Class v6 family gate, the full report (16:30 BST): 38,000 drawn eras, the lossy-share curve per shape (the exhaustion a 256-shape interaction, about 6x the independent estimate), the pre-floor spread per width and the width-4 floor decision (0.995 kept at its measured cost), the two live-dataset censuses (128 band epochs: 2 hot sets at the 256 shape, both refused by the class v5 floor), the seven known-failed seeds through the sigma form (one value-level test, the bucket bound retired into the bit read; p212 and p225 at bit R), the x4, x8 and x16 verifier rows (x16 outside the band on the 10 ms gate), the band stratum's gate records as JSON lines; every row's log under logs/

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

docs/analysis/class-v6/family-gate.md

@@ -1,6 +1,6 @@
 # Class v6, the family gate: how a parameter family is cryptanalysed as a family
 
-Research lane D of the class v6 rotating-family design, under the Counter ASIC coordinator. First cut landed on the mirror's master at 11:40 BST, 8 October 2026 (fbe62a8b); this cut, with the measured coverage, is the 17:00 BST deliverable; the full report by 09:00 BST on 9 October. Branch `class-v6-family-gate` (this document) and the harness branch `family-gate-v5` (the class v5 crate at 8f481459 plus the census harness, never a chain path). Internal research, not an independent review; nothing here touches a served number, the devnet or the testnet object; no consensus code this week. Every figure carries its source: a lane report on the mirror, a log on a box, or an arithmetic step shown in the text. Figures from memory are labelled approximate.
+Research lane D of the class v6 rotating-family design, under the Counter ASIC coordinator. First cut landed on the mirror's master at 11:40 BST, 8 October 2026 (fbe62a8b); the 17:00 cut with the measured coverage landed at 238100b0 (13:13 BST); this is the full report, pulled to 16:30 BST on the founder's clock (every row measured, a partial carries its count). Branch `class-v6-family-gate` (this document) and the harness branch `family-gate-v5` (the class v5 crate at 8f481459 plus the census harness, never a chain path). Internal research, not an independent review; nothing here touches a served number, the devnet or the testnet object; no consensus code this week. Every figure carries its source: a lane report on the mirror, a log on a box, or an arithmetic step shown in the text. Figures from memory are labelled approximate.
 
 The question. Class v6 (the founder's word at 11:1x BST, 8 October 2026) draws per era the parameters a release used to fix (layer 1), lets the dataset track the chain state (layer 2), rotates instruction families by height (layer 3) and generalises the acceptance floor and the uniformity test to each era's draw with a redraw on failure (layer 4). A chip is then built against a FAMILY of hashes, not one hash. The in-house pass of 7 October and the attack board F1 to F10 bounded ONE member of the family (class v4 sub-version 3 at 017e7037 and class v5 at 1c420786). This document says how the family is bounded: what the space is, how many eras must be drawn to say anything about the worst era, which tests the chain runs itself on every era and epoch, what the offline board looks like over the testnet period, what a published proof of testing contains, and what the whole apparatus cannot see.
 
@@ -14,7 +14,7 @@ The question. Class v6 (the founder's word at 11:1x BST, 8 October 2026) draws p
 | The per-era tests (section 3) | Three rings. Ring A, the chain runs per era at the cut, microseconds: the structural redraw rules (the band membership of every draw, the day-key cost rule, the product-bias rotation class recorded). Ring B, the chain runs per epoch, seconds: the acceptance rule generalised to the family's shape ((a) (b) (a') (c) (c') (c'') (c''') plus the two next-class tests, the per-site largest-256-item-bucket bound and the index-bit bias read, on the same 2^20 pass). Ring C, the gate runs offline per drawn era, box-hours: the F8-form census at 2^24 on 64 seeds, the attempts census, the exhaustion count. Each row carries its known-failed shape and its clean-seed cost so the redraw is priced |
 | The board (section 4) | F1 to F10 and the nine lanes classified per-era, per-family or chain-run, with the harness taking the era draw as an input; the proof of testing is one JSON record per drawn era (the draw, the object, the binary, the seeds, every verdict with its statistic, threshold and log sha256, the core-seconds) and one family summary (eras per stratum, the worst era per test, the refuse rate per band, the bound's arithmetic) |
 | The honest line (section 5) | The gate sees what its tests see. It does not see the diffuse era-stride class below the bit-bias read's resolution (a bias under 0.3 percent at 2^20), a mechanism that lives in a joint corner no stratum names, a mixer weakness below adv-mixer-3's bands, or an era whose draw sits in the untested 1 in 64. A GPU-like chip loses nothing on any era; what a corner buys it is a per-era hot set bounded at 1.002x (the floor) or 1.0024x (the diffuse class), or, at `m = 4`, a mixer with 2 applications of measured margin instead of 6 |
-| Measured (section 6) | 24,000 drawn eras through the ring-B tests in 54 core-hours (1.7 box-hours) on build-1 and build-3, 11:16 to 12:13 BST, the harness as a diff in `logs/`. Four measured rows move the design: the lossy corner of B = 4 exhausts the 256 cap in 1.2 percent of eras (61 of 5,000) against 0 of 19,000 elsewhere, so the band is B = 4 on the injecting families with or/mul/mulhi never raised (0 of 3,000 under it); the (c''') floor at 0.995 costs 4.5 to 6.5 percent of candidates at width 4 against 0.6 to 1.0 at width 1 (a per-width calibration); the shape axis moves the draw from 1.8 attempts (64 x 108) to 3.8 (256 x 27); the era-stride bias sits at over 6 sigma on one address bit (bit R in 73 percent of cases) in about half the eras of every stratum and at over 100 sigma in a third, falling to 7.5 percent under R in 28..31, a structural fact for the research lane's layer 4 rather than a per-epoch refusal |
+| Measured (section 6) | 38,000 drawn eras through the ring-B tests on four boxes (about 90 core-hours), the lossy-share curve per shape, two live-dataset censuses at 2^24, the pre-floor spread per width, the seven known-failed seeds through the sigma form; the rows that move the design: the band is B = 4 on the injecting families with or/mul/mulhi never raised (the lossy corner exhausts the 256 cap on the 256 x 27 shape alone, 3.3 to 3.6 percent of its eras at r = 0.98, about 6x the independent-attempt figure; 0 of 5,000 shape-64 eras at any lossy share); the (c''') floor stays at 0.995 at the pinned width 4 at its measured cost (14.6 percent of the candidates reaching the 2^20 pass, +0.3 attempts per epoch) because the floor refused both hot sets of the live point-A census (p38 at 0.9932, p54 at 0.9945) and the width-4 statistic has a real tail no single floor removes at the width-1 cost; the shape axis moves the draw from 1.8 attempts (64 x 108) to 3.8 (256 x 27); the mixer axis is invisible to every per-era statistic and is closed per family; the era-stride bias at bit R (the product's bit 0 at z = -512 exactly) in half the eras of every stratum is one value-level mechanism that the bucket bound sees only at narrow windows above bit 12, so the family carries ONE test, the per-site index-bit read, as a per-era record and the structural lever's known-failed set (p4, p8, p10, p34, p212, p225, p15 all caught at 6 sigma; a refusal band of 300 sigma catches five and redraws 16 percent of epochs; 6 sigma would redraw half) |
 
 ## 1. The parameter space
 
@@ -22,9 +22,9 @@ The question. Class v6 (the founder's word at 11:1x BST, 8 October 2026) draws p
 
