{ "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" }