{ "slug": "causalqwen", "title": "CausalQwen MSP Paper v0.1 + Runnable Evidence", "last_updated": "2026-05-30 22:16 CST", "status": "runnable_evidence_gate_a_b_passed_gate_c_designed_blocked", "evidence_level": "E2 + toy runnable evidence", "public_url": "https://research.wehub.us/ghy/tmp/causalqwen/", "claim_boundary": "design_and_toy_runnable_evidence_not_reproduced_full_causalqwen_experiment_no_proven_forgetting_mitigation", "artifacts": { "paper_markdown": "research-projects/causalqwen/paper/causalqwen-msp-paper-v0.1.md", "paper_status": "research-projects/causalqwen/paper/paper-status-v0.1.md", "site_index": "site/ghy/tmp/causalqwen/index.html", "progress_json": "site/ghy/tmp/causalqwen/progress.json", "source_manifest": "research-projects/causalqwen/source-bundle/manifest-v0.1.json", "gate_a_cauchy_math": "research-projects/causalqwen/evidence/gate_a_cauchy_math/results.json", "gate_b_fake_hx_num_gate": "research-projects/causalqwen/evidence/gate_b_fake_hx_num_gate/results.json", "gate_c_inheritance_anchor_drift_design": "research-projects/causalqwen/evidence/gate_c_qwen_inheritance_anchor_drift/results.json" }, "motivation": [ "Bridge LLM conditional next-token learning with individual-level causal/interventional action semantics.", "Reuse Qwen as a pretrained FeatureNetwork instead of relearning all language knowledge.", "Unify text token decisions and numeric regression through a shared latent causal representation U." ], "core_formal_content": { "pipeline": "x -> Qwen FeatureNetwork -> h(x) -> AbductionNetwork -> U ~ Cauchy(mu_U, gamma_U) -> ActionNetwork", "abduction": "mu_U = W_mu h + b_mu; gamma_U = softplus(W_gamma h + b_gamma)", "cauchy_stability": "sum_j w_j U_j + b ~ Cauchy(sum_j w_j mu_j + b, sum_j |w_j| gamma_j)", "token_score": "S_k = A_k · U + B_k", "token_probability": "P_k = 1/2 + 1/pi · arctan((loc_Sk - C_k) / scale_Sk)", "classification_loss": "L_cls = -sum_k [y_k log P_k + (1-y_k) log(1-P_k)]", "numeric_loss": "L_reg_gated = I(y_true_id=_ID) · P_ · L_cauchy_nll", "alignment_loss": "L_align = sum_i sum_{k in K_teacher,i} (P_CausalQwen_{i,k} - P_Qwen_{i,k})^2", "total_design_loss": "L_total = L_task + lambda_align L_align + lambda_anchor L_anchor + lambda_mech L_mech" }, "forgetting_mitigation_design": { "boundary": "design hypothesis only; not experimentally verified", "mechanisms": [ "Qwen backbone / feature inheritance", "offline teacher Top-K feature distillation", "dense Top-K alignment gradients for causal heads", "future anchor-set drift regularization", "future mechanism-level intervention/action distribution preservation" ], "required_evidence_gate": "teacher-vs-student anchor drift must be measured before claiming anti-forgetting evidence" }, "milestones": [ { "id": "source_bundle", "label": "29-file source bundle indexed", "done": true }, { "id": "msp_paper_v0_1", "label": "MSP paper v0.1 drafted", "done": true }, { "id": "motivation_forgetting_math_refine", "label": "Motivation, anti-forgetting design, and formulas refined", "done": true }, { "id": "public_review_surface", "label": "Public review/paper surface created", "done": true }, { "id": "mathjax_formula_render_fix", "label": "MathJax formula render errors fixed on public review surface", "done": true }, { "id": "cauchy_math", "label": "Gate A Cauchy math contract verification passed", "done": true }, { "id": "fake_h_proof", "label": "Gate B fake-h(x) gated proof passed", "done": true }, { "id": "qwen_inheritance", "label": "Real Qwen weight/logit alignment check", "done": false }, { "id": "qwen_inheritance_anchor_drift_design", "label": "Gate C