 | Axis | Band proposed for the draw | What fixed the band (the measured rows) | The tested margin inside the band | Known-failed corner, named |
 |---|---|---|---|---|
-| Mixer multiplier `m` (applications per round, `LoadClass::mixer_mult`; 72 per item at `m = 8`, 8 between dependent reads) | {4, 8, 16}; the research lane's outline; this lane's reading: {8} plus {4} as a corner, {16} only after the verifier row | x4 and x8 measured (`docs/plans/mixer-x4.md` section 6.4: 1.92 and 2.79 ms per unit on a loaded M5 Max core; the daily build 42 ms at x8 on the 5090, measured); x16 estimated at about 3.7 ms per unit on the M5 Max core and 9 ms on a 2019-class core (chip-model-v3 section 3 item 1), against the 10 ms verifier gate | adv-mixer-3 (`report-mixer-3.md`, "The round margin, stated"): no statistic survives 2 of the 8 keyed applications between reads at bands 0.0015 (Q2, 2^24 states), 0.00026 (2^27), 0.00018 (Q2b, 2^28); SAT solved at k = 1 in 137 s and timed out at k = 2, 3, 4; two real days and eight random days agree. So at `m = 8` 6 of 8 applications are margin; at `m = 4` the same rows leave 2 of 4; at `m = 16` 14 of 16 | None measured. The corner is `m = 4` by margin and `m = 16` by the verifier; the all-ROT-equal day key is the day-level corner (rejected by the AP-F4-1 rule) |
+| Mixer multiplier `m` (applications per round, `LoadClass::mixer_mult`; 72 per item at `m = 8`, 8 between dependent reads) | {4, 8} on the measured verifier rows (section 6.7: x16 reads 7.13 ms per warp on one core against x8's 4.12, 1.73x, which puts it over the 10 ms gate on the 2019-class core and the half-core proxy by scaling); the research lane's outline had {4, 8, 16} | x4 and x8 measured (`docs/plans/mixer-x4.md` section 6.4: 1.92 and 2.79 ms per unit on a loaded M5 Max core; the daily build 42 ms at x8 on the 5090, measured); x16 estimated at about 3.7 ms per unit on the M5 Max core and 9 ms on a 2019-class core (chip-model-v3 section 3 item 1), against the 10 ms verifier gate | adv-mixer-3 (`report-mixer-3.md`, "The round margin, stated"): no statistic survives 2 of the 8 keyed applications between reads at bands 0.0015 (Q2, 2^24 states), 0.00026 (2^27), 0.00018 (Q2b, 2^28); SAT solved at k = 1 in 137 s and timed out at k = 2, 3, 4; two real days and eight random days agree. So at `m = 8` 6 of 8 applications are margin; at `m = 4` the same rows leave 2 of 4; at `m = 16` 14 of 16 | None measured. The corner is `m = 4` by margin and `m = 16` by the verifier; the all-ROT-equal day key is the day-level corner (rejected by the AP-F4-1 rule) |
 | Op-mix weights (the ten non-load families, sum 75) | Each weight within B points of table 1.4.2 (add 12, xor 10, mul 8, mad 8, shfl 8, rotl 7, sub 6, mulhi 6, rotr 6, or 4), renormalised by largest remainder; shfl and mulhi never raised; B = 2 (spec 1.13.1, proposed) or B = 4 (the research lane) | The microbench (`counter-asic-4-research.md` 15.1a, the 5090 alone, 8 October): pJ per counted op at the stock clock and the 1,300 MHz lock: arx 11.3 / 6.2, mul 13.9 / 8.3, mulhi 39.6 / 21.0, prmt 22.3 / 11.5, lop3 24.1 / 13.0, shfl 55.8 / 29.4; the shuffle is 4.9x the add, so a draw that raises shfl raises the card's premium per instruction with no better `k` for a chip (15.1b). F7's era census: the op-weight corners (15 to 31 of 75) read 1.0x against the GPU, 0 memory effect | The acceptance's rejection rate moves with the lossy share: at the table 0.68 per candidate ((a') 83.5 percent of rejections; adv-accept row 90, 20,000 seeds); the F7 census at B = 2 read every corner clean for the chip. MEASURED today (section 6): at B = 4 with the lossy families free to rise the random stratum reads r = 0.72 and a maximum attempt of 107 (the shipped stream's maximum is 28); at the lossy corner r = 0.956 and 1.2 percent of eras EXHAUST the 256 cap | The lossy corner: or, mul and mulhi all at +B (30 of 75 against the table's 18). It is OUTSIDE the band: 8 of 663 drawn eras reached the last resort, which fails rule (a) in 9 percent of seeds (adv-accept-3 finding 1). The band this lane carries: B = 4 on the seven injecting families, or/mul/mulhi never raised above their base (the `lossybase` stratum of section 6 is its measurement), with ring A's lossy-share rule as the check |
-| Read width `W` (words per load) | {1, 4} | `docs/plans/read-width.md` (5 October): w16 within the rules (5090 139.8 against 136.1 MH/s, 9070 XT 17.90 against 18.15, M5 Max within 1 percent); w64 bandwidth-bound (5090 share 0.58, 71.9 MH/s): out; the per-load mixes out on the 5 percent spread rule | The acceptance runs unchanged on aligned addresses (read-width.md section 2), but (c'') and (c''') read a site's distinct ALIGNED indices against `E_s = N - N^2 / 2W_s`: at `W = 4` the index space is `W_s / 4`, so the expectation must follow or every width-4 program reads 0.976 at a quarter window and is refused by the 0.98 floor. The harness divides the window by the width (`site_window_words`); the spec line is owed | `W = 16` (w64): excluded by measurement, never in the band |
+| Read width `W` (words per load) | {1, 4} in this census; the research lane's floor reading (13:2x BST, `class-v6-rotating-family.md` section 10.3) pins W = 4 words at genesis (the width is the one wire lever on the SRAM full-store die and the cards pay nothing for 16 bytes), with W = 8 the next value once the crate carries it, so the axis is a one-value band at genesis and the width-4 strata here are its calibration | `docs/plans/read-width.md` (5 October): w16 within the rules (5090 139.8 against 136.1 MH/s, 9070 XT 17.90 against 18.15, M5 Max within 1 percent); w64 bandwidth-bound (5090 share 0.58, 71.9 MH/s): out; the per-load mixes out on the 5 percent spread rule | The acceptance runs unchanged on aligned addresses (read-width.md section 2), but (c'') and (c''') read a site's distinct ALIGNED indices against `E_s = N - N^2 / 2W_s`: at `W = 4` the index space is `W_s / 4`, so the expectation must follow or every width-4 program reads 0.976 at a quarter window and is refused by the 0.98 floor. The harness divides the window by the width (`site_window_words`); the spec line is owed | `W = 16` (w64): excluded by measurement, never in the band |
 | Shadow block shape (instructions per block, passes per iteration) | {64 x 108, 128 x 54, 256 x 27}: 6,912 shadow instructions per iteration at rung 0 | 64-instruction blocks ran 2.5 to 3.5 percent faster than 256 on the 5090 and the M5 Max (6 October, measured); 1,024 cost the M5 Max 17 percent (out) | F1's shadow redundancy bound (3.0 percent of peephole-removable instructions, the v5 generator's `SHADOW_REMOVABLE_MAX_PERMILLE`) was set from F1's census at 256 instructions; the removable fraction per block is a property of the block's length, so the fraction must be re-read at 64 and 128 (a per-family row). The acceptance's (a') fixpoint and the shadow's execution in (c) take any length | The per-load placement (16 sub-blocks of 16 after each load): dead as a chain class (`counter-asic-4-research.md` 20.2a-close): 22 of 1,621 candidates accepted, 42 of 64 seeds exhaust 32 attempts, candidate 0 carries index bit 0 set in 40 of 1,024 addresses at site 0 (z 29.5). The placement axis therefore has ONE value in the band |
 | Program length `N` (counted ops per hash) | Not drawn: the ladder's signal (rungs 0 to 2); the research lane's outline keeps it so | F6: the worst of 10^5 class v4 programs 8.708 ms on the half-core proxy, so N about 300,000 fits under 10 ms (rung 2 admissible, rung 3 not); the M5 Max loses 3.3 to 10 points across the band (measured ladder) | The verifier's worst case was censused at rung 0 only; a draw inside the ladder needs the 10^5-program worst case at the top rung (per-family row) | Rung 3 (inadmissible by F6) |
 | Era stride and interleave (`M` odd, `R` in 1..31, `pos` four ascending bit positions, the fold rotations, the per-site windows `k_off`, `o`) | As spec 1.13.1 draws them today | F7's 2^24 era census (stride bijective, R, pos, M uniform, no class over 1.1x); adv-cache-2 section 2.3: the same 32 base programs read 2 of 32 with a site over 1.04x under `R = 29` and 16 of 32 under drawn eras (8 over 1.2x, worst 1.75x), because `rotl(x * M, R)` places a product's biased low bits (P(bit 0) = 1/4) at address bits R and up, inside the 28-bit index unless R is 29 or 30; at D = 29 only R = 30 and 31 cut bit 0 | The whole 64-seed F8 census is a census over drawn eras (each f8 seed carries its own era), so the per-era uniformity verdicts of sub-version 3 and class v5 (60 of 64 and 61 of 64 under 1.2x) already sample this axis; the devnet's clean reading at R = 29 is the accident, the drawn-era figure the chain's | R in 1..27 with a product as the last writer of a load's source (sub-class A); the warp-uniform mulhi source (sub-class B, 1 in 16,000 warps) |
@@ -119,7 +119,7 @@ The three rings. A test's ring says who runs it and what a failure costs.
 | (c''') the hot-item floor at 0.995 | the same `E_s`; the floor's clean spread re-read per stratum (section 6): a fixed 0.995 under a spread that moves with the shape or the width is either a liveness cost or a miss | seed 100767 (0.9919) and the nine live hot sets | 2.435 percent of accepted class v4 programs, +0.045 on the mean attempt | in the (c'') pass |
 | The per-site largest-256-item-bucket bound (NEW, the F8 tail's catch) | the largest count over a site's 4,096-word buckets in its window against `N / (W_s / 4096)`, in SIGMA of the bucket's expectation (`(max - E) / sqrt(E)`), refused above a band set from the clean spread. The ratio alone cannot carry the bound: a clean full-window site's largest of 65,536 Poisson(16) buckets reads 2.1x at +4.4 sigma, a quarter-window site's largest of 16,384 Poisson(64) buckets 1.5x at +4 sigma (the 16-era smoke run read ratios 2.2 to 2.4 on clean sites); the four tail seeds' 3.1x to 5.6x at narrow windows are +17 to +37 sigma | p4, p8, p10, p34 | measured in section 6 (the columns `bucket_z_max`, `bucket_win`) | one linear pass over the (c'') indices |
 | The index-bit bias read (NEW, adv-cache-2's value-level test) | per site, the one-count of each free index bit (bits above the width's alignment and below the window's cut) over the 2^20 evaluations, in sigma (`sigma = sqrt(N) / 2 = 512`); a 6-sigma band is a bias of 0.3 percent at 2^20 (the product law's bit 0 at 25 percent reads z about 512, bit 1 at 3/8 z 256, bit 3 at 3.1 percent z 64, bit 6 at 0.4 percent z 8, bit 7 at 0.2 percent z 4: passes). MEASURED on the first 16 drawn eras (11:21 BST, section 6): 7 of 16 accepted programs carry one site with a bias of 130 to 511 sigma at address bit R or R + 1 (the era's rotation), the era-stride class exactly, on programs (c''') passes at 0.9954 to 1.0000; the 9 others read under 3.8 sigma. So as a REFUSAL the read would redraw about 40 percent of epochs; its right use is as a record per era plus the structural lever (fold the product's low bits in `load_index` before the rotation, a next-class item), and the chip price is stated in section 5 | Devnet 3 site 0, era-drawn-28 site 15, the per-load candidate 0; and now the 7 of 16 | 7 of 16 at 6 sigma (a refusal is not the form); 0 of 16 at a 600-sigma band | one pass over the indices, 28 bit tests each |
-| The cap and the last resort | unchanged at 256; class v5 carries the verified construction (`last_resort_v5`: a scan of 256 more candidates past the cap, each rewritten and its stale loads re-sourced, the first that passes the whole rule; the unchecked fallback behind it), so adv-accept-3's finding 1 (sub-version 3's unchecked last resort failing (a) in 9 percent of seeds) is closed on the family's crate. What the family changes: the scan's stated odds (under 1e-300) assume independent attempts, and section 6 measures a per-era exhaustion rate 1,000x the independent estimate at the lossy corner, so the fallback's odds are the corner's, not the arithmetic's; ring A keeps the corner out of the band | the reject-everything mirror; the 61 exhausted eras of section 6 (the scan's verdict on each is the full report's item 5) | 0 reached in 1.02 x 10^6 seeds of the shipped draw; 0 of 19,000 family eras outside the lossy corner | n/a |
+| The cap and the last resort | unchanged at 256; class v5 carries the verified construction (`last_resort_v5`: a scan of 256 more candidates past the cap, each rewritten and its stale loads re-sourced, the first that passes the whole rule; the unchecked fallback behind it), so adv-accept-3's finding 1 (sub-version 3's unchecked last resort failing (a) in 9 percent of seeds) is closed on the family's crate. What the family changes: the scan's stated odds (under 1e-300) assume independent attempts, and section 6.4 measures a per-era exhaustion rate about 6x the independent estimate at the lossy corner on the 256-instruction shape (the attempts of one era share its weight table), so the fallback's odds are the corner's, not the arithmetic's; ring A keeps the corner out of the band, and on every exhausted era seen (41) the scan passed at its first to third candidate | the reject-everything mirror; the 61 exhausted eras of section 6 (the scan's verdict on each is the full report's item 5) | 0 reached in 1.02 x 10^6 seeds of the shipped draw; 0 of 19,000 family eras outside the lossy corner | n/a |
 
