Ctrl+K

1 contributor

v164: COACH parallel-submission configs (licensed, plus_role) reconcile exactly; wscollapse dropped; 715-token reimplementation defect caught (part 3)

1ec26da verified about 1 hour ago

configs
SOMA CoT compression suite - 12 algorithms with 94.77% compression 1 day ago
.gitattributes
1.52 kB
initial commit 1 day ago
BENCHMARK_ANALYSIS.md
8.49 kB
v3: Added comprehensive benchmark analysis and quick reference guide 1 day ago
COMPARISON_MATRIX.md
1.79 kB
v4: Added deployment guide, benchmark script, and comparison matrix 1 day ago
DEPLOYMENT_GUIDE.md
6.97 kB
v4: Added deployment guide, benchmark script, and comparison matrix 1 day ago
QUICK_REFERENCE.md
2.94 kB
v3: Added comprehensive benchmark analysis and quick reference guide 1 day ago
README.md
5.35 kB
v2: Added 3 specialized compressors - Performance achieves 99.04% compression! 1 day ago
achievable_retention.json
704 Bytes
v28: tool_result loss is pattern-bound (90.6%) not budget-bound; CODEISH harvest rejected as no-op; 103 tests about 23 hours ago
achievable_retention_eval.py
3.68 kB
v28: tool_result loss is pattern-bound (90.6%) not budget-bound; CODEISH harvest rejected as no-op; 103 tests about 23 hours ago
acon_axis_report.json
6.97 kB
v99: ACON directional scopes map the frontier; doc-unique proven a strict subset of forward-unique; fwd/bwd union covers 20/25 lost terms; 501 tests about 15 hours ago
ad_first75.json
1.24 kB
v19: size-dependent payload budget (+0.0031 term, t=2.83); fixed cap-scaling blowout 1 day ago
ad_second75.json
1.25 kB
v19: size-dependent payload budget (+0.0031 term, t=2.83); fixed cap-scaling blowout 1 day ago
adaptive_budget_eval.py
5.95 kB
v19: size-dependent payload budget (+0.0031 term, t=2.83); fixed cap-scaling blowout 1 day ago
adaptive_compressor.py
4.83 kB
v41: fixed silent message deletion in 6 legacy compressors; message-count invariant pinned across all 11; 147 tests about 23 hours ago
adaptive_structural_compressor.py
5.65 kB
v41: fixed silent message deletion in 6 legacy compressors; message-count invariant pinned across all 11; 147 tests about 23 hours ago
aggressive_hybrid_compressor.py
4.17 kB
v41: fixed silent message deletion in 6 legacy compressors; message-count invariant pinned across all 11; 147 tests about 23 hours ago
all.json
493 Bytes
v123: BEST FREE RESULT - ws+tab collapse component +0.0038355487 at retention exactly 0.0, 0 content chars touched, 10.3x the v118 strip. Blank-line collapse LOSES (-13913 chars but +6636 tokens): a run of N newlines costs ONE token, so the audit average rate pointed the wrong way. Fixed a backslash-doubling bug in my own safety CHECK; 825 tests about 11 hours ago
all2.json
491 Bytes
v123: BEST FREE RESULT - ws+tab collapse component +0.0038355487 at retention exactly 0.0, 0 content chars touched, 10.3x the v118 strip. Blank-line collapse LOSES (-13913 chars but +6636 tokens): a run of N newlines costs ONE token, so the audit average rate pointed the wrong way. Fixed a backslash-doubling bug in my own safety CHECK; 825 tests about 11 hours ago
all_variant_trim.json
1.23 kB
v42: full package audit vs official metric - trim trap documented, lead margin quantified at 2.04%, dialogue no-op flagged; 151 tests about 23 hours ago
allboth.json
190 Bytes
v152: the elegant clamp-gated repair policy is WRONG ON THE MERITS - it cuts cost 38.7x but 3 of 4 answer-bearing tilde references sit exactly where it declines to repair. Fragment count was the wrong axis. Content-triggered repair dominates: full coverage for 8.6x less than ungated. Caught an impossible-sign attribution error and 5 guard gaps incl. a new class - a magnitude comparison cannot detect a vanished quantity. 1202 tests about 4 hours ago
always.json
140 Bytes
v146: 64.2% of the stack's token savings score EXACTLY zero, spent past the clamp. Gating on the challenge's own pre-stack ratio gives an IDENTICAL component for 51.6% less text rewritten - same reward, half the risk surface. Inverted control scores exactly 0.0 while removing 11601 tokens. Corrected an impossible-sign headroom formula. Recovered a result file I silently clobbered. 1116 tests about 7 hours ago
angle.json
308 Bytes
v136: reopened the v125 delimiter family (closed for the wrong reason) and closed it properly. Characters equally cheap IN ISOLATION are not equally cheap IN CONTEXT - pilcrow has 0 collisions and is 1 token alone yet costs +6762. Elision gains 2025 tokens but 36.65% of span bodies contain newlines, so it is undecodable. 982 tests about 9 hours ago
answer_retention_eval.py
6.32 kB
