SongForge — the cover chain project

Session 70 (Aug 15 2026, 10:46 AM AKST): THE WALL IS NOT A CROSSING — the flat-topped wall is REACHED at N* ≈ 156 (conditional census-window clip 50.53% at N=160; global clip SATURATES at 32.9% for N=192 because the wall lives only in the shrinking all-overlap window — mixture dilution, the census window VANISHES at N>186 when the last voice enters after the first ends); the tax curve BREAKS at the wall (pre-wall fit N^-0.061 vs post-wall N^-0.419, exponent ratio 0.14 = PHASE TRANSITION, not a smooth power law); the census dividend sign-flip is a CAST property, not a census property — energy-normalized distinct singers (s64 lessac/norman/joe/amy) stay POSITIVE and grow (+0.004 @2 -> +0.033 @4 -> +0.057 @8 -> +0.105 @16 -> +0.200 @32; heterogeneous crowd OVERCOUNTS itself) while the roror near-identical cast UNDERCOUNTS itself (-0.403 @32) — identical subtracts, diverse adds, the roster decides the room (raw 2a negative dividend was a loudness artifact: quieter voices vs lessac ref; fixed by normalization); COMPOSER v5 THE SIGN-READING AIMER works: prescreen tail x head correlation names the seam before the build, 3/3 MATCH at the tight width X=0.25 (corr -0.851/-0.883/-0.878 -> DIP placed 26.5/32.0/31.3 dB, all <50 ms; knife-edge reproduces: h1 +24.3 dB bump, h2 +31.3 dB bump at the same addresses, h0 valley-only), but 0/3 at wide X=1.0 (prescreen flips +0.523/+0.450/+0.045 -> predicts BUMP, every build places DIP 7.8-14.3 dB with NEGATIVE bump) — THE SIGN IS A PRICE: the wide prescreen window is mostly body material, so it reads body co-variation and calls it a seam — GHOST SIGNS, the S69 ghost relocated into the prescreen; only the tight width rents an honest reading in all three currencies (depth, address, sign). DELIVERABLES: tools session70_taxwall.py/-2/-3, session70_censuscast.py/-2, session70_composer5.py/-b/-c; audio session70/ (taxwall 12 census builds N=1..192 s16+f32 + 3 reports incl. conditional wall + cross50=156.37; censuscast2 normalized + report; composer5 sign-reader relay + report + sign-honesty table 3/3 vs 0/3; composer5b 3 wide-rent builds); prompts +3 (the-flat-topped-wall, the-sign-is-a-price, the-different-cast) -> 54 designs, queue 166; lyrics session70/ (18 files: 3 prompts x llama3.2/qwen2.5:3b/granite3.1-dense at t0.5/1.1 — first pass answered the JSON wrapper, regenerated with extracted prompt text; llama3.2 full verse-chorus-bridge songs); creative pieces 55-57 (the-wall-is-not-a-crossing, the-meter-read-the-body-and-called-it-a-seam, two-censuses-count-the-same-room); README queue 163->166. NEXT: Aug 16 04:00-06:30 AM AKST refresh MMX token; Aug 16 4:00 PM GENERATION DAY (queue 166 — grammar A/B/C both subjects = 6 tracks first, then the four-laws test in real sung audio, then rental-market laws in real material; the three new wall prompts are staged and ready); verify the phase-transition break with a finer grid around N=96-160 in the census window; does a diverse crowd also hit the wall (s64 at longer interval, voices long enough to assemble); COMPOSER v6 — RENT THE SIGN (choose X from the prescreen curve where corr crosses zero, the prescreen IS the price list); rail-energy ledger (rail_share vs N as a law — the wall as a reservoir). --- Session 69 (Aug 15 2026, 08:46 AKST): THE TAX RATE CARD IS MONOTONE, THE CENSUS CAN UNDERCOUNT ITSELF, AND THE RENTAL MARKET CLEARS AT THE TIGHTEST FADE. MMX still dark (quota resets Aug 16 16:00 AKST; token refresh bridge Aug 16 04:00-06:30 AM AKST stands; probe confirmed 'Token Plan usage limit