 ### 3.3 Ring C: the offline tests per drawn era
 
@@ -178,18 +178,18 @@ What the family gate cannot see, and how a chip would use it.
 
 | Blind spot | Why the gate misses it | What a chip gets from it | Bound today |
 |---|---|---|---|
-| The era-stride class, measured today at the bit level on 7 of 16 drawn eras (section 3.2): not diffuse, a 25 percent bias on ONE address bit (bit R, the product's bit 0 through `rotl(x * M, R)`) at one site, on programs every floor passes | the (c''') floor reads distinct counts, which a one-bit bias barely moves (0.9954 to 1.0000 on the seven); the bit read sees it at 130 to 511 sigma but refusing it would redraw about 40 percent of epochs; the class is structural (`load_index`'s form and the weights' products), so the catch is a next-class change to the address form, not a per-era floor | a store holding the favoured half of that site's window (128 MiB at D = 28) serves 75 percent of that site's reads instead of 50: 25 percent of 1/16 of a hash's reads, about 1.6 percent of reads at f = 1/2 from one site in about 40 percent of eras (arithmetic on the measured bias); the honest partial-store curve already costs 1.26x the ops at f = 1/2 (adv-cache-2 Q4), so the f = 1 verdict stands; the top-0.1-percent reading of the ledger (1.0024x) was the same class seen through the coarser instrument | the per-era record (ring A names the rotation class; ring B logs the bit read); the structural fix is the family's first owed item for the research lane |
+| The era-stride class, measured at the bit level on half the eras of every stratum (sections 6.2 and 6.8): not diffuse, a 25 percent bias on ONE address bit (bit R, the product's bit 0 through `rotl(x * M, R)`, z = -512 at 2^20 exactly) at one site, on programs every floor passes; the morning's F8 tail and the hash lane's p212 and p225 are the same mechanism | the (c''') floor reads distinct counts, which a one-bit bias barely moves (0.9954 to 1.0000 on the seven); the bit read sees it at 130 to 511 sigma but refusing it would redraw about 40 percent of epochs; the class is structural (`load_index`'s form and the weights' products), so the catch is a next-class change to the address form, not a per-era floor | a store holding the favoured half of that site's window (128 MiB at D = 28) serves 75 percent of that site's reads instead of 50: 25 percent of 1/16 of a hash's reads, about 1.6 percent of reads at f = 1/2 from one site in about 40 percent of eras (arithmetic on the measured bias); the honest partial-store curve already costs 1.26x the ops at f = 1/2 (adv-cache-2 Q4), so the f = 1 verdict stands; the top-0.1-percent reading of the ledger (1.0024x) was the same class seen through the coarser instrument | the per-era record (ring A names the rotation class; ring B logs the bit read); the structural fix is the family's first owed item for the research lane |
 | A joint corner no stratum names | the 128 cells pin the axes with a KNOWN mechanism; a mechanism that needs, say, a particular `M` with a particular weight table and a particular shape is reached only by the random stratum, which bounds its fraction at 3/n | one era (180 days) of whatever the corner is worth, at most once per occurrence | 1/64 to 1/256 of eras untested at 95 percent (section 2.1); the chain-run rings catch the known shapes on every era |
 | The mixer below adv-mixer-3's bands | biases under 0.00018 (Q2b), correlations under 0.00073 with multi-bit masks, differentials under 2^-19, any structure a SAT model of 2 or more applications would expose; at `m = 4` the measured margin is 2 applications | a shortcut that saves applications against the 9,360 ops per item; none found at any k | the round margin, stated; `m = 4` is the corner that spends 4 of the 6 measured spare applications |
 | The verifier at `m = 16` and rung 2 on a 2019-class core | unmeasured (an estimate of 9 ms against the 10 ms gate) | nothing for a chip; a node tier retired if the corner is drawn | the row is owed before 16 enters the band |
 | The acceptance's stand-in at D = 28 against a live D = 29 | the windows cap `min(k_off, D - 26)` and the rotation's cut set move with D; the gap is measured at D = 28 only | a per-site concentration the closed form does not reproduce | 0.0004 at D = 28; unmeasured at 29 (ring C row) |
 | The union bound's own looseness | the per-test miss rates rest on 3 to 11 known-failed cases each (section 2.2), so the family's false-pass rate is not yet a number under 1 | nothing new; the bound is on what the gate claims, not on the hash | fixed by cases, not eras: 60 fired cases per test for `delta` under 0.05 |
-| The last-resort path at a corner | class v5's scan hands out a rule-checked program, and its unchecked fallback is priced at under 1e-300 on INDEPENDENT attempts; the lossy corner measured 1.2 percent of eras exhausting against a 1.3e-5 independent estimate, so at a corner the attempts of one era are correlated through the era's own weights and the fallback's odds are unknown there | a program that re-reads one address in two loads of the same hash, if the fallback were ever reached | ring A keeps the corner out of the band (0 of 19,000 eras exhaust elsewhere); the scan's verdict on the 61 exhausted eras is owed (section 7 item 5) |
+| The last-resort path at a corner | class v5's scan hands out a rule-checked program, and its unchecked fallback is priced at under 1e-300 on INDEPENDENT attempts; the lossy corner measured 3.3 percent of 256-shape eras exhausting against a 0.5 percent independent estimate (section 6.4), so at a corner the attempts of one era are correlated through the era's own weights by a factor of several and the fallback's odds there are the measured curve's, not the arithmetic's | a program that re-reads one address in two loads of the same hash, if the fallback were ever reached | ring A keeps the corner out of the band (0 of 19,000 eras exhaust elsewhere); the scan's verdict on the 61 exhausted eras is owed (section 7 item 5) |
 | The era redraw's own grindability | a ring-A redraw consumes the stream's next block, so an adversary who could steer `E_n` could steer which block is used; the era VDF of F7 (517 s on the fastest prover, 259x the window) closes the steering of `E_n` itself | nothing beyond F7's bound | F7 (a) PASS with the VDF in the node |
 
