| { |
| "axis": "ACON arXiv:2510.00615 - keep what later steps need - applied to the routing problem I had failed at from four directions", |
| "arxiv_citations": [ |
| { |
| "title": "ACON: Optimizing Context Compression for Long-horizon LLM Agents", |
| "id": "arXiv:2510.00615", |
| "venue": "ICML 2026", |
| "date": "2025-10-01" |
| }, |
| { |
| "title": "Distilling LLM Agent into Small Models with Retrieval and Code Tools", |
| "id": "arXiv:2505.17612" |
| }, |
| { |
| "title": "Structured Agent Distillation for Large Language Models", |
| "id": "arXiv:2505.13820" |
| } |
| ], |
| "acon_gap": "ACON fits its compressor from outcome feedback (compare success with full vs compressed context). I have no run outcomes - no OpenRouter key - so I took only its structural principle: directionality", |
| "challenge_level_rates_useless_again": { |
| "results": { |
| "fwd": -0.0081, |
| "bwd": -0.0017, |
| "doc": -0.0092 |
| }, |
| "reading": "all three separations are near zero and NEGATIVE at the challenge level; I ran this first out of habit and correctly refused to conclude from it, because v95 established that a challenge-level rate is the wrong instrument for a span-level rule. Second time in three turns it pointed the opposite way from term coverage" |
| }, |
| "span_lattice": { |
| "results": { |
| "fwd": 452, |
| "bwd": 509, |
| "doc": 127, |
| "doc_subset_of_fwd": true, |
| "doc_minus_fwd": 0, |
| "doc_subset_of_bwd": true, |
| "fwd_and_bwd": 138, |
| "fwd_or_bwd": 823, |
| "fwd_only": 314, |
| "bwd_only": 371 |
| }, |
| "finding": "document-unique is a STRICT SUBSET of both forward-unique and backward-unique (0 spans outside either) - logically necessary since \"appears nowhere else\" implies \"appears nowhere later\"", |
| "how_i_found_it": "I tried to write a fwd-OR-doc union variant and saw the disjunction collapse; the subset relation was only visible once I wrote the predicate down", |
| "reframing": "this corrects how I described the forward scope last turn. It is not a DIFFERENT feature that beat mine, it is a strict RELAXATION of mine. My document-wide rule was over-restrictive at 127 spans when the useful set was 452. The gain came from loosening a too-tight condition, not from directionality per se", |
| "non_vacuous_pair": "fwd and bwd share only 138 of 823 spans, so they are genuinely different and their union is the one combination worth building" |
| }, |
| "term_lattice": { |
| "results": { |
| "total_lost_terms": 25, |
| "fwd": 16, |
| "bwd": 15, |
| "doc": 8, |
| "fwd_and_bwd": 11, |
| "fwd_or_bwd": 20, |
| "fwd_only": 5, |
| "bwd_only": 4, |
| "doc_subset_of_fwd_terms": true, |
| "uncovered_by_either": 5 |
| }, |
| "finding": "fwd covers 16 of 25 lost terms, bwd 15, sharing only 11 - so 5 are forward-only and 4 backward-only, and the union reaches 20 of 25 (80%)", |
| "why_it_matters": "the first combination on this axis with headroom over its parts; only 5 lost terms are unreachable by either scope" |
| }, |
| "frontier": { |
| "points": [ |
| { |
| "variant": "shipped", |
| "spans": 0, |
| "terms_covered": 0, |
| "char_pct": 78.136, |
| "term_10dp": 0.935724891 |
| }, |
| { |
| "variant": "doc", |
| "spans": 127, |
| "terms_covered": 8, |
| "char_pct": 80.896, |
| "term_10dp": 0.9344322611 |
| }, |
| { |
| "variant": "forward", |
| "spans": 452, |
| "terms_covered": 16, |
| "char_pct": 80.656, |
| "term_10dp": 0.9348915453 |
| }, |
| { |
| "variant": "combined", |
| "spans": 823, |
| "terms_covered": 20, |
| "char_pct": 80.274, |
| "term_10dp": 0.9351396283 |
| } |
| ], |
| "monotone": "every step up in coverage buys retention and costs compression with no reversals: 8/16/20 terms maps onto retention 0.93443/0.93489/0.93514 and char 80.896/80.656/80.274", |
| "coverage_predicted_ordering": "coverage got the ORDERING right here even though it has never predicted magnitudes" |
| }, |
| "combined_multiseed": { |
| "results": { |
| "per_seed": [ |
| { |
| "seed": 11, |
| "shipped": { |
| "char_pct": 79.119, |
| "term_10dp": 0.931462117 |
| }, |
| "combined": { |
| "char_pct": 81.129, |
| "term_10dp": 0.9309158796 |
| } |
| }, |
| { |
| "seed": 29, |
| "shipped": { |
| "char_pct": 76.17, |
| "term_10dp": 0.945231517 |
| }, |
| "combined": { |
| "char_pct": 78.369, |
| "term_10dp": 0.9448618065 |
| } |
| }, |
| { |
| "seed": 47, |
| "shipped": { |
| "char_pct": 78.453, |
| "term_10dp": 0.933831047 |
| }, |
| "combined": { |
| "char_pct": 80.695, |
| "term_10dp": 0.9331452292 |
| } |
| }, |
| { |
| "seed": 73, |
| "shipped": { |
| "char_pct": 77.266, |
| "term_10dp": 0.9440895859 |
| }, |
| "combined": { |
| "char_pct": 79.43, |
| "term_10dp": 0.9431722632 |
| } |
| }, |
| { |
| "seed": 101, |
| "shipped": { |
| "char_pct": 79.151, |
| "term_10dp": 0.9301182346 |
| }, |
| "combined": { |
| "char_pct": 81.241, |
| "term_10dp": 0.9298644773 |
| } |
| } |
| ], |
| "term_delta_mean": -0.0005545691, |
| "term_delta_stdev": 0.0002615082, |
| "char_delta_mean": 2.141, |
| "char_delta_stdev": 0.092, |
| "seeds_losing_retention": 5 |
| }, |
| "significance": "2.12 sigma, losing on all 5 seeds" |
| }, |
| "gate": { |
| "disabled_identical_to_shipped": 150, |
| "invariants": { |
| "unbalanced": 0, |
| "tildes_even": true, |
| "grew": 0, |
| "backrefs": 315, |
| "unresolved": 0 |
| }, |
| "decoder_forward": { |
| "role_recovery": 1.0, |
| "tool_recovery": 1.0, |
| "pair_recovery": 1.0, |
| "order_fidelity": 1.0, |
| "challenges": 150, |
| "event_count_matches": 150 |
| }, |
| "decoder_combined": { |
| "role_recovery": 1.0, |
| "tool_recovery": 1.0, |
| "pair_recovery": 1.0, |
| "order_fidelity": 1.0, |
| "challenges": 150, |
| "event_count_matches": 150 |
| } |
| }, |
| "decision": "shipped remains the lead; forward and combined are NOT promoted and default OFF. The combined rule is the best aggressive policy on record (+2.141 char for -0.00055 term). Still a trade, and I cannot convert a retention delta into run-failure probability - the 7.6% break-even from v82 is unmeasurable without an OpenRouter key", |
| "what_changed": "the frontier is now MAPPED - four points, monotone, multi-seed variance on each - so a future decision with a real retention-to-failure mapping can pick a point off this curve instead of re-deriving it", |
| "revert_proofs": "22 checks: 3 LIVE source mutations (combined default enabled, forward default enabled, union collapsed to the subsumed doc-unique conjunction) and 19 report mutations, all firing" |
| } |
