Mens Continuous Training, Multi-Spoke Data Flywheel & Evaluation SSoT (2026)

This document defines the canonical architecture for the Mens five-spoke continuous training, evaluation, and audit-to-corpus flywheel across VoxLang, RustWorkspace, StructuredShell, McpToolChain, and AgentHarness. It specifies the TrainingRecordV2 schema and VoxDb Trusty dataset snapshots (CORPUS-01..06), deterministic text-in/text-out round-trip verification oracles (SPOKE-01..06), inverse-pass-rate curriculum weighting and QLoRA plus DPO adapter calibration (TRAIN-01..05), paired multi-spoke promotion gates (EVAL-01..05), and continuous audit-to-DPO feedback ratchets (AUDIT-01..05, LOOP-01..04).

1. Overview & Five-Spoke Architecture

Mens trains and evaluates as a unified multi-spoke coding and orchestration model. Every record admitted into the training corpus, every curriculum batch, and every promotion gate operates across five canonical spokes:

Spoke Enum (SpokeKind / MensEvalSpoke)	Identifier	Deterministic Verification Oracle
VoxLang	vox_lang	vox_compiler lexer, parser, typechecker (check_module), and retired-decorator grammar guard (CORPUS-02, SPOKE-01)
RustWorkspace	rust_workspace	Balanced-delimiter syntax check, item signature extraction, and AGENTS.md policy guard banning direct crypto imports (CORPUS-03, SPOKE-02)
StructuredShell	structured_shell / shell_command	ShellSafetyClassification (Safe, Forbidden, Chained) enforcing single-command discipline and secret redaction (CORPUS-04, SPOKE-03)
McpToolChain	mcp_tool_chain / mcp_tool	JSON schema and tool-registry argument validator (SPOKE-04)
AgentHarness	agent_harness	Multi-step plan and tool-trajectory validator (SPOKE-05)

2. Persistent Multi-Spoke Training Data Schema & VoxDb Snapshots (CORPUS-01..06)

All training records conform to TrainingRecordV2 (schema_version = "vox.training_record.v2") in crates/vox-corpus/src/training/contract.rs:

• SpokeKind: Explicit 5-spoke discriminator (vox_lang, rust_workspace, structured_shell, mcp_tool_chain, agent_harness).
• OracleReceipt: Carries spoke, passed: bool, stage, diagnostics, and exit_code. No record with passed == false is admitted to positive SFT training.
• RecordProvenance: Tracks origin, source_id, task_family, and difficulty.
• Deduplication & Decontamination (CORPUS-05): deduplicate_records_v2 and decontaminate_against_holdout in crates/vox-corpus/src/corpus/eval_split.rs apply per-spoke token-trigram Jaccard near-duplicate filtering and holdout decontamination; the Jaccard threshold is a caller-supplied parameter (max_near_dup_jaccard / max_jaccard), not a fixed constant.
• Immutable VoxDb Snapshots (CORPUS-06): build_trusty_snapshot, persist_snapshot_to_db, and load_snapshot_from_db in crates/vox-corpus/src/dataset_snapshot.rs seal datasets under content-addressed trusty://vox-dataset/<digest> URIs in VoxDb.

3. Text-In / Text-Out Round-Trip Oracles (SPOKE-01..06)

crates/vox-corpus/src/training/spoke_roundtrip.rs provides deterministic round-trip verification (prompt -> candidate -> oracle -> MensSpokeScorecardV1):

• verify_voxlang_roundtrip (SPOKE-01)
• verify_rust_roundtrip / verify_rust_workspace_roundtrip (SPOKE-02)
• verify_shell_roundtrip / verify_structured_shell_roundtrip (SPOKE-03)
• verify_mcp_roundtrip (SPOKE-04)
• verify_agent_harness_roundtrip (SPOKE-05)
• evaluate_multi_spoke_suite & canonical_five_spoke_golden_suite (SPOKE-06), emitting contracts/reports/mens-spoke-scorecard.v1.json.