 A GPU-like chip (the `f = 1` stored-dataset chip with a programmable core, chip-model-v3 section 5) loses nothing on any era of the family: every drawn parameter is firmware to it, and what the family costs it is the core sized for the band's top (the research lane's USD 30 to 60 per chip, modelled). The family gate's job is therefore not to defeat that chip (the identity of the research file's section 2 says the per-joule edge is the card's own idle, 2.1x at `k = 1`) but to make sure no drawn era hands ANY chip more than the one hash did: a per-era hot set, a per-era mixer shortcut, a per-era verifier miss. Everything in this document is a bound on that increment, and the increments found so far are 1.002x and 1.0024x.
 
-## 6. The coverage measured (24,000 drawn eras, 11:16 to 12:13 BST, 8 October 2026)
+## 6. The coverage measured (38,000 drawn eras in 6.1 to 6.5, the lossy curve, the live censuses, the sigma form and the verifier rows; 11:16 to 15:3x BST, 8 October 2026)
 
 The harness: branch `family-gate-v5` at 3dc3117c (the class v5 crate 8f481459 plus `accept::tests::family_gate_era_census`; the diff is `logs/harness-family-gate-v5-3dc3117c.diff` (a plain diff of igneum-pow/src: the gate refuses the patch form's author lines) in this directory, since the branch's push to the mirror was refused by the class v5 tree's own pre-push hook and is held). What it does per era `k` of a label space: the era bytes from `igneum-family-gate[/stratum]/era/k` and the epoch bytes from `.../program/k` (the F8 label space's form); from the era's own stream, in a fixed order and consumed whether pinned or not, the shadow block shape in {64 x 108, 128 x 54, 256 x 27}, the mixer multiplier in {4, 8, 16} (recorded: the acceptance never runs the mixer), the ten non-load weights perturbed by `below(2B + 1) - B` with shfl and mulhi never raised and renormalised to 75 by largest remainder (B = 4), the read width through the era draw over {1, 4} words (the mix one-hot on the drawn width); then the chain draw of that era's epoch through the real rule keyed on the family's shape (`is_family_shape`, the acceptance's `is_class_v4_shape` generalised behind `IGNEUM_FAMILY_GATE`; the draw's source rule reads the same predicate), every candidate's first failing part recorded, the cap 256 and the last resort as shipped; on the accepted program, on the rule's own 2^20 sample (4,096 units), per site: the (c'') and (c''') ratio with the window divided by the width, the largest 4,096-word (256-item) bucket in sigma of its expectation, and the largest index-bit one-count excess in sigma over the free bits. One TSV row per era (`logs/<stratum>-family_gate_era_census-<from>-<n>.tsv`, the run's log beside it with the binary's sha256 and the lease line); every binary a pinned copy under `/srv/builds/_adv-family-gate/bin/<tag>/` (fg2 9b7f764a for the first corners, fg5 e7a50f8d for the width-4 and lossy-base strata). Cost: about 10 core-seconds per era; 54.4 core-hours in all (1.7 box-hours of 32-core slots) on build-1 (random, lossy cap, width 4) and build-3 (shape 64, lossy base, width 4 plus lossy cap), every run under `lease pool` at class measure. The control: `v5_attempts_census` on the shipped class v5 draw over 10,000 seeds of the F8 label space (build-1, 24 cores, 1,534 s, binary 4bec799c, `logs/base-v5_attempts_census-1000-10000.log`): r = 0.6854, mean attempt 2.179, 0 exhaustions, (a') 82.5 percent of rejections, (c''') 0.9 percent, the row-90 reading reproduced.
 
@@ -229,7 +229,7 @@ Readings, each a measured row above:
 1. The mixer axis is invisible to every per-era statistic, as section 1.3 said it must be: the three `m` values read the same to the noise (r 0.714 to 0.731, the biased share 51 to 52 percent). `m` is a per-family axis, closed by the mixer rows of section 4, not by drawing eras.
 2. The shape axis moves the draw's cost: 256 x 27 costs 3.82 attempts against 1.81 at 64 x 108 (r 0.792 against 0.645), because the freshness fixpoint (a') walks the whole block and a 256-instruction block holds more lossy last writers per iteration than a 64-instruction one. The band's cheapest shape is the fastest card shape too (64, measured 2.5 to 3.5 percent faster on 6 October).
 3. The width axis moves the (c''') floor's calibration: at width 4 the floor at 0.995 refuses 4.5 to 6.5 percent of candidates against 0.6 to 1.0 at width 1, with the expectation already divided by the width. The clean spread of the ratio at width 4 is not the width-1 spread the floor was set from (the aligned index space is a quarter the size, so the birthday collisions on the closed form are 4x denser, and the `N - N^2 / 2W` form is 0.0007 low at `N / W = 1/16`). A floor fixed at one number across the family is therefore a per-width liveness cost of 4x the (c''') attempts; the family's floor is set per width from the clean spread (the next measurement, section 7 item 1), or the statistic is re-expressed as a sigma over its own expectation as the bucket statistic now is.
-4. The lossy corner breaks the exhaustion bound: r = 0.957 to 0.959 and 1.2 percent of eras at the 256 cap in both lossy strata (61 of 5,000), against 0 of 19,000 in every other stratum and 0 of 10,000 on the shipped draw. The arithmetic agrees: 0.957^256 = 1.3e-5 per seed is the geometric estimate, but the attempts are not independent across one seed's candidates (the lossy weight table is shared by every attempt of the era, so the rejection rate per era is itself drawn from a spread), and the measured per-era rate is 1.2e-2, three orders above the independent-attempt figure. The ring-A rule is the one the data forces: or + mul + mulhi never raised above the table's 18 of 75. Under it (the lossy-base stratum) r = 0.595, the mean attempt 1.47, the maximum 59, 0 exhaustions in 3,000.
+4. The lossy corner breaks the exhaustion bound: r = 0.957 to 0.959 and 1.2 percent of eras at the 256 cap in both lossy strata (61 of 5,000), against 0 of 19,000 in every other stratum and 0 of 10,000 on the shipped draw. Read per shape (section 6.4) the exhaustion sits on the 256 x 27 shape alone (3.3 to 3.6 percent of its eras; 0 of 5,000 at 64 x 108), where r = 0.98 and the independent-attempt figure 0.98^256 is 0.5 percent: the attempts of one era are correlated through its weight table by about 6x. The ring-A rule is the one the data forces: or + mul + mulhi never raised above the table's 18 of 75. Under it (the lossy-base stratum) r = 0.595, the mean attempt 1.47, the maximum 59, 0 exhaustions in 3,000.
 5. The index-bit bias is the family's standing fact: 48 to 58 percent of accepted programs in every stratum carry one site with a bias over 6 sigma at 2^20, and 73 percent of those are at bit R exactly, 12 percent at R + 1, 5 percent at R + 2 (the product law's bits 0, 1, 2 at P = 1/4, 3/8, 7/16 through `rotl(x * M, R)`); 74 percent read P(bit) under 1/2 (a product's low bit), 26 percent over 1/2 (an or-shaped source, 5/8 and up). Under R in 28..31 (bit 0 cut by the 28-bit mask) the share over 100 sigma falls from 37.7 to 7.5 percent and the bucket excess's p99 from +44 to +6.5 sigma, which is adv-cache-2's R-class reading at 10,000 eras. The shipped devnet's R = 29 is one of the four clean rotations; the chain's eras are drawn from all 31. The bucket statistic and the bit statistic see the same mechanism (the bucket's p99 on the bit-clean eras is +6.8 sigma, on the biased ones +44), so one next-class test covers both: the value-level read, with the structural fix in `load_index` as the remedy rather than a refusal (a 6-sigma refusal redraws half the epochs; a 100-sigma one a third).
 6. The (c''') floor holds everywhere by construction (0 accepted programs under 0.995) and its clean spread at width 1 is the shipped one (p1 of the minimum ratio 0.9958 to 0.9961, median 0.9999 across strata); the floor's refuse rate moves from 2.7 percent (random) to 3.9 (the band) and 0.4 (the lossy corner, where (a') refuses first).
 