v100: 5 residual terms are one Django settings cluster; four scope variants used a positional model disproved in v84; corrected gist drop-set variant improves both axes by 0.14 sigma; 513 tests about 15 hours ago
aoff_woff.json
416 Bytes
v124: FIRST ADDITIVE PAIR - allocator +0.0083662574 retention and ws-collapse +0.0038355487 component compose EXACTLY (diffs 0.0). Reason is structural: all 1042 extra substitutions land on CLAMPED challenges and 0 on unclamped, so the two are paid on DISJOINT populations (74 clamped retention / 76 unclamped ratio). New variant structural_wscollapse_compressor.py, default OFF; 838 tests about 11 hours ago
aoff_won.json
446 Bytes
v124: FIRST ADDITIVE PAIR - allocator +0.0083662574 retention and ws-collapse +0.0038355487 component compose EXACTLY (diffs 0.0). Reason is structural: all 1042 extra substitutions land on CLAMPED challenges and 0 on unclamped, so the two are paid on DISJOINT populations (74 clamped retention / 76 unclamped ratio). New variant structural_wscollapse_compressor.py, default OFF; 838 tests about 11 hours ago
aon_woff.json
453 Bytes
v124: FIRST ADDITIVE PAIR - allocator +0.0083662574 retention and ws-collapse +0.0038355487 component compose EXACTLY (diffs 0.0). Reason is structural: all 1042 extra substitutions land on CLAMPED challenges and 0 on unclamped, so the two are paid on DISJOINT populations (74 clamped retention / 76 unclamped ratio). New variant structural_wscollapse_compressor.py, default OFF; 838 tests about 11 hours ago
aon_won.json
475 Bytes
v124: FIRST ADDITIVE PAIR - allocator +0.0083662574 retention and ws-collapse +0.0038355487 component compose EXACTLY (diffs 0.0). Reason is structural: all 1042 extra substitutions land on CLAMPED challenges and 0 on unclamped, so the two are paid on DISJOINT populations (74 clamped retention / 76 unclamped ratio). New variant structural_wscollapse_compressor.py, default OFF; 838 tests about 11 hours ago
aonly.json
166 Bytes
v157: THE FIRST FREE NARROWING. rolemark's 150 U-markers save EXACTLY ZERO tokens (unanimously, all 150, one identified cause) while carrying 36.3% of the transform's sigil ambiguity - 13.3x more ambiguous per substitution than the A-marker that holds 100% of the value. Refines v149 rather than contradicting it: the U-marker is not a trade because it has no gain to trade. Scope audit now complete for all four stack members. 1259 tests, 33/33 mutations fire about 4 hours ago
backref_direction.json
214 Bytes
v142: RETRACTED v130's span census. Its 1033-span pattern required a space after the label, and the space-form is single-line BY CONSTRUCTION - a filter that guaranteed its own answer. All 3381 spans are 53.83% multi-line, mean 1.8953, max 13, with 17 nesting depths. STRUCTURAL EXHAUSTION withdrawn, line-relation family reopens with a 4847-line surface. Backrefs verified 315/315 backward. 1057 tests about 8 hours ago
backref_intra.json
1.09 kB
v142: RETRACTED v130's span census. Its 1033-span pattern required a space after the label, and the space-form is single-line BY CONSTRUCTION - a filter that guaranteed its own answer. All 3381 spans are 53.83% multi-line, mean 1.8953, max 13, with 17 nesting depths. STRUCTURAL EXHAUSTION withdrawn, line-relation family reopens with a 4847-line surface. Backrefs verified 315/315 backward. 1057 tests about 8 hours ago
backref_l.json
170 Bytes
v34: backref round-trip eval - encoder correct (315/0), 3 of my 4 parsers wrong, bad fix reverted; 122 tests about 23 hours ago
backref_r.json
172 Bytes
v34: backref round-trip eval - encoder correct (315/0), 3 of my 4 parsers wrong, bad fix reverted; 122 tests about 23 hours ago
backref_report.json
7.08 kB
v142: RETRACTED v130's span census. Its 1033-span pattern required a space after the label, and the space-form is single-line BY CONSTRUCTION - a filter that guaranteed its own answer. All 3381 spans are 53.83% multi-line, mean 1.8953, max 13, with 17 nesting depths. STRUCTURAL EXHAUSTION withdrawn, line-relation family reopens with a 4847-line surface. Backrefs verified 315/315 backward. 1057 tests about 8 hours ago
backref_roundtrip_eval.py
3.04 kB
v34: backref round-trip eval - encoder correct (315/0), 3 of my 4 parsers wrong, bad fix reverted; 122 tests about 23 hours ago
backref_s.json
166 Bytes
v34: backref round-trip eval - encoder correct (315/0), 3 of my 4 parsers wrong, bad fix reverted; 122 tests about 23 hours ago
backref_spans.json
119 Bytes
