Autonomous Telemetry-Driven Routing Learning, Arena Comparison & Axis Harness v3.3

Vox Milestone v3.3 closes the loop between centralized telemetry intelligence, interactive human preference evaluation, and workspace harness ergonomics across vox-orchestrator, vox-orchestrator-mcp, and vox-gui. It introduces signed remote Community Intelligence Bundle ingestion (Ed25519 + Empirical Bayes $k \ge 20, n \ge 50$), Side-by-Side Arena turn comparison with online Bradley-Terry updates and DPO tuple harvesting, turn-level conversation branching paired with Jujutsu (vox-vcs) checkpoint rewind, zero-LLM-cost sub-transcript tool-result clearing, and deterministic .vox lifecycle hooks with Oracle tier escalation.

1. Closed-Loop Centralized Telemetry & Model Routing Auto-Improvement (ROUTE-LEARN-01..04)

1.1 Remote Community Intelligence Sync (intelligence_bundle.rs)

• Endpoint & Secret Resolution: sync_community_intelligence_from_remote in crates/vox-orchestrator/src/models/intelligence_bundle.rs resolves the upload endpoint via vox_secrets::resolve_secret(SecretId::VoxTelemetryUploadUrl) (with optional SecretId::VoxTelemetryUploadToken Bearer auth), derives /v1/intelligence/bundle.json, and fetches the signed bundle over vox_http_client::client_builder().
• Cryptographic & Privacy Verification: Before persisting to ~/.vox/cache/community-intelligence.v1.json, CommunityIntelligenceBundle::verify enforces:
  1. Valid Ed25519 signature over the canonical signing payload via vox_crypto::signing::PublicKey.
  2. Strict $k$-anonymity ($k \ge 20$ distinct installations) and minimum sample size ($n \ge 50$) on every ModelTaskPrior and ModelBenchSpokePrior cell.

1.2 Empirical Bayes Priors & Scoreboard Feedback (scoring.rs & selection_queries.rs)

• Spoke-Specific Priors: ModelRegistry::apply_community_intelligence_bundle blends Empirical Bayes model_bench_spoke_priors into each candidate’s ModelScore::quality_score and seeds cold-start contextual bandit arm_stats pseudo-counts (n_calls == 0) so Thompson sampling starts from community-verified priors.
• Graded Quality Feedback Boost: scoreboard_feedback_boost in crates/vox-orchestrator/src/models/scoring.rs incorporates graded quality_score values in (0.0, 1.0) (0.6 * measured + 0.4 * arena_quality) while keeping ungraded 0.0/1.0 defaults neutral.
• Automatic Refresh: Both ModelRegistry::inject_evidence_from_db and Orchestrator::refresh_model_scoreboard load and apply the verified cache bundle automatically.

---

2. Side-by-Side Arena Turn Comparison & DPO Preference Learning (ARENA-COMPARE-01..04)

• Dual-Arm Turn Dispatch: chat_arena_compare (crates/vox-gui/src/commands/chat_arena_branch.rs) dispatches a prompt to Arm A (primary auto-routed model) and Arm B (challenger model, defaulting to the local 5-spoke MENS model mens/qwen3.5-27b:five-spoke-qlora-r2), returning ArenaTurnComparisonDto.
• Online Bradley-Terry & Bandit Updates: ModelRegistry::record_arena_preference_outcome and chat_arena_vote apply an online Bradley-Terry step ($K = 0.12$) to quality_a / quality_b (clamped to [0.05, 0.95]), update contextual bandit (successes, failures) posteriors, persist the winning reply into the session transcript, and record a DPO preference tuple (prompt, chosen, rejected).
• GUI Surface: ArenaComparisonCard.tsx and ChatTranscript.tsx render a toggleable split view (data-testid="chat-arena-split-card") with 1-click voting buttons (Prefer A, Prefer B, Tie / Both Good).

---

3. Turn-Level Conversation Branching & Jujutsu Checkpoint Rewind (BRANCH-REWIND-01..04)

• Non-Destructive Session Forking: chat_fork_session (fork_session_impl) clones the message prefix up to fork_from_message_id into a new ChatSessionDto with parent_session_id and fork_from_message_id lineage metadata preserved.
• Deterministic Turn Rewind: chat_rewind_to_turn (rewind_to_turn_impl) rewinds the session cursor to target_message_id and restores or synthesizes the turn’s vcs_checkpoint_op_id (vox-vcs / Jujutsu operation ID).
• GUI Controls: Every user/assistant MessageBubble in ChatTranscript.tsx exposes ⑂ Branch here (data-testid="chat-turn-fork-btn-${id}") and ↺ Rewind (data-testid="chat-turn-rewind-btn-${id}"), with active lineage displayed in data-testid="chat-session-branch-badge".

---

4. Sub-Transcript Surgical Tool-Result Clearing (CONTEXT-SURGERY-01..04)

• Zero-LLM-Cost Compaction: clear_stale_tool_results_in_messages in crates/vox-orchestrator-mcp/src/chat_tools/chat/conversation.rs walks the conversation history and replaces stale role == "tool" or verbose JSON tool outputs older than keep_recent_tool_outputs with [cleared to save context — N BPE tokens reclaimed], leaving all user prompts, assistant reasoning, and recent tool outputs intact.
• Interactive Context Meter Control: ContextWindowMeter.tsx and ChatExecutionRail.tsx expose a 1-click Clear Tools button (data-testid="context-meter-clear-tools-btn") wired to clear_session_tool_results, displaying cumulative reclaimed BPE tokens.

---

5. Deterministic .vox Lifecycle Hooks & Oracle Model Escalation (HOOKS-ORACLE-01..04)

• Out-of-Context Lifecycle Guardrails: crates/vox-orchestrator-mcp/src/chat_tools/chat/agent_loop/lifecycle_hooks.rs executes deterministic PreToolUse, PostToolBatch, and Stop hooks inside dispatch_tool_calls_batch with zero LLM token overhead:
  • Blocks read/write access to .env* secret files (except .env.example).
  • Blocks direct manual edits to generated SSOT artifacts (docs/src/*.generated.md, contracts/reports/gui-surface-*.v1.json, .cursorignore, .aiignore, .aiexclude).
  • Automatically rewrites cargo fmt --all invocations to vox run scripts/fmt.vox -- --all.
  • Injects a post-mutation verification reminder after file edits and blocks empty or unverified stop states.
• Oracle Tier Escalation: consult_oracle_model_advisory resolves the highest-precision reasoning model (SelectionAxes::QUALITY_FIRST, complexity 9) via vox_orchestrator::models::select to provide structured architectural and root-cause verification advisories (OracleConsultationAdvisory), surfaced in ChatExecutionRail.tsx (data-testid="execution-rail-hooks-status").