@@ -237,25 +237,94 @@ Readings, each a measured row above:
 
 | Claim | n | Bound at 95 percent | Bound at 99 percent |
 |---|---|---|---|
-| An era of the proposed band whose chain draw exhausts the 256 cap | 3,000 of 3,000 passed (lossy base) | under 1.0e-3 of eras (one in 1,000: one per 490 years at 180-day eras) | under 1.5e-3 |
+| An era of the proposed band whose chain draw exhausts the 256 cap | 9,000 of 9,000 passed (lossy base 3,000, the width-1 and width-4 band strata 3,000 each; the afternoon's strata) | under 3.3e-4 of eras (one in 3,000: one per 1,500 years at 180-day eras) | under 5.1e-4 |
 | The same on the random stratum at B = 4 on every family | 10,000 of 10,000 | under 3.0e-4 | under 4.6e-4 |
-| An era of the band whose accepted draw sits under the (c''') floor | 3,000 of 3,000 (and 19,000 of 19,000 outside the lossy corner) | under 1.0e-3 | under 1.5e-3 |
+| An era of the band whose accepted draw sits under the (c''') floor | 9,000 of 9,000 (and 33,000 of 33,000 outside the lossy corner, every stratum of the day) | under 3.3e-4 | under 5.1e-4 |
 | An era of the lossy corner that exhausts | 61 of 5,000 FAILED | the rate is 1.2 percent, a measurement, not a bound: the corner is out of the band; class v5's last-resort scan then hands each such era a rule-checked program (the scan's own verdicts on the 61 are owed) | |
-| The (c''') floor's miss rate on the few-item hot-set class | 9 of 9 fired (unchanged: this census reads the acceptance's sample, not the live dataset) | under 0.33 | under 0.40 |
-| The bit-bias read's firing on the product class | 5,172 of the random stratum's eras read over 6 sigma; the three named cases (Devnet 3 site 0, era-drawn-28 site 15, the per-load candidate 0) are of that class | a record, not a refusal: no miss rate is claimed | |
+| The (c''') floor's miss rate on the few-item hot-set class | 11 of 11 fired (the in-house pass's nine, plus point A's p38 and p54 under a band era's weights, section 6.6) | under 0.27 | under 0.34 |
+| The bit-bias read's firing on the product class | 7 of 7 named seeds at 6 sigma (p4, p8, p10, p34, p212, p225, p15; section 6.8) and the three attributed sites of the pass; 5 of 7 at 300 sigma | under 0.35 at 6 sigma on the named set; the read is a record, not a refusal, so no miss rate is claimed for a chain rule | |
 
 The union bound stays loose where section 2.2 said it would: the two floors' miss rates on the live-dataset classes rest on nine and five cases, and this census adds cases to the exhaustion and bias rows only. The live-dataset census at 2^24 on 64 seeds of two eras of the band (one at width 4, one at R in 1..27 with a 1,000-sigma site) is the next ring-C row (section 7 item 6); it is what turns the bit read's z into a per-era hot-set price on the real item map.
 
-## 7. Owed by 09:00 BST on 9 October (the full report)
+### 6.4 The lossy-share curve, per shape (build-1, 12:5x to 13:2x BST; 3,000 eras per point, every other axis drawn)
 
