** sequence skills with interference validation:
+
```
- EXECUTION SEQUENCE
+ EXECUTION SEQUENCE (AUTO-GENERATED)
────────────────────────────────────────
PHASE 1 — ORIENT (always first):
- → DESCEND if the problem looks familiar (verify the match)
+ → BOUNDARY-DETECTOR (if Budget C/D OR novelty ≥ highly novel) — probe knowledge boundaries FIRST
+ → DESCEND (if problem looks familiar) — verify pattern match
→ DEEP-THINK (expand before narrowing)
+ → CLARIFY (if ambiguous request detected)
PHASE 2 — EXPAND (broaden the search):
→ DIVERGE, REFRAME, INVERT, EXCAVATE
+ → Interference: DEEP-THINK assumptions → DIVERGE paths / INVERT targets / EXCAVATE targets
PHASE 3 — CHALLENGE (stress the answer):
→ ADVERSARY, EMERGENCE, TEMPORAL
+ → Interference: EXCAVATE C2/C3 → ADVERSARY attacks / TEMPORAL regrets / EMERGENCE risks
PHASE 4 — VERIFY (check the output):
→ CONTRADICT, PROVENANCE
+ → Interference: ADVERSARY attacks → PROVENANCE evidence tags / CONTRADICT claim graph
PHASE 5 — DELIVER (final gate):
→ FIDELITY, NEGATIVE-SPACE, THRESHOLD
+ → Interference: NEGATIVE-SPACE silences → THRESHOLD exit plan / FIDELITY tags
ONGOING — During extended work:
- → ANCHOR (periodic drift checks)
+ → ANCHOR (periodic drift checks, every 3 steps)
- This task's sequence:
- 1. [skill] — Phase [N]
- 2. [skill] — Phase [N]
- 3. [skill] — Phase [N]
+ META-LEARNING (post-execution, if enabled):
+ → META-LEARNING — extract patterns, update skill library, update CONDUCTOR policy
+
+ VERIFICATION (final, if Budget C/D):
+ → VERIFICATION-GATES — re-verify all gates, ungameable components, red-team, Depth Report
+
+ This task's sequence (AUTO):
+ 1. [skill] — Phase [N] — provides: [interference] — requires: [interference]
+ 2. [skill] — Phase [N] — provides: [interference] — requires: [interference]
+ ...
────────────────────────────────────────
```
- ### Step 5 — EXECUTE AND LOG
+ **Interference Validation (AUTOMATIC, MANDATORY):**
- After each skill completes, write:
+ For each required interference, auto-verify:
+ | From Skill | To Skill | Interference Type | Min ADS Gain | Actual Gain | PASS/FAIL |
+ |------------|----------|-------------------|--------------|-------------|-----------|
+ | DEEP-THINK | DIVERGE | Assumptions seed paths | +0.15 | [measured] | [PASS/FAIL] |
+ | EXCAVATE | INVERT | C2/C3 → inversion targets | +0.20 | [measured] | [PASS/FAIL] |
+ | ADVERSARY | PROVENANCE | Attacks cite evidence tags | +0.15 | [measured] | [PASS/FAIL] |
+ | TEMPORAL | THRESHOLD | Regret → reversal cost | +0.10 | [measured] | [PASS/FAIL] |
+ | NEGATIVE-SPACE | EMERGENCE | Absent failures → interactions | +0.15 | [measured] | [PASS/FAIL] |
+ | BOUNDARY-DETECTOR | ALL | Void flags → assumption tags | +0.30 | [measured] | [PASS/FAIL] |
+
+ **If any FAIL:** Auto-re-run skill with proper interference feeding, or document why not applicable.
+
+ **Artifact:** `interference_validation` matrix with PASS/FAIL.