4. Native QLoRA + DPO Pipeline, Curriculum Weighting & Adapter Calibration (TRAIN-01..05)

• Snapshot Streaming (TRAIN-01): ChosenRejectedPairV2, SftDpoBatch, and stream_snapshot_batches in crates/vox-corpus/src/training/mix_prepare.rs stream token-bounded SFT + DPO batches directly from TrustyDatasetSnapshot URIs.
• Inverse-Pass-Rate Curriculum Weighting (TRAIN-02): SpokeCurriculumWeights and compute_curriculum_weights in crates/vox-corpus/src/training/curriculum.rs up-weight weaker spokes (weight_i ∝ max(floor, 1.0 - pass_at_1_i)) while guaranteeing a minimum 0.10 (DEFAULT_MIN_SPOKE_WEIGHT_FLOOR) sampling floor per spoke to prevent catastrophic forgetting.
• Checkpoint Lineage (TRAIN-03): MensCheckpointManifestV1, persist_checkpoint_to_db, and load_checkpoint_from_db in crates/vox-corpus/src/training/adapter_registry.rs record snapshot_uri, adapter_uri, curriculum_weights, sft_step_count, dpo_step_count, initial_loss, final_loss, loss_monotonic_decreased, and spoke_adapter_gains.
• Model-Registry Calibration (TRAIN-04): calibrate_adapter_from_scorecard and build_adapter_registry_entry register verified Mens adapters (MensAdapterRegistryEntry) with empirical quality_score = overall_pass_at_1 and promotion eligibility gated on zero spoke regressions.
• Preflight Safety Gate (TRAIN-05): validate_snapshot_safety_gate in crates/vox-corpus/src/training/preflight.rs fails closed on empty spokes, unverified SFT records, unredacted secret patterns, or holdout contamination (holdout_overlap_count > 0).

5. Multi-Spoke Empirical Evaluation, Longitudinal Ledger & Promotion Gate (EVAL-01..05)

• 5-Spoke Scorecard (EVAL-01): MensFiveSpokeScorecardV1 and evaluate_mens_five_spokes in crates/vox-eval/src/mens.rs.
• Paired Comparison & Promotion Gate (EVAL-02, EVAL-04): compare_mens_five_spokes, MultiSpokeComparisonV1, and PromotionGateVerdict (Promote, RejectRegression, HoldNoImprovement) in crates/vox-eval/src/compare.rs enforce per-spoke non-inferiority (losses == 0 and delta_pass_at_1 >= -max_spoke_regression) alongside strict overall pass@1 improvement.
• Longitudinal Ledger (EVAL-03): MensLongitudinalScorecardV1 and MensLongitudinalLedger in crates/vox-eval/src/trial_flywheel.rs verify monotonic progression across flywheel cycles.
• Report Rendering (EVAL-05): render_multi_spoke_comparison_markdown and render_multi_spoke_comparison_json in crates/vox-eval/src/report.rs.

6. Continuous Audit-to-Corpus Flywheel & Autonomous Burn-In (AUDIT-01..05, LOOP-01..04)

• Rejected Attempt Harvesting (AUDIT-01): harvest_rejected_attempts in crates/vox-audit/src/subcommands/corpus_feedback.rs.
• Verified Repair DPO Pairing (AUDIT-02): pair_verified_repair_to_dpo_and_sft and pair_verified_repair_with_failure in crates/vox-audit/src/subcommands/repair_corpus.rs.
• Hard-Negative AST Mutation (AUDIT-03): mint_hard_negative_dpo_pair in crates/vox-corpus/src/ast_mutator.rs.
• Per-Spoke Regression Ratchets (AUDIT-04): SpokeRegressionRatchetV1 and verify_spoke_regression_ratchet in crates/vox-audit/src/subcommands/regression_budget.rs, backed by the committed contracts/eval/spoke-regression-ratchet.v1.json contract.
• Behavioral Golden Promotion (AUDIT-05): PromotedSpokeGolden and promote_verified_repair_to_golden in crates/vox-audit/src/subcommands/behavioral_goldens.rs.
• Two-Cycle Autonomous Burn-In (LOOP-01..03): run_two_cycle_autonomous_burnin, persist_burnin_report_to_db, and load_burnin_report_from_db in crates/vox-corpus/src/flywheel.rs execute Cycle 0 -> Cycle 1 -> Cycle 2, proving monotonic overall pass@1 gain (0.40 -> 0.80 -> 1.00), zero spoke regressions, and full VoxDb CAS lineage in contracts/reports/mens-flywheel-burnin.v1.json.