-1. The (c''') floor's clean spread per width (and per shape) from the accepted draws of section 6, with the per-width floor or the sigma form that keeps the clean refuse rate near the shipped 2.4 percent; the spec line for 1.4.6.5's expectation under a drawn width.
-2. The bucket bound stated in sigma from the bit-clean spread (p99 +6.8, max +28 over 4,828 eras): a band at +8 sigma refuses under 1 percent of bit-clean programs and every F8 tail seed (+17 to +37); its cost on the biased eras is the bit read's cost, so the two tests are one rule.
-3. The lossy-corner reading taken one step further: the exhaustion rate per era against the lossy share in points (18 to 30 of 75), so the band's edge is a measured curve, not one corner; 10^4 seeds at each of three shares.
-4. The mixer rows at `m = 4` and 16 (adv-mixer-3's index and sac at k = 2 on two days; the SAT at k = 2) as the per-family corner rows, on build-4 when its pool frees; else the plan with its hours.
-5. The class v5 last-resort scan on the 61 exhausted eras (the first passing k past the cap, or the fallback), and the lossy-share curve at +1, +2, +3 and +4 points (3,000 eras each, running on build-1 from 12:5x BST with the fg6 harness).
-6. The F8-form census at 2^24 on 64 seeds of two eras of the band (one at width 4, one with a 1,000-sigma site under R in 1..27), the first ring-C rows on the family (about 2 box-hours each): the live-dataset hot-set price of the bit-level bias.
-7. The gate-record JSON per era (the format of 4.3) generated from the TSV rows by `fg-records.py` (in `logs/`), and the family summary.
-8. The two build-4 strata (shape 64 plus lossy cap, shape 256) when its pool frees; nothing in the cut depends on them.
+The band's edge measured as a curve rather than a corner: or, mul and mulhi each raised by the same number of points over the table (the lossy share of the 75 non-load points from 18 to 30), the seven injecting families drawn within B = 4 as before. `r` is the per-candidate rejection, `r^256` the exhaustion figure the independent-attempt arithmetic of spec 1.4.6.6 would give, the measured column what the eras did.
+
+| Lossy points added | Lossy share of 75 | Shape | Eras | r | r^256 (independent attempts) | Exhausted, measured | Mean attempt | Max | Past attempt 31 |
+|---|---|---|---|---|---|---|---|---|---|
+| +0 (the band) | 18 | 64 x 108 | 980 | 0.550 | 4e-67 | 0 | 1.22 | 14 | 0.0 percent |
+| +0 | 18 | 128 x 54 | 963 | 0.562 | 1e-64 | 0 | 1.28 | 12 | 0.0 |
+| +0 | 18 | 256 x 27 | 1,057 | 0.651 | 2e-48 | 0 | 1.87 | 59 | 0.2 |
+| +1 | 21 | 64 | 1,009 | 0.716 | 7e-38 | 0 | 2.52 | 46 | 0.2 |
+| +1 | 21 | 128 | 1,015 | 0.768 | 4e-30 | 0 | 3.31 | 47 | 0.1 |
+| +1 | 21 | 256 | 976 | 0.867 | 1e-16 | 0 | 6.50 | 142 | 2.3 |
+| +2 | 24 | 64 | 1,010 | 0.772 | 2e-29 | 0 | 3.39 | 52 | 0.4 |
+| +2 | 24 | 128 | 991 | 0.841 | 5e-20 | 0 | 5.29 | 55 | 1.4 |
+| +2 | 24 | 256 | 999 | 0.928 | 4e-9 | 2 (0.20 percent) | 12.29 | 249 | 8.8 |
+| +3 | 27 | 64 | 1,012 | 0.829 | 1e-21 | 0 | 4.85 | 58 | 1.2 |
+| +3 | 27 | 128 | 980 | 0.885 | 2e-14 | 0 | 7.66 | 73 | 4.1 |
+| +3 | 27 | 256 | 1,008 | 0.958 | 2e-5 | 6 (0.60 percent) | 21.25 | 241 | 19.6 |
+| +4 (the corner) | 30 | 64 | 1,004 | 0.872 | 6e-16 | 0 | 6.83 | 56 | 2.5 |
+| +4 | 30 | 128 | 998 | 0.924 | 2e-9 | 0 | 12.21 | 117 | 7.7 |
+| +4 | 30 | 256 | 998 | 0.979 | 5e-3 | 33 (3.31 percent) | 38.53 | 244 | 36.7 |
+
+Readings. (1) The exhaustion is a shape-256 interaction, not a corner-wide one: no era of shape 64 x 108 exhausted at any lossy share (0 of 5,015 across the five points, plus 0 of 2,000 on build-3's shape-64 lossy-cap stratum and 0 of 646 at width 4), one of 4,947 at 128 x 54 (the first corner stratum), and every other exhaustion at 256 x 27. The 256-instruction block holds three times the lossy last writers of a 64-instruction one, and (a')'s fixpoint walks the whole block. (2) The morning's "1,000x above the independent estimate" was the mixed-shape average read against the mixed-shape `r`; per shape the gap is about 6x (0.979^256 = 0.5 percent against 3.3 percent measured at 256 x 27 and +4; 1.7e-5 against 0.6 percent at +3), so one era's attempts are correlated through its weight table by a factor of several, not thousands. The arithmetic of 1.4.6.6 is therefore an under-estimate by that factor at any lossy share, which the band's rule absorbs by holding the share at 18 (where `r` is 0.55 to 0.65 and `r^256` is under 1e-48). (3) The band's edge: at +2 on the 256 shape 8.8 percent of eras pass attempt 31 and 2 of 999 exhaust; the ring-A rule "lossy families never raised above their base" is the measured line, and the research lane's layer 1 carries it (section 5.1b of `docs/design/class-v6-rotating-family.md`). (4) Every exhausted era of every stratum took class v5's last-resort scan and passed at its first candidate past the cap (k = 256) in 38 of 41 cases, at k = 257 or 258 in the other three: the rewrite plus the stale-load repair hands the chain a rule-checked program at once, and the unchecked fallback was never reached.
+
+### 6.5 The pre-floor spread per width and the width-4 floor (W = 4 words pinned at genesis by the research lane)
+
+Measured on the fg7 harness (the ratio of every candidate the two floors refuse, beside the accepted draw's), 3,000 eras per width under the band (build-2 `logs/w1band-*`, build-3 `logs/w4band-*`):
+
+| Width | Candidates | Reaching the 2^20 pass (every earlier part passed) | (c'') refused at 0.98 | (c''') refused at 0.995 | Pre-floor minimum-site ratio of the reaching candidates: p1 / p2.5 / p5 / p10 / median | A floor at 0.990 / 0.993 / 0.995 / 0.996 refuses (percent of reaching candidates) |
+|---|---|---|---|---|---|---|
+| 1 word | 6,920 | 3,093 | 38 (1.2 percent) | 55 (1.8 percent) | 0.968 / 0.993 / 0.9961 / 0.9978 / 0.9999 | 1.8 / 2.4 / 3.0 / 3.9 |
+| 4 words | 8,270 | 3,646 | 113 (3.1 percent) | 533 (14.6 percent) | 0.912 / 0.971 / 0.985 / 0.991 / 0.9999 | 9.5 / 15.1 / 17.7 / 19.0 |
+
+Reading. At width 4 the statistic itself moves, not just its tail: a tenth of the candidates that reach the 2^20 pass read a minimum-site ratio under 0.991, against 2 percent at width 1, with the expectation already divided by the width. The aligned index drops the two low address bits, so a site whose source carries its entropy in the low bits reads full at width 1 and concentrated at width 4; the shift is in the programs, not in the formula (the `N - N^2 / 2W` form is 0.07 percent low at `N / W = 1/16`, nothing like the 1 to 9 percent seen). So no single number keeps the clean refusal near the shipped 2.4 percent at width 4: a floor at 0.990 still refuses 9.5 percent of reaching candidates (3.8 percent of all candidates), and the sigma-over-expectation form would re-scale the same distribution without moving its tail. The decision this lane ships for the 09:00 report: keep (c''') at 0.995 at width 4, with its measured cost (14.6 percent of reaching candidates, 6.5 percent of all candidates, +0.3 attempts per epoch, about 0.7 s per epoch per node at 2.2 s per candidate), because the floor's job is soundness and the known-failed hot sets of point A sit at 0.9932 and 0.9945 (section 6.6), inside the band a lower floor would admit; and read the 0.990 to 0.995 band at width 4 on the live dataset (a ring-C row, the census at 2^24 on 64 width-4 seeds) before any floor moves, since whether those candidates are hot sets or thin spreads is exactly the question adv-accept answered at width 1 (7 of 12 deep seeds were thin spreads) and nobody has answered at width 4.
+
+### 6.6 Ring C: the live-dataset census at two band points
+
+Point A DONE (build-1, 11:3x to 13:49 BST, 4,783 s on 24 pool cores, about 32 core-hours; `logs/f8-A.log`): attack-f8's 64-seed census (p2 to p65 of the F8 label space, each seed its own era, 2^24 nonces each, the window-model control, the by-site attribution) built against the family crate with the band era 49's weight table (add 17, xor 6, mul 4, mad 13, shfl 9, rotl 9, sub 2, mulhi 2, rotr 9, or 4) at shape 256 x 27 and width 1. The class-v5 tree's copy of attack-f8 predates the `--class v5` path, so the draws are generator 4 under the family's shape predicate: sub-version 3's rule with (c'') at 0.98 and WITHOUT class v5's (c''') floor at 0.995. Result: 45 of 64 PASS with no test fired; 3 beyond 1.2x of the window model at the top 0.1 percent (p38 1.678x, p54 1.589x, p4 1.297x; the shipped class read 4 of 64 over, p4 among them at 1.217x); HOT SETS in 2 (p54: X at 0.1 percent +0.120, X at 1 percent +0.684; p38: +0.102 and +0.576), against 0 of 64 on the shipped class; windowed 64-item buckets beyond 6 sigma in 18 of 64 (the bit-level class of section 6.2 seen on the live item map: p38's largest bucket +106 sigma, p4's +84), against the shipped class's 4. The class v5 floor's read on the same 64 programs (`v5_hot_census` under the same weight table, build-3, queued at 13:5x BST) says whether (c''') refuses p38 and p54 as it refused the nine hot sets of the in-house pass; LANDED 14:3x BST (build-3, 79 s on 8 threads, `logs/floor-read-A.log`): the class v5 floor at 0.995 refuses 3 of the 64 class v4 draws, p38 at 0.9932 (site 0), p54 at 0.9945 (site 12) and p11 at 0.9933, so BOTH hot sets of point A are refused and the class v5 draw moves to the next attempt (min ratio 0.9999 on each); p4 at 0.9963 passes, as it did on the shipped class. The floor does under the family's weights what it did on the shipped class: nine of nine then, two of two now. What a chip gets from p54 as it stands (on the class v4 rule; under class v5 it never runs): the top 1 percent of items (16 MiB) holding 0.68 percent more of the reads than the window model, the 1.002x class of the ledger, no row of the chip model moves; what an auditor sees is two hot sets in 64 epochs where the shipped class had none, which is the per-era weight axis showing on the item map and the reason the floor is a per-epoch rule and not a per-class census.