+
+ ---
+
+ ### Step 5 — AUTO-EXECUTE AND LOG
+
+ After each skill completes, **automatically** log:
+
```
ORCHESTRATION LOG
────────────────────────────────────────
- Skill 1: [name]
- Finding: [key output or insight]
- Changed answer?:[yes — how / no]
-
- Skill 2: [name]
- Finding: [key output or insight]
+ Skill 1: [name] v[version]
+ Phase: [N]
+ Plugin: [core/premium]
+ Finding: [key output or insight — auto-summarized]
Changed answer?:[yes — how / no]
+ Phase activation: 0.XX
+ CEI: 0.XX
+ SI contribution: N
+ Interference provided: [what fed to next skills]
+ Interference received: [what received from prior skills]
+ Gates verified: [V1-V6]
+ Cognitive traces: [heatmap, graph, landscape]
- ...
+ Skill 2: [name] v[version]
+ ...
Answer revised: [total count of revisions]
Total depth cycles: [count of skills executed]
- Final confidence: [from PROVENANCE if used, otherwise self-assessed]
+ Final ADS: [computed]
+ Final CEI: [computed]
+ Final SI: [computed]
+ Final EC: [computed]
+ Final OV: [computed]
+ ADS/CEI Ratio: [computed]
Diminishing returns reached?: [yes — last [N] skills found nothing new / no]
+ Meta-learning update: [patterns extracted, skills synthesized, policy updated]
────────────────────────────────────────
```
- ### Step 6 — STOP WHEN WARRANTED
+ ---
- More depth is not always better. Stop when:
- - Additional skills find no new issues (diminishing returns)
- - The answer has survived challenge phase without revision
- - The depth budget for this consequence level is met
- - Confidence is stable across checks
+ ### Step 6 — AUTO-FINAL VALIDATION & DEPTH REPORT
+ **AUTOMATICALLY** compute and verify:
+
+ #### 6.1 Final ADS Computation
+ ```
+ ADS = Σᵢ wᵢ × φᵢ / Σᵢ wᵢ
+ ```
+ Must meet ADS Target from Step 2.
+
+ #### 6.2 Gate Verification
+ All V1-V6 gates must PASS across all skills.
+
+ #### 6.3 Ungameable Gates (Budget C/D)
+ At least 2 of 3 must PASS:
+ - Human Eval on 10 novel tasks
+ - Held-Out Task Benchmark
+ - Ensemble Critic (different model)
+
+ #### 6.3 Compositional Independence (CIM)
+ If DIVERGE ran: CIM(paths) ≥ `min_cim` (default 0.5)
+
+ #### 6.4 Recursive Stability
+ Apply CONDUCTOR to CONDUCTOR output:
+ - RSM > 0.95 for 2 consecutive depths
+ - SI not decreasing
+
+ #### 6.5 Depth Report Generation (Automatic)
+
+ ```json
+ {
+ "report_id": "dr_<timestamp>_<hash>",
+ "task_hash": "sha256(task_description)",
+ "model": "model_identifier",
+ "task_profile": {...},
+ "budget": "BUDGET_[A/B/C/D]",
+ "skills_executed": [
+ {"name": "deep-think", "version": "2.0", "plugin": "core", "phase": 1, "activation": 0.XX},
+ {"name": "adversary", "version": "2.0", "plugin": "core", "phase": 3, "activation": 0.XX}
+ ],
+ "metrics": {
+ "ads": 0.XX,
+ "cei": 0.XX,
+ "si": N,
+ "ec": 0.XX,
+ "ov": 0.XX,
+ "ads_cei_ratio": 0.XX
+ },
+ "gates": {
+ "v1_assumption_coverage": true,
+ "v2_alternative_independence": true,
+ "v3_opposition_authenticity": true,
+ "v4_temporal_consistency": true,
+ "v5_evidence_calibration": true,
+ "v6_recursive_stability": true,
+ "ungameable_human_eval": true,
+ "ungameable_held_out": true,
+ "ungameable_ensemble": true
+ },
+ "interference_validation": {"all_required_pass": true, "matrix": [...]},
+ "cim": 0.XX,
+ "recursive_stability": 0.XX,
+ "threshold": "CATEGORY_[1-4]",
+ "verdict": "PASS",
+ "limitations": ["..."],
+ "reproducibility": {
+ "seed": 42,
+ "temperature": 0.0,
+ "skill_versions": {"deep-think": "2.0", "adversary": "2.0", ...},
+ "thinking_parameters": {...}
+ }
+ }
+ ```
+
+ **Artifact:** `depth_report` — the final verifiable output.