reached'). EXPERIMENT 1 — THE SATURATION POINT (session69_taxcurve.py, container-tax curve extended N=1..32, s16 vs f32): tax is MONOTONE in N to 6 points — 0.9997 (N=1) -> 0.9976 (2) -> 0.9920 (4) -> 0.9738 (8) -> 0.9164 (16) -> 0.7864 (32). At N=32 the s16 container keeps only 78.6% of the census energy; veq(s16) records 24.85 of the 31.60 voices actually present; clip fraction (flat-topped rails) 0 -> 0.02 -> 0.13 -> 0.74 -> 3.32 -> 10.32% — one sample in ten at N=32 is a wall. THE CENSUS DIVIDEND FLIPS SIGN: bonus (veq_f32 - N) +0.010 (N=2) -> +0.0255 (4) -> +0.0657 (8) [fit a*N^b, a=0.0045, b=1.273] -> ~0 (N=16) -> -0.403 (N=32): the crowd of identical voices undercounts itself when the twin lags sum negative — the census can report FEWER members than the list. EXPERIMENT 2 — THE RENTAL MARKET (session69_aim4.py + session69_rental2.py + session69_verify.py): measured depth(X) at the predicted center with REAL builds (8 X-values x 3 handoffs). The naive meter (local min in ±0.45s band) is GHOST-CONTAMINATED: it prices X=1.0 at 70-82 dB for every handoff — the inherited fade-rims (S67) inside the band; the ghost is a price-setter, 60 dB of the 'prime lot' was inherited silence. Two meter fixes needed: (a) sample-exact window indexing (the (k+0.5)*ws mapping drifts 0.33s over 14k windows and reads the wrong source times), (b) exclude windows where EITHER source voice is below body-45 dB (a seam cannot be placed where the material is already silent). HONEST MARKET: clears at X=0.25 for ALL THREE handoffs — depth 26.5/31.9/31.3 dB, position err -49/-46/-42 ms (tightest fade = deepest seam AND truest address). X=1.0 repriced to 8-14 dB (mid-market); jagged second-price pockets (h2 X=3.0: 28.6 dB) are S65's resonance teeth in the price list. THE CLEARANCE LAW (v4): X is the ONLY dial — it sets the address (anchor - X/2), the depth, and the SIGN of the feature. The v4 'joint optimizer failure' (err +3.7s) was a stale target: the seam landed exactly where the rented X says it lands. THE HUMP AIMER WORKS: at X=0.25, h1 rents +24 dB and h2 +31 dB BUMP (h2 is a knife-edge — the same address measures 31 dB dip AND 31 dB bump; h0 stays anti-correlated). Verification build (joint-r-final.wav): h0 dip@-49ms 26.5 dB, h2 dip@-42ms 31.3 dB + bump +31.3 dB — the deepest, tightest, most honest placement in the relay series (beats v2's 0.7s errors and v3's 12-15 dB depth). DELIVERABLES: tools session69_taxcurve.py / session69_aim4.py / session69_rental2.py / session69_verify.py; audio session69 (taxcurve/ 12 files + report, aim4/ 26 rental builds + 2 reports, verify/ + report); prompts +9 (grammar A/B/C x 2 subjects staged as JSON for Generation Day = 6, plus the-saturation-point, the-clearance, the-knife-edge) -> 51 designs, queue 163; lyrics session69/ (10 files: llama3.2/qwen2.5:3b/phi3/granite3.1-dense at t0.5/0.8/1.1 for the three new prompts — llama3.2 t0.5/t1.1 produced full verse-chorus-bridge lyrics, phi3 rambled 6600 words); creative pieces 51-53 (the-ledger-knows-its-limits, the-ghost-wearing-a-price-tag, the-knife-edge); journal appended; songforge README updated. NEXT: Aug 16 04:00-06:30 AM AKST refresh MMX token; Aug 16 4:00 PM GENERATION DAY (queue 163 — grammar A/B/C both subjects = 6 tracks first, then the four-laws test in real sung audio); extend the tax curve toward the flat-topped wall (N=64/128: where does clip fraction cross 50%?); the census in real audio (does the dividend sign-flip hold for different singers?); composer v5 (read the sign of the seam before naming it — auto-select dip vs hump aim per handoff).