{ "artifacts": { "adapter_dir": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/adapter", "best_checkpoint_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/best_checkpoint_manifest.json", "epoch_history": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/epoch_history.json", "eval_jsonl": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/rollback_eval.jsonl", "host_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/host_manifest.json", "plan_json": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/rollback_adapter_plan.json", "rollback_head": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/rollback_head.pt", "tokenizer_dir": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/tokenizer", "train_jsonl": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/rollback_train.jsonl", "training_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/rollback_qwen36_rootfix_run1/rollback_training_manifest.json" }, "dataset": { "cause_vocab": [ "none", "evidence_contradiction", "execution_negation", "validator_rejection", "patch_regression", "worktree_conflict", "high_risk_unsupported", "tool_selfcheck_failure", "reverse_engineering_speculation" ], "eval_size": 45, "mode_vocab": [ "contradiction", "execution_failure", "validator_rejection", "patch_regression", "worktree_conflict", "high_risk_unsupported", "tool_selfcheck_repair", "reverse_engineering_drift", "validated_stable" ], "modes": [ "contradiction", "execution_failure", "high_risk_unsupported", "patch_regression", "reverse_engineering_drift", "tool_selfcheck_repair", "validated_stable", "validator_rejection", "worktree_conflict" ], "need_vocab": [ "none", "advisory", "required" ], "rewrite_vocab": [ "do_not_rollback", "local_patch_only", "validator_scoped_revision", "minimum_diff_patch", "do_not_overwrite_worktree_state", "fail_closed", "repair_via_bounded_selfcheck_loop", "bounded_observation_first", "local_conflict_preserving_revision" ], "train_size": 126 }, "eval_metrics": { "adapter_exported": true, "auto_lora_from_ia3": false, "avg_binary_accuracy": 0.8194444444444444, "best_epoch": 3, "best_quality_score": 0.3986773560841878, "cause_accuracy": 0.4888888888888889, "count": 45, "eval_batches": 45, "eval_loss": 7.656409973568387, "fail_closed_accuracy": 0.7777777777777778, "minimal_diff_accuracy": 0.5555555555555556, "mode_accuracy": 0.3333333333333333, "need_accuracy": 0.7777777777777778, "peft_method": "lora_narrow", "preserve_conflict_accuracy": 0.7777777777777778, "quality_score": 0.3986773560841878, "reverse_engineering_bounded_accuracy": 0.8888888888888888, "rewrite_accuracy": 0.4888888888888889, "selfcheck_route_accuracy": 0.5555555555555556, "session_continuity_accuracy": 1.0, "used_peft": true, "validator_aligned_accuracy": 1.0, "worktree_safe_accuracy": 1.0 }, "plan": { "backbone": "/public/wang_libo/veriloop_coder_e1/model", "dataset_summary": { "cause_vocab": [ "none", "evidence_contradiction", "execution_negation", "validator_rejection", "patch_regression", "worktree_conflict", "high_risk_unsupported", "tool_selfcheck_failure", "reverse_engineering_speculation" ], "eval_size": 45, "mode_vocab": [ "contradiction", "execution_failure", "validator_rejection", "patch_regression", "worktree_conflict", "high_risk_unsupported", "tool_selfcheck_repair", "reverse_engineering_drift", "validated_stable" ], "modes": [ "contradiction", "execution_failure", "high_risk_unsupported", "patch_regression", "reverse_engineering_drift", "tool_selfcheck_repair", "validated_stable", "validator_rejection", "worktree_conflict" ], "need_vocab": [ "none", "advisory", "required" ], "rewrite_vocab": [ "do_not_rollback", "local_patch_only", "validator_scoped_revision", "minimum_diff_patch", "do_not_overwrite_worktree_state", "fail_closed", "repair_via_bounded_selfcheck_loop", "bounded_observation_first", "local_conflict_preserving_revision" ], "train_size": 126 }, "excluded_surfaces": [ "(^|\\.)lm_head($|\\.)::Do not retune final token head; too broad and evaluation-heavy.", "(^|\\.)embed_tokens($|\\.)::Embedding surgery risks broad semantic drift.", "(^|\\.)norm($|\\.)::Global norm tuning can destabilize calibration across scenes.", "attnres|attention_residual::Block AttnRes may be mounted structurally but is never a PEFT target.", "dualpath::DualPath is serving/runtime infrastructure only.", "mhc|hyper[-_]?connection::mHC-inspired stability hooks remain structural, not PEFT surfaces.", "rope|rotary::RoPE/context surgery is handled architecturally, not by narrow PEFT here.", "kvcache|kv_cache::KV-cache runtime surfaces are not PEFT targets.", "(^|\\.)memory(_store|_bank)?($|\\.)::Persistent memory stores are harness/runtime policy surfaces, not PEFT targets." ], "notes": [ "Primary route is host-surface-first rollback training.", "Validator receipts, failure routing, patch state, worktree continuity, self-check repair, and reverse-engineering boundedness are first-class signals.", "DualPath, Block AttnRes, mHC hooks, and MoE routers/experts remain structurally excluded.", "This adapter should improve rollback obedience, not general coding free-formity." ], "peft_method": "lora_narrow", "product_line": "veriloop_coder", "recipe": { "adapter_family": "validator_rollback", "backbone": "/public/wang_libo/veriloop_coder_e1/model", "backbone_family": "qwen_dense", "excluded_patterns": [ "(?i)\\bdualpath\\b", "(?i)\\bmhc\\b", "(?i)\\bfull[_\\- ]?attnres\\b", "(?i)\\battnres(_full)?\\b", "(?i)\\brouter\\b", "(?i)\\bexperts?\\b", "(?i)\\bmoe\\b.*\\b(gate|router|expert)\\b", "(?i)\\brope\\b.*\\b(freq|inv_freq|theta|rotary)\\b", "(?i)\\bkvcache\\b", "(?i)\\bposition_embedding\\b", "(?i)\\bembed(tokens|ding)?\\b", "(?i)\\blm_head\\b" ], "harness_constraints": [ "Harness Engineering remains the primary convergence layer.", "Adapter must not bypass runtime orchestrator / validator / rollback loops.", "Adapter outputs remain subordinate to VeriLoop control-plane decisions.", "Adapter must not create hidden prompt-style memory authority.", "Adapter must preserve validator-first correction and minimal-diff rollback discipline.", "Patch production must remain subordinate to validator receipts and failure routing." ], "hyperparams": { "alpha": 16, "bias": "none", "dropout": 0.08, "fan_in_fan_out": false, "modules_to_save": [ "input_layernorm" ], "r": 8, "task_type": "CAUSAL_LM" }, "merge_policy": "side_load", "metadata": { "allow_backbone_bridge": false, "allow_vla_action_expert": false, "harness_first": true, "prefer_explicit_heads": true, "prefer_qlora_for_backbone_bridge": false, "require_harness_first": true, "reverse_engineering_readiness": true, "rollback_training": true, "selector_group_count": 2, "selfcheck_loop_readiness": true, "strict_narrow_scope": true, "trainer": "veriloop.rollback_adapter_trainer.v9.qwen36" }, "notes": [ "Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op.", "Backbone family inferred as qwen_dense.", "PEFT method resolved as lora_narrow.", "Recipe is harness-first: runtime convergence remains in VeriLoop control-plane + harness, not in broad weight surgery.", "Block AttnRes, DualPath, mHC hooks, RoPE, KV-cache, and broad MoE routing remain structurally excluded." ], "peft_method": "lora_narrow", "precision_policy": "auto", "product_line": "veriloop_coder", "regression_requirements": [ "Must pass PEFT regression guard structural policy checks.", "Must not introduce forbidden backbone/serving structural targets.", "Must preserve harness regression envelope for the selected product