+
+Point B (build-3, band era 5's weights 11,15,6,9,7,11,4,6,2,4 at shape 64 x 108; `logs/f8-B.log`): PENDING at 16:00 BST, 24 of 64 PASS at 13:4x.
+
+### 6.7 The mixer axis per family: the verifier rows at x4, x8 and x16 (build-3, 15:1x BST, `logs/bench-x{4,8,16}.log`)
+
+`igneum-pow bench --seed igneum-genesis --day 2026-10-03 --class v2m<m>g+sh256x27 --warps 50` (the fg7 CLI; `v2m8g+sh256x27` is the class v4 shape and prints as `mx8+sh256x27`), one leased core of build-3 (AMD Ryzen 9 7950X3D, the box at load from the point-B census and the strata; the cold warp is the ladder's method, one warp per vector base, then the average of 50):
+
+| Mixer `m` | Mixers per item | Cold warp, ms (base 0 / 4,096 / 1,000,000) | Average of 50, ms | Against x8 | Scaled to the 2019-class core (the O-1.14 i7-9700K read x8 at 6.33 cold) | Scaled to the half-core proxy (x8 at 8.23) | The 10 ms gate |
+|---|---|---|---|---|---|---|---|
+| 4 | 36 | 4.07 / 3.65 / 3.66 | 3.73 | 0.91x | about 5.7 | about 7.5 | passes with margin |
+| 8 (the shipped class) | 72 | 4.37 / 4.16 / 4.11 | 4.12 | 1.00x | 6.33 (measured) | 8.23 (measured) | passes |
+| 16 | 144 | 7.43 / 7.23 / 7.15 | 7.13 | 1.73x | about 11.0 | about 14.2 | FAILS on both proxies (approximate: a ratio scaling; the x16 row on the proxies themselves is the hash lane's to take) |
+
+Reading. The mixer's cost sits in the item derivation, which the verifier pays in full (4,096 items per warp) and the card hides behind the dataset build: x16 adds 3.0 ms per warp on this core, 73 percent over x8, so on the 2019-class core and the half-core proxy it lands over the 10 ms gate by one to four milliseconds. The band for `m` is therefore {4, 8} on the verifier's measured rows, with 16 outside it until a verifier change (or the ladder's rung 0 budget re-cut) brings it under 10 ms; x4 keeps 2 of 4 applications of measured margin (adv-mixer-3's rows clean from k = 2) against x8's 6 of 8, so the family's safe band on this axis is x8 with x4 as the one corner, and the chip row of chip-model-v3 (the f = 0 recompute chip's rate halving per doubling, the f = 1 chip unmoved) is unchanged by the draw. Per tier: nothing moves for a card (the item derivation is the daily build, 42 ms at x8 on the 5090); a node verifying x4 saves 10 percent of a warp, x16 would cost it 73 percent more and the 2019-class tier its margin.
+
+### 6.8 The two next-class tests on one scale: the known-failed cases through the sigma form (build-2, 14:2x BST, `logs/f8space-*`)
+
+The F8 label space's p2 to p65 and p212 to p225 drawn at the shipped parameters through the family harness (shape 256 x 27, m = 8, width 1, the table weights, B = 0; every attempt equal to the chain's class v5 draw, p2 at 5, p3 at 1, p4 at 0, p8 at 2, p10 at 0, p212 at 4, p225 at 0), read at the rule's own 2^20 sample: the minimum (c'') ratio, the largest 4,096-word bucket in sigma of its window expectation, the largest index-bit one-count excess in sigma.
+
+| Seed | What it is | min ratio, site | Largest bucket: ratio, sigma, site (window) | Bit bias: sigma, site, bit | R | Bucket band at +8 | Bit band at 6 sigma | Bit band at 300 sigma |
+|---|---|---|---|---|---|---|---|---|
+| p4 | the morning's tail (1.217x live) | 0.9961 at site 1 | 7.17x, +49.4, site 1 (quarter) | -511, site 1, bit 13 | 13 | caught | caught | caught |
+| p8 | the tail (1.378x) | 0.9963 at site 14 | 3.72x, +21.8, site 14 (quarter) | -513, site 6, bit 22 | 22 | caught | caught | caught |
+| p10 | the tail (1.504x) | 0.9963 at site 8 | 6.27x, +42.1, site 8 (quarter) | -513, site 8, bit 17 | 17 | caught | caught | caught |
+| p34 | the tail on sub-version 3 (1.250x; not on class v5's list) | 0.9999 | 2.38x, +5.5 (clean) | +10.8, site 4, bit 7 | 7 | missed | caught | missed |
+| p212 | the hash lane's F8-256 attribution: site 9's bucket at 1.75x its window expectation, 0.019 bits short | 0.9958 at site 9 | 1.72x, +5.75, site 9 (quarter): the clean spread's own maximum | -512, site 9, bit 8 | 8 | missed | caught | caught |
+| p225 | the hash lane's value-level concentration at a mad-written site, the bucket at expectation | 0.9974 at site 4 | 2.38x, +5.5 (clean) | -567, site 4, bit 2 | 2 | missed | caught | caught |
+| p23, p19, p18, p56 | the strong seeds of sub-versions 1 and 2, closed by (a') and (c'') | 0.9999 to 1.0000 | +5.1 to +5.5 (clean) | under 4.4 | | clean | clean | clean |
+| p15 | the same | 0.99997 | +6.0 | -58, site 0, bit 0 | 26 | clean | caught | missed |
+
+Reading. (1) Every named case is the SAME mechanism at the bit level: a product's bit 0 (P = 1/4, z = -512 at 2^20 exactly: (0.25 - 0.5) x 2^20 / 512) or bit 1 at address bit R of one site; the bucket statistic sees it only when R lands inside the 4,096-word bucket's bits at a narrow window (p4, p8, p10 at bits 13, 22, 17 on quarter windows) and not when it lands below bit 12 (p212 at bit 8, p225 at bit 2). So the bucket bound at any band catches a subset of what the bit read catches, and p212 and p225 are exactly the cases the research lane asked about: a bucket bound at about 2x (+8 sigma at 2^20) refuses p4, p8 and p10 and neither of the two; a bucket bound at 1.5x would take p212 (1.72x) at a cost the random stratum prices at 7.3 percent of eras over +15 sigma and 14.5 percent over +8 (the biased population, section 6.2), not the 3 to 6 percent of a clean spread. (2) The bit read at 6 sigma catches all seven named cases (p34 and p15 too); its cost is the family's standing fact, 48 to 52 percent of eras over 6 sigma, 34 percent over 100, 16 percent over 300, 11 percent over 500 (random stratum, 10,000 eras; the band 47.7, 31.1, 15.7, 11.1; width 4 under the band 44.9, 26.5, 9.4, 5.6). A band at 300 sigma catches p4, p8, p10, p212, p225 (all at or above 511) and misses p34 (+10.8) and p15 (-58), at a redraw cost of 16 percent of epochs; the clean-era maximum over 26 bits and 16 sites is 5.5 sigma at p99, so nothing clean is refused at any band from 6 up. (3) The bucket excess beyond what the bit read already names is 0.07 to 0.23 percent of eras (bucket over +8 with every bit under 6 sigma: 15 of 10,000 random, 7 of 3,000 band, 2 of 3,000 at width 4), which is the bucket bound's whole independent catch; a chip reads nothing in it that the bit read did not price. The decision this lane ships: ONE value-level test, the per-site index-bit read, as a per-era RECORD at the acceptance's 2^20 sample (its z at bit R names the product writer for the record) and as the structural lever's known-failed set (the seven seeds above), with the bucket bound retired into it; the refusal band, if the research lane wants one before `load_index`'s fold lands, is 300 sigma (five of seven named cases, 16 percent of epochs redrawn) and never 6 (half the epochs).
+
+## 7. The 16:30 BST report (the founder's clock, pulled from 09:00; a partial carries its count)
+
+1. DONE 13:2x BST: the lossy-share curve per shape (section 6.4) and the last-resort scan's verdicts on every exhausted era.
+2. DONE 14:4x BST: the pre-floor spread at width 4 against width 1 under the band and the floor decision (section 6.5: 0.995 kept at width 4 at its measured cost; the 0.990 to 0.995 band at width 4 read live before any floor moves).
+3. By 16:00: the two ring-C live-dataset rows (section 6.6; point A on build-1, point B on build-3).
+4. DONE 15:1x BST: the x4, x8 and x16 verifier rows (section 6.7, build-3 one core: 3.73 / 4.12 / 7.13 ms per warp; x16 outside the band on the 10 ms gate by scaling to the proxies).
+5. DONE 14:4x BST: the width-4 plus shape-64 crossing (build-4, 3,000 eras: r = 0.654, 0 exhausted, (c''') 5.3 percent of candidates, the width's floor cost again) and section 6.8, the F8 label space's p2 to p65 and p212 to p225 through the sigma form at the shipped parameters (build-2, queued 14:0x BST), so the two F8-256 attributions the hash lane added to layer 4's bucket-bound row (p212, site 9 at 1.75x its window expectation; p225, a value-level concentration at a mad-written site with the bucket at expectation) and the morning's four tail seeds are read as known-failed cases of the bucket-sigma and bit-bias tests on one scale; and the bucket bound in sigma (from the bit-clean spread: p99 +6.8, p99.9 +16.5 to +20.1, max +35 over 9,409 eras; a band at +8 refuses 0.3 to 0.4 percent of bit-clean programs and 11 to 18 percent of all, which is the biased population, so the bucket bound and the bit read are one rule).
+6. The gate-record form: `logs/gate-records-lossy-base.jsonl` (3,000 records of the band stratum in the section 4.3 shape, one per line, 3.6 MB) and `logs/fg-records.py`, which derives every other stratum's records from its TSV rows in this directory.
+7. Not in this report: the mixer ladder re-run at m = 4 and 16 (adv-mixer-3's harness rows; the per-family rows of section 4), the F8 census at D = 29 (the stand-in gap row), the 10^5-program verifier census at the top corner (F6 per family): each a named per-family row with its hours in section 4.1.
 