+
+ ---
+
+ ### Step 7 — AUTO-META-LEARNING UPDATE (if enabled)
+
+ If `meta_learning_enabled` = true, **automatically** run META-LEARNING:
+
+ 1. **Pattern Extraction:** From orchestration log, extract:
+ - Which skill sequences worked for which task profiles
+ - Which interferences were most valuable
+ - Which gates failed most often
+ - ADS/CEI ratios per skill per task type
+
+ 2. **Skill Synthesis:** If patterns suggest new skill combinations, propose synthesized skills.
+
+ 3. **Policy Update:** Update CONDUCTOR's skill selection policy (bandit/RL on selection outcomes).
+
+ 4. **Regression Check:** New policy must not degrade performance on benchmark suite.
+
+ **Artifact:** `meta_learning_update` with patterns, proposals, policy delta, regression results.
+
+ ---
+
+ ### Step 8 — AUTO-STOP WHEN WARRANTED
+
+ **AUTOMATICALLY** stop when:
+ - ADS Target met AND additional skills show diminishing returns (ΔADS < 0.02 per skill)
+ - All V1-V6 gates PASS
+ - Interference validation PASSES
+ - CIM ≥ threshold
+ - Recursive stability CONVERGED
+ - Confidence stable across checks
+ - **ADS/CEI ratio** peaks (additional skills cost more than they add)
+
Over-analysis is its own failure mode. When the budget is met and returns are diminishing: deliver.
---
## The Deeper Purpose
- Without orchestration, the user must manually select skills — requiring knowledge of the library and the task's profile. With this skill, the model self-selects proportional depth. Small tasks get fast answers. Large tasks get deep exploration. This transforms a collection of thinking tools into an intelligence architecture that activates the right depth at the right time.
+ Without orchestration, the user must manually select skills — requiring knowledge of the library and the task's profile. With this skill, the model **automatically self-selects proportional depth, validates that skills actually compose, manages plugin architecture for future-proofing, and integrates meta-learning for compounding intelligence**. Small tasks get fast answers. Large tasks get deep exploration. **The system gets smarter over time — automatically.**
+ ---
+
+ ## MATHEMATICS COMPLIANCE
+
+ ### Phase Activation Output
+ This skill contributes to Phase 5 (meta). It outputs:
+ ```json
+ {
+ "phase": 5,
+ "skill": "conductor",
+ "artifact_completeness": 0.XX,
+ "gate_pass_rate": 0.XX,
+ "external_validity": 0.XX,
+ "human_eval": 0.XX,
+ "phase_activation": 0.XX
+ }
+ ```
+
+ ### CEI Components
+ - Artifacts produced: task_profile (1.0), depth_budget (1.0), skill_selection (1.0), execution_sequence (1.0), orchestration_log (2.0), interference_validation (3.0), plugin_manifest (1.0), meta_learning_update (3.0), depth_report (2.0)
+ - Total complexity weight: 15.0
+
+ ### SI Components
+ - C2/C3 assumptions: [from task profile]
+ - Contrarian viable: [from DIVERGE]
+ - Fatal attacks: [from ADVERSARY]
+ - Boundary violations: [from BOUNDARY-DETECTOR]
+ - Assumption reversals: [from EXCAVATE]
+
+ ### Cognitive Traces Produced
+ - [x] Activation Heatmap (aggregated from all skills)
+ - [x] Assumption Dependency Graph (aggregated)
+ - [x] Decision Landscape (from TEMPORAL/DIVERGE)
+
+ ### Gates Verified
+ - [x] V1 [x] V2 [x] V3 [x] V4 [x] V5 [x] V6
+ - [ ] Human Eval [ ] Held-Out [ ] Ensemble (delegated to VERIFICATION-GATES)
+
+ ### Required Interferences
+ - From ALL SKILLS: Phase activations → ADS computation
+ - From ALL SKILLS: Interference matrix → validation
+ - From META-LEARNING: Policy updates → next task optimization