line.", "Validator feedback uptake must not regress.", "Rollback minimality and revision fidelity must not regress." ], "target_groups": [ { "alpha": 16, "dropout": 0.05, "name": "group_1_custom_runtime_harness_bridge", "rank": 8, "rationale": "Runtime / harness obedience should attach to explicit interface bridges before any backbone fallback.", "surface": "custom_runtime_harness_bridge", "target_modules": [ "failure_signal_bridge.rollback_bridge", "request_normalizer", "request_normalizer.adapter", "rollback_adapter", "rollback_adapter.head", "runtime_harness_adapter", "runtime_harness_adapter.bridge", "tool_protocol_adapter", "tool_protocol_adapter.bridge", "toolspec_bridge", "toolspec_bridge.adapter", "toolspec_head", "toolspec_head.param_schema_adapter", "toolspec_head.postcondition_adapter", "toolspec_head.precondition_adapter", "toolspec_head.receipt_formatter", "toolspec_head.trigger_gate", "toolspec_head.validator_gate", "validator_feedback_loop.rollback_adapter" ] }, { "alpha": 16, "dropout": 0.0, "name": "group_2_custom_validator_bridge", "rank": 8, "rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.", "surface": "custom_validator_bridge", "target_modules": [ "failure_signal_bridge", "failure_signal_bridge.rollback_bridge", "rollback_adapter", "rollback_adapter.head", "sandbox_rollback_bridge", "sandbox_rollback_bridge.adapter", "validator_feedback_bridge", "validator_feedback_bridge.adapter", "validator_feedback_loop.rollback_adapter" ] } ], "target_modules": [ "failure_signal_bridge.rollback_bridge", "request_normalizer", "request_normalizer.adapter", "rollback_adapter", "rollback_adapter.head", "runtime_harness_adapter", "runtime_harness_adapter.bridge", "tool_protocol_adapter", "tool_protocol_adapter.bridge", "toolspec_bridge", "toolspec_bridge.adapter", "toolspec_head", "toolspec_head.param_schema_adapter", "toolspec_head.postcondition_adapter", "toolspec_head.precondition_adapter", "toolspec_head.receipt_formatter", "toolspec_head.trigger_gate", "toolspec_head.validator_gate", "validator_feedback_loop.rollback_adapter", "failure_signal_bridge", "sandbox_rollback_bridge", "sandbox_rollback_bridge.adapter", "validator_feedback_bridge", "validator_feedback_bridge.adapter" ], "version": "veriloop.lora_recipe_veriloop.v2" }, "selected_surfaces": [ "custom_runtime_harness_bridge", "custom_validator_bridge", "custom_memory_boundary_bridge" ], "selected_target_modules": [ "failure_signal_bridge.rollback_bridge", "rollback_adapter.head", "runtime_harness_adapter.bridge", "sandbox_rollback_bridge.adapter", "toolspec_head.validator_gate", "validator_feedback_bridge.adapter", "validator_feedback_loop.rollback_adapter", "memory_boundary_guard.adapter", "memory_boundary_guard.rollback_filter", "request_normalizer.adapter", "tool_protocol_adapter.bridge", "toolspec_bridge.adapter", "toolspec_head.postcondition_adapter", "toolspec_head.precondition_adapter", "toolspec_head.receipt_formatter", "toolspec_head.trigger_gate" ], "selection_mode": "minimal", "version": "veriloop.rollback_adapter_trainer.v9.qwen36", "warnings": [ "Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.", "Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op." ] }, "status": "trained", "train_metrics": { "adapter_exported": true, "auto_lora_from_ia3": false, "best_epoch": 3, "best_quality_score": 0.3986773560841878, "epochs_completed": 4, "loss": 0.4452953128114579, "micro_batches": 126, "micro_batches_total": 504, "optimizer_steps": 8, "optimizer_steps_total": 32, "peft_method": "lora_narrow", "used_peft": true }, "version": "veriloop.rollback_adapter_trainer.v9.qwen36", "warnings": [ "Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.", "Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op." ] }