v142: RETRACTED v130's span census. Its 1033-span pattern required a space after the label, and the space-form is single-line BY CONSTRUCTION - a filter that guaranteed its own answer. All 3381 spans are 53.83% multi-line, mean 1.8953, max 13, with 17 nesting depths. STRUCTURAL EXHAUSTION withdrawn, line-relation family reopens with a 4847-line surface. Backrefs verified 315/315 backward. 1057 tests about 8 hours ago
backref_struct.json
345 Bytes
v142: RETRACTED v130's span census. Its 1033-span pattern required a space after the label, and the space-form is single-line BY CONSTRUCTION - a filter that guaranteed its own answer. All 3381 spans are 53.83% multi-line, mean 1.8953, max 13, with 17 nesting depths. STRUCTURAL EXHAUSTION withdrawn, line-relation family reopens with a 4847-line surface. Backrefs verified 315/315 backward. 1057 tests about 8 hours ago
backref_surface.json
166 Bytes
v142: RETRACTED v130's span census. Its 1033-span pattern required a space after the label, and the space-form is single-line BY CONSTRUCTION - a filter that guaranteed its own answer. All 3381 spans are 53.83% multi-line, mean 1.8953, max 13, with 17 nesting depths. STRUCTURAL EXHAUSTION withdrawn, line-relation family reopens with a 4847-line surface. Backrefs verified 315/315 backward. 1057 tests about 8 hours ago
band_10k.json
3.27 kB
v117: COACH orders executed - tail-preserving is the best PAID policy (eff 0.294561, 1.71x global) and first positional win; both LOSE when gated free (search beats fixed rule); density targeting falsified; lambda=0.5 not SCORING_ALPHA=0.8 after 4 wrong explanations; 753 tests about 12 hours ago
band_1k.json
3.15 kB
v117: COACH orders executed - tail-preserving is the best PAID policy (eff 0.294561, 1.71x global) and first positional win; both LOSE when gated free (search beats fixed rule); density targeting falsified; lambda=0.5 not SCORING_ALPHA=0.8 after 4 wrong explanations; 753 tests about 12 hours ago
band_1k_192.json
3.43 kB
v117: COACH orders executed - tail-preserving is the best PAID policy (eff 0.294561, 1.71x global) and first positional win; both LOSE when gated free (search beats fixed rule); density targeting falsified; lambda=0.5 not SCORING_ALPHA=0.8 after 4 wrong explanations; 753 tests about 12 hours ago
bareonly.json
177 Bytes
v155: narrowing pathprefix's scope is a PRICED trade (17.3% of gain for 15.2x less retention loss), and my risk assignment was BACKWARDS - the 277 bare mentions are lossless while the 1714 continuations destroy an answer term on 41 challenges. Both mechanism hypotheses were rejected individually and both were partly right, covering only 45% of the gap. Corrected two of my own claims mid-turn. 1235 tests, 34/34 mutations fire about 4 hours ago
barsweep_report.json
4.11 kB
v162: audited my own correction. Per-subset break-even bars (0.3304%-0.3939%) change NO verdict - path+relabel's own bar is stricter than the one I borrowed and it still clears on a zero bound. The retracted 8.685% never licensed a decision. v139's one-flip multiple was wrong the OTHER way (12x vs 1.69x). Formula error N*G/C vs N*G/(G+C) found by an internal contradiction. 1315 tests, 20/20 mutations fire about 3 hours ago
base.json
273 Bytes
v160: wscollapse scope audit closes the stack. Tabs are 15.4% of the gain from 4.9% of runs - but 91.0% of that is ONE challenge, so AN AGGREGATE CAN BE ONE DATA POINT. Both mechanism hypotheses rejected by measurement. Seventh isolated-probe failure: both classes save 0 tokens standalone while the arms save 673. New defect class: A UNITTEST SKIP EXITS 0. 1296 tests, 40/40 mutations fire about 3 hours ago
base_rate_test.json
638 Bytes
v90: loss mechanism pinned (100% of lost terms only in call bodies); two wrong conclusions corrected; held-out gain +0.259 char, a tenth of the oracle; 419 tests about 16 hours ago
base_rate_test.py
1.93 kB
v90: loss mechanism pinned (100% of lost terms only in call bodies); two wrong conclusions corrected; held-out gain +0.259 char, a tenth of the oracle; 419 tests about 16 hours ago
basename_novelty.json
141 Bytes
v85: no official harness exists; pathkeep retention NULL (retracts v84 +0.0009); 89.2% of restored basenames already present; 3 compressors registered; 373 tests about 17 hours ago
basename_novelty.py
1.8 kB
v85: no official harness exists; pathkeep retention NULL (retracts v84 +0.0009); 89.2% of restored basenames already present; 3 compressors registered; 373 tests about 17 hours ago
benchmark.py
3.21 kB
v4: Added deployment guide, benchmark script, and comparison matrix 1 day ago
bet_clamp_measure.py
1.24 kB
v69: COACH paired run + SCORING_R_MAX ceiling binds on 74/150 challenges, killing 63% of cap-28 gain; 263 tests about 19 hours ago