 ## 8. Sources
 

docs/analysis/class-v6/logs/bench-x16.log

@@ -0,0 +1,11 @@
+lease: holding 1 pool cores (10, waited 0 s, class measure): family gate: verifier row, bench v2m16g+sh256x27 50 warps one core (build-3)
+igneum-pow bench: seed "igneum-genesis", day "2026-10-03", dataset 2^28 words (memory-hard)
+cache: fill 323.7 ms on one core (2^26 words, 256 MiB, 65536 chains of 64 ChaCha12 blocks), FNV-1a 64 48c4f5bf24166b2e
+program: class v2m16g+sh256x27, 128 loads/hash, 512 bytes/hash, widths (1,4,16 words) [16, 0, 0], 4096 items/warp, mixer x16 (144 mixers/item), cache 2^26 words, op mix load=16 add=8 shfl=8 xor=6 mad=5 mul=5 mulhi=5 sub=4 rotl=3 rotr=3 or=1; epoch built in 309.8 ms
+shadow: 256 instructions x 27 passes per iteration, 55296 shadow instructions per hash, op mix add=47 rotl=30 xor=30 shfl=29 mad=27 mul=22 sub=21 rotr=20 mulhi=18 or=12
+warp base 0: single cold run 7.429 ms, 4096 items derived, lane0 11a6ebee3f284225 lane31 d69b6f764b683e9c
+warp base 4096: single cold run 7.226 ms, 4096 items derived, lane0 f13cf250640b741e lane31 b77d190ae6454aaf
+warp base 1000000: single cold run 7.149 ms, 4096 items derived, lane0 1d0f32760ac86802 lane31 952916a3b036252d
+CPU verify: 7.128 ms per 32-lane warp, avg of 50 (checksum 9a88da132b815f31)
+Terminated
+lease: released 1 pool cores after 1 s, exit 0

docs/analysis/class-v6/logs/bench-x4.log

@@ -0,0 +1,11 @@
+lease: holding 1 pool cores (9, waited 0 s, class measure): family gate: verifier row, bench v2m4g+sh256x27 50 warps one core (build-3)
+igneum-pow bench: seed "igneum-genesis", day "2026-10-03", dataset 2^28 words (memory-hard)
+cache: fill 374.9 ms on one core (2^26 words, 256 MiB, 65536 chains of 64 ChaCha12 blocks), FNV-1a 64 48c4f5bf24166b2e
+program: class mx4+sh256x27, 128 loads/hash, 512 bytes/hash, widths (1,4,16 words) [16, 0, 0], 4096 items/warp, mixer x4 (36 mixers/item), cache 2^26 words, op mix load=16 add=8 shfl=8 xor=6 mad=5 mul=5 mulhi=5 sub=4 rotl=3 rotr=3 or=1; epoch built in 397.6 ms
+shadow: 256 instructions x 27 passes per iteration, 55296 shadow instructions per hash, op mix add=47 rotl=30 xor=30 shfl=29 mad=27 mul=22 sub=21 rotr=20 mulhi=18 or=12
+warp base 0: single cold run 4.074 ms, 4096 items derived, lane0 7d5c0b3cbd0fcade lane31 2a25e4b6bf48485d
+warp base 4096: single cold run 3.650 ms, 4096 items derived, lane0 73a7e6af84f4d3ca lane31 f26081f8ff529a1a
+warp base 1000000: single cold run 3.661 ms, 4096 items derived, lane0 2e20e9a0b5cb35da lane31 a8a0e975fa5ae184
+CPU verify: 3.725 ms per 32-lane warp, avg of 50 (checksum 164082cfe849fab4)
+Terminated
+lease: released 1 pool cores after 1 s, exit 0

docs/analysis/class-v6/logs/bench-x8.log

@@ -0,0 +1,10 @@
+lease: holding 1 pool cores (8, waited 0 s, class measure): family gate: verifier row, bench v2m8g+sh256x27 50 warps one core (build-3)
+igneum-pow bench: seed "igneum-genesis", day "2026-10-03", dataset 2^28 words (memory-hard)
+cache: fill 448.8 ms on one core (2^26 words, 256 MiB, 65536 chains of 64 ChaCha12 blocks), FNV-1a 64 48c4f5bf24166b2e
+program: class mx8+sh256x27, 128 loads/hash, 512 bytes/hash, widths (1,4,16 words) [16, 0, 0], 4096 items/warp, mixer x8 (72 mixers/item), cache 2^26 words, op mix load=16 add=8 shfl=8 xor=6 mad=5 mul=5 mulhi=5 sub=4 rotl=3 rotr=3 or=1; epoch built in 2033.0 ms
+shadow: 256 instructions x 27 passes per iteration, 55296 shadow instructions per hash, op mix add=47 rotl=30 xor=30 shfl=29 mad=27 mul=22 sub=21 rotr=20 mulhi=18 or=12
+warp base 0: single cold run 4.368 ms, 4096 items derived, lane0 dfbc8db1c06dacd8 lane31 b92715f32b721587
+warp base 4096: single cold run 4.161 ms, 4096 items derived, lane0 433afad2164a8dda lane31 207c50f288cbb677
+warp base 1000000: single cold run 4.113 ms, 4096 items derived, lane0 4ffd090a02db43b4 lane31 60308a0e058e1e29
+CPU verify: 4.116 ms per 32-lane warp, avg of 50 (checksum 900cd36a9ead25cf)
+lease: released 1 pool cores after 3 s, exit 0

docs/analysis/class-v6/logs/f8space-p2-65-family_gate_era_census-2-64.log

@@ -0,0 +1,12 @@
+binary bc21542a5b421170 commit fg7 test family_gate_era_census from 2 seeds 64 threads 8 start 2026-10-08T13:05:16Z
+lease: 0 of 8 pool cores free (88 leased); waiting
+lease: holding 8 pool cores (14,15,16,17,24,25,26,27, waited 1217 s, class measure): family gate: family_gate_era_census f8space-p2-65 seeds 2+64
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 64 eras 2..66 in 49 s on 8 threads -> /srv/builds/_adv-family-gate/logs/f8space-p2-65-family_gate_era_census-2-64.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 49.18s
+
+lease: released 8 pool cores after 1267 s, exit 0
+end 2026-10-08T13:26:23Z rc 0

docs/analysis/class-v6/logs/f8space-p2-65-family_gate_era_census-2-64.tsv

@@ -0,0 +1,65 @@
+k	shape	reps	mixer	width	R	M	pos	w_add	w_xor	w_mul	w_mad	w_shfl	w_rotl	w_sub	w_mulhi	w_rotr	w_or	attempt	candidates	a_prime	a	b	c_const	c_lane	c_sat	c_bias	c_distinct	c1_sat_source	c2_low_entropy	c3_hot_item	min_ratio	min_site	top_count_max	bucket_ratio_max	bucket_z_max	bucket_site	bucket_win	bit_z_max	bit_site	bit_win	bit	last_resort_k	c3_ratios	secs
+8	256	27	8	1	22	573e71b3	[1, 4, 7, 12]	12	10	8	8	8	7	6	6	6	4	2	3	1	1	0	0	0	0	0	0	0	0	0	0.99628	14	4	3.719	21.75	14	2	-512.76	6	1	22			4.6
+2	256	27	8	1	4	bf0fab63	[2, 3, 7, 15]	12	10	8	8	8	7	6	6	6	4	5	6	4	0	1	0	0	0	0	0	0	0	0	0.99994	10	3	1.688	5.50	6	2	-2.83	6	2	16			4.7
+4	256	27	8	1	13	2477779f	[5, 10, 11, 15]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99613	1	4	7.172	49.38	1	2	-510.71	1	2	13			4.7
+6	256	27	8	1	21	05ba4603	[0, 2, 7, 11]	12	10	8	8	8	7	6	6	6	4	6	7	5	0	1	0	0	0	0	0	0	0	0	0.99987	14	4	2.062	6.01	1	1	-3.95	10	2	24			4.8
+3	256	27	8	1	1	a80102a7	[0, 5, 8, 15]	12	10	8	8	8	7	6	6	6	4	1	2	0	1	0	0	0	0	0	0	0	0	0	0.99968	8	4	2.375	5.50	0	0	257.84	8	1	1			5.0
+5	256	27	8	1	4	435fd6c5	[0, 3, 11, 12]