{ "schema": "consumption-october-2026-summary-v1", "project": "Simone Consumption Engine", "repository": "https://github.com/thisisntjon/consumption", "head_commit": "2694413", "reviewed_date": "2026-10-24", "scope_boundary": "Author-run engineering and evaluation receipts from github.com/thisisntjon/consumption through 2026-10-24 (commit 2694413). No independent external reproduction has been logged yet. Code-brick rebuild results (100.00% on ApprovalBench in P1a/P1b) test source-blind Python feature reconstruction from extracted AST contracts and do not establish general superiority across arbitrary un-extracted repositories. Earlier negative and mixed studies (September 30 skill-selection 0 autonomous calls, October 7 37-task study 48.8% and 49.0% vs 53.8% ordinary, October 20 P-extract1 0.0 line F1, October 21 P-rom1 3/28 held-out recall, and October 23 P-docatlas3 6/10 intruder check vs 8/10 bar) remain part of the record.", "tracks": { "code_bricks": { "summary": "AST-based Python unit extraction, feature mapping, execution-pruned oracle synthesis, mechanical self-maintenance, and source-blind feature reconstruction.", "milestones": [ { "phase": "P0", "date": "2026-10-10", "commit": "9076077", "title": "Single-home repository layout and deterministic foundation checks", "verdict": "PASS", "metrics": { "tracked_files_checked": 1138, "foundation_rules_enforced": 6, "planted_violations_caught": "4/4" }, "scope": "Enforces single-repository governance rules (R1, R2, R3, R7, R9, R10) and sprawl scanning." }, { "phase": "TFE1", "date": "2026-10-11", "commit": "48129dd", "title": "Binary differential judge falsification check", "verdict": "PASS", "metrics": { "add3_toolchain_vectors_failed": "4/29", "prior_false_pass_caught": 1 }, "scope": "Repaired binary differential judge so an incomplete add3 toolchain rebuild fails 4/29 vectors instead of passing with 0 round-trip vectors." }, { "phase": "P1a", "date": "2026-10-14", "commit": "86ff4a7", "title": "Print-head probe on ApprovalBench (Rung 1 Subtract & Rung 2 Rebuild)", "verdict": "PASS", "metrics": { "target_program": "ApprovalBench (61 AST units, 74 tests: 61 S + 13 U)", "rung1_subtract_arm_b_score": 100.0, "rung1_bar": 95.0, "rung2_arm_a_bricks_runs": 3, "rung2_arm_c_docs_runs": 3, "rung2_mean_advantage_over_docs": 51.5188, "rung2_max_docs_only_score": 48.3588, "independent_same_project_reproductions": "7/7" }, "scope": "Compares source-blind feature reconstruction from hand-extracted code bricks plus thin docs (Arm a, 3 runs) against thin docs alone (Arm c, 3 runs) on ApprovalBench; all 3 brick runs beat all 3 docs-only runs by a mean of 51.52 points." }, { "phase": "P1b", "date": "2026-10-17", "commit": "c662675", "title": "Automated AST deconstruct-and-select on ApprovalBench", "verdict": "PASS", "metrics": { "ast_units_verified": "61/61", "features_mapped": 4, "visible_tests_used": 27, "hidden_h1_tests_leaked": 0, "canary_hits": "0/20", "arm_a_auto_runs": 3, "arm_a_auto_mean_score": 100.0, "arm_a_hand_runs": 3, "arm_a_hand_mean_score": 97.7382, "auto_vs_hand_delta": 2.2618 }, "scope": "Zero-human-touch AST brick extractor restricted to 27 visible tests matched and slightly exceeded hand-built bricks (100.0000 vs 97.7382 mean across 3 interleaved runs per arm)." }, { "phase": "P1c", "date": "2026-10-18", "commit": "bfc29ff", "title": "Blind oracle synthesis, mechanical self-maintenance, and qudbud slice workflow", "verdict": "PASS", "metrics": { "raw_test_executability": "162/164 (98.78%)", "ast_branch_coverage": "179/180 (99.44%)", "ast_mutant_kill_rate": "168/182 (92.31%)", "shipped_suite_kill_recall": "107/107 (100.0%)", "hand_mutants_killed": "15/15", "cli_untested_agreement": "27/30 (10/10 families)", "sm_behaviour_changing_flagged": "15/15 (100.0%)", "sm_cosmetic_false_stale": "0/12 (0.0%)", "sm_refactor_false_broken": "0/7 (0.0%)", "sm_affected_bricks_reproved": "22/22 (100.0%)", "p1c7_qudbud_feature_jaccard": 1.0, "p1c7_qudbud_sibling_leaks": 0 }, "scope": "Synthesizes pytest oracles with tests/ hidden and evaluates mechanical self-maintenance across 34 edits (15 behaviour-changing, 12 cosmetic, 7 refactors) with 0 human edits, plus connected workflow verification on qudbud." } ] }, "document_and_corpus_atlas": { "summary": "Read-only local corpus inventory, secret/key refusal before file open, SHA-256 duplicate collapse, SQLite FTS5 indexing, and negative findings on unsupervised clustering and word-only feature extraction.", "milestones": [ { "phase": "P-extract1", "date": "2026-10-20", "commit": "ac571d4", "title": "Word-only component feature location on unseen assembly/C repositories (NEGATIVE / LOSS)", "verdict": "LOSS", "metrics": { "unseen_features_tested": 5, "line_f1": 0.0 }, "scope": "Negative finding: locating code features by lexical word matching alone scored 0.0 line F1 on 5 unseen features (L35), motivating behaviour-trace and AST-dependency slicing." }, { "phase": "P-game1", "date": "2026-10-20", "commit": "6b0cab1", "title": "Two-game 6502 routine transplant with engine-made address adaptation", "verdict": "PASS", "metrics": { "oracle_checks_passed": "16/16" }, "scope": "Transplanted a freeze-ability routine between two homebrew 6502 games with automated address adaptation (16/16 oracle checks passed); initial component identification was manual." }, { "phase": "P-rom1", "date": "2026-10-21", "commit": "ec1467d", "title": "Blind behaviour-only RAM labelling on held-out 16 KB mapper-0 binaries (NEGATIVE / LOSS)", "verdict": "LOSS", "metrics": { "held_out_ram_recall": "3/28 (0.107, positions only)", "gate_g5_score": "0/12", "routine_matcher_accuracy": 0.955 }, "scope": "Negative finding: blind behaviour-only RAM role labelling achieved 0.107 recall (3/28) and 0/12 on G5 (CLOSED LOSS); routine-matching component (0.955) carried forward into P-parts1." }, { "phase": "P-parts1", "date": "2026-10-21", "commit": "317bc7e", "title": "Cross-workspace source-pinned lesson and outcome chains", "verdict": "PASS", "metrics": { "verified_cross_workspace_chains": 2, "checksum_manifest_files_verified": "19/19" }, "scope": "Verified prospective before-work brief retrieval and after-work lesson writeback across two workspaces with SHA-256 source-span verification." }, { "phase": "P-accept1-docatlas2", "date": "2026-10-23", "commit": "166ce68", "title": "Read-only Project Atlas and Document Atlas over 883,486 local files", "verdict": "PASS", "metrics": { "total_files_mapped": 883486, "total_corpus_gb": 84.5, "top_level_entries": 122, "exact_duplicate_groups": 24513, "reclaimable_duplicate_gb": 33.74, "never_read_secret_files_opened": 0 }, "scope": "Read-only census mapping 883,486 files and 24,513 exact duplicate groups without opening any secret, key, credential, or database content." }, { "phase": "P-docatlas3", "date": "2026-10-23", "commit": "a7bfae0", "title": "Unsupervised TF-IDF/LSA topic clustering evaluation (NEGATIVE / NULL NEAR MISS)", "verdict": "NULL_NEAR_MISS", "metrics": { "clusters_k": 30, "gate_g1_largest_family": "14 (<= 5% bar, PASS)", "gate_g2_silhouette": "0.209 (PASS)", "gate_g3_intruder_score": "6/10 (z = 1.53 vs 8/10 preregistered bar, FAIL)" }, "scope": "Negative finding: unsupervised TF-IDF/LSA clustering passed family-size (G1) and silhouette (G2) checks but scored 6/10 on the blinded intruder check (G3 bar 8/10) and was rejected in favor of explicit project/path taxonomy and FTS5 search." } ] }, "whole_pipeline_integration": { "summary": "Connected Atlas-to-Engine CLI intake, SHA-256 object store and SQLite/FTS5 catalogue, ordinary-operation lifecycle, and local workbench verification.", "milestones": [ { "phase": "P-mvp1-4", "date": "2026-10-24", "commit": "05f5d02", "title": "End-to-end Atlas-to-Engine intake and verification on 5 live projects", "verdict": "PASS", "metrics": { "projects_ingested": 5, "eligible_files": 768, "keepers_ingested": 699, "duplicates_skipped": 69, "duplicate_bytes_avoided": 304894, "unsupported_extension_excluded": 489, "singleton_entries_ingested": 125, "canonical_items_admitted": 7727, "flagged_items": 55, "store_objects_verified": 811, "store_records_verified": 8180, "store_verification_errors": 0, "live_span_sha256_checks": "8013/8013 (100.0%)", "bounded_retrieval_questions_passed": "3/3", "raw_source_mb": 11.91, "bricks_jsonl_mb": 27.58, "bricks_expansion_ratio": 2.316, "catalogue_sqlite_mb": 147.54 }, "scope": "Connected ingest-atlas -> ce_storage verify -> prepare-agent across 5 projects (699 keepers, 69 duplicates skipped, 0 secret bytes read, 8,013/8,013 span hashes verified). Discloses 2.316x JSONL and 147.54 MB SQLite storage expansion." }, { "phase": "P-mvp-close", "date": "2026-10-24", "commit": "2694413", "title": "Whole-pipeline storage, closure-rendering, and local workbench verification", "verdict": "PASS", "metrics": { "phases_closed": "4/4 (MVP-Close-1 through MVP-Close-4)", "storage_feedback_tests": "25/25", "agent_preparation_tests": "60/60", "agent_printing_tests": "25/25", "ordinary_operations_tests": "2/2", "ui_backend_and_launcher_tests": "23/23", "red_team_findings_remediated": "5/5", "foundation_tracked_files_ok": 1138 }, "scope": "Closes storage/outcome replay checks (PS07-PS10), ce-evidence-closure-rendering-v1 range support, ordinary-operation CLI lifecycle, and same-origin/bootstrap session token hardening. Remaining boundaries: native OS file-picker modal click requires a human desktop session, two brief producers share ce-agent-brief-v1, and heavy external worker binaries remain optional." } ] } } }