5.68 MB

Ctrl+K

1 contributor

v181: shipped lead raised TypeError on the platform's own compress_messages call shape; 28/28 fixed, outputs byte-identical, tokens unchanged at 278973 (part 4)

b3f4cc3 verified 2 months ago

configs
SOMA CoT compression suite - 12 algorithms with 94.77% compression 2 months ago
.gitattributes

1.52 kB
initial commit 2 months ago
BENCHMARK_ANALYSIS.md

8.49 kB
v3: Added comprehensive benchmark analysis and quick reference guide 2 months ago
COMPARISON_MATRIX.md

1.79 kB
v4: Added deployment guide, benchmark script, and comparison matrix 2 months ago
DEPLOYMENT_GUIDE.md

6.97 kB
v4: Added deployment guide, benchmark script, and comparison matrix 2 months ago
QUICK_REFERENCE.md

2.94 kB
v3: Added comprehensive benchmark analysis and quick reference guide 2 months ago
README.md

5.35 kB
v2: Added 3 specialized compressors - Performance achieves 99.04% compression! 2 months 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 2 months 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 2 months 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 2 months ago
ad_first75.json

1.24 kB
v19: size-dependent payload budget (+0.0031 term, t=2.83); fixed cap-scaling blowout 2 months ago
ad_second75.json

1.25 kB
v19: size-dependent payload budget (+0.0031 term, t=2.83); fixed cap-scaling blowout 2 months ago
adaptive_budget_eval.py

5.95 kB
v19: size-dependent payload budget (+0.0031 term, t=2.83); fixed cap-scaling blowout 2 months 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 2 months ago
adaptive_structural_compressor.py

5.79 kB
v181: shipped lead raised TypeError on the platform's own compress_messages call shape; 28/28 fixed, outputs byte-identical, tokens unchanged at 278973 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months ago
audit_cfgA.json

897 Bytes
v172: COACH orders executed - split-invariant audits + king gate located at metagraph registration (403); burn unobtained 2 months ago
audit_cfgB.json

911 Bytes
v172: COACH orders executed - split-invariant audits + king gate located at metagraph registration (403); burn unobtained 2 months ago
audit_coach_v172.json

4.5 kB
v172: COACH orders executed - split-invariant audits + king gate located at metagraph registration (403); burn unobtained 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months 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 2 months ago
basegen_cot.json

271 Bytes
v169: licensed config proved base-independent on three bases; identical token delta explained by disjoint support; headroom-is-not-merit caveat 2 months ago