7. High-Performance 25-Problem Five-Spoke Before / After Benchmark (MENS-PERF-01..04)

canonical_five_spoke_benchmark_suite_25 in crates/vox-corpus/src/training/spoke_roundtrip.rs and run_before_after_five_spoke_evaluation (MensBeforeAfterEvaluationV1) in crates/vox-corpus/src/flywheel.rs evaluate 25 realistic problems (5 per spoke) across three calibration stages (contracts/reports/mens-before-after-comparison.v1.json):

Spoke	Cases	Before (Cycle 0) pass@1	Mid (Cycle 1 SFT+DPO+Curriculum) pass@1	After (Cycle 2 Verified Repair+Rebalance) pass@1	Delta pass@1	Before pass@3	After pass@3
vox_lang	5	40.0%	80.0%	100.0%	+60.0%	60.0%	100.0%
rust_workspace	5	20.0%	80.0%	100.0%	+80.0%	40.0%	100.0%
structured_shell	5	20.0%	60.0%	100.0%	+80.0%	40.0%	100.0%
mcp_tool_chain	5	20.0%	80.0%	100.0%	+80.0%	40.0%	100.0%
agent_harness	5	20.0%	60.0%	100.0%	+80.0%	40.0%	100.0%
Overall (25 cases)	25	24.0%	72.0%	100.0%	+76.0%	44.0%	100.0%

8. Axis Harness (vox-orchestrator), CLI (vox-cli / vox-ml-cli), and Axis GUI (vox-gui) Wiring (MENS-HARNESS-01..04, MENS-SURFACE-01..04)

1. Axis Model Registry & Spoke Routing (crates/vox-orchestrator/src/models/registry.rs, scoring.rs, catalog.rs):
   • ProviderType::VoxLocal and ProviderType::PopuliMesh receive a first-class 0.92 local latency score fallback in latency_score.
   • MensCatalog::refresh and ModelRegistry::register_mens_local_candidates populate full 5-spoke StrengthTags (Generalist, Codegen, UiCodegen, Parsing, Debugging, Review, Security, Logic, InterAgent, Planning) and ModelCapabilities (supports_json, supports_jsonl, supports_tool_use, supports_native_tools, supports_reasoning, writes_vox, latency_p50_ms: Some(140)).
   • ModelRegistry::register_calibrated_mens_adapter ingests MensAdapterRegistryEntry, populates BenchEvidence (ModelConfidence::Confirmed), ModelScore, and arm_stats, and ModelRegistry::route_spoke_task(spoke) routes all five SpokeKind tasks to the calibrated Mens adapter.
2. CLI Surfaces (crates/vox-cli/src/commands/model/eval.rs, crates/vox-ml-cli/src/commands/mens/populi/mens_tail_subcommands.rs):
   • vox model eval --mens-five-spokes [--flywheel-burnin] [--write-contracts] executes the 25-problem 5-spoke before/after evaluation, verifies Axis ModelRegistry routing across all 5 spokes, and renders the Markdown delta table.
   • vox mens eval-gate --five-spokes runs the 5-spoke promotion gate and writes mens_spoke_scorecard.json.
3. Axis GUI Surface (crates/vox-gui/src/commands/mens_run_reports.rs, MensTrainingView.tsx):
   • mens_run_reports returns spoke_scorecard: Option<MensSpokeScorecardDto> (including "workspace" contract loading from contracts/reports/mens-before-after-comparison.v1.json).
   • MensTrainingView.tsx provides the Load 5-Spoke Benchmark action and renders the data-testid="mens-spoke-scorecard" before/after pass@1, delta, and curriculum weight table.