Qwen inheritance + anchor drift minimal metrics designed; real run blocked on env", "done": true }, { "id": "anti_forgetting_anchor", "label": "Real teacher-vs-student anchor drift measured", "done": false } ], "next_gates": [ "Run real Gate C on a machine/env with transformers + local Qwen2.5-0.5B weights", "Measure initial shared-vocab cls_loc vs Qwen logits before any training", "Only after target improvement + bounded anchor drift may anti-forgetting wording be upgraded", "If real Qwen gate is unavailable, keep CausalQwen at E2 toy runnable evidence boundary" ], "validation": { "local_files_updated": true, "local_json_checked": true, "public_deploy_verified": false, "html_checked": false, "json_checked": false, "browser_checked": false, "console_errors": null, "source_public_root_sha256_parity": false, "verified_at": null, "html_mathjax_source_checked": true, "mathjax_formula_errors": 0, "control_chars_in_html": 0, "html_raw_lt_t_fixed": true, "motivation_inline_math_checked": true, "public_mathjax_browser_verified": true, "public_console_errors": 0, "public_mjx_merrors": 0, "public_raw_latex_visible": false, "public_formula_count": 8, "public_math_nodes": 20, "runnable_evidence_gate_a_passed": true, "runnable_evidence_gate_b_passed": true, "gate_c_designed_blocked_on_env": true, "docs_tasks_site_sync_pending_public_verification": true, "site_public_evidence_bundle_included": true }, "deployed_to": { "source_mirror": "dgx2:/home/cms/wehub-sites/research/ghy/tmp/causalqwen/", "public_root": "dgx2:/var/www/research.wehub.us/ghy/tmp/causalqwen/", "parent_index": "https://research.wehub.us/ghy/tmp/", "public_url": "https://research.wehub.us/ghy/tmp/causalqwen/", "progress_json": "https://research.wehub.us/ghy/tmp/causalqwen/progress.json" }, "runnable_evidence": { "boundary": "Gates A/B are standalone toy/contract-level checks; Gate C is metric design/readiness only until transformers + local Qwen weights are available.", "gate_a": { "status": "passed", "artifact": "research-projects/causalqwen/evidence/gate_a_cauchy_math/results.json", "checks": [ "parameter_formula", "sampling_stability", "ovr_probability", "cauchy_nll_gradients" ], "sample_count": 250000, "max_scale_relative_error": 0.002384556217894077 }, "gate_b": { "status": "passed", "artifact": "research-projects/causalqwen/evidence/gate_b_fake_hx_num_gate/results.json", "classification_accuracy": 0.9300000071525574, "num_precision": 0.9567099567099567, "num_recall": 1.0, "num_mae": 0.0493672713637352, "non_num_regression_gradient_norm": 0.0 }, "gate_c": { "status": "designed_blocked", "real_qwen_gate_status": "blocked_on_missing_transformers_or_weights", "artifact": "research-projects/causalqwen/evidence/gate_c_qwen_inheritance_anchor_drift/results.json", "transformers_available": false, "qwen_weight_path_found": false, "mock_topk_overlap_near": 1.0, "mock_topk_overlap_far": 0.05 } }, "public_evidence_artifacts": { "gate_a_results": "https://research.wehub.us/ghy/tmp/causalqwen/evidence/gate_a_cauchy_math/results.json", "gate_a_readme": "https://research.wehub.us/ghy/tmp/causalqwen/evidence/gate_a_cauchy_math/README.md", "gate_b_results": "https://research.wehub.us/ghy/tmp/causalqwen/evidence/gate_b_fake_hx_num_gate/results.json", "gate_b_readme": "https://research.wehub.us/ghy/tmp/causalqwen/evidence/gate_b_fake_hx_num_gate/README.md", "gate_c_results": "https://research.wehub.us/ghy/tmp/causalqwen/evidence/gate_c_qwen_inheritance_anchor_drift/results.json", "gate_c_readme": "https://research.wehub.us/ghy/tmp/causalqwen/evidence/gate_c_qwen_inheritance_anchor_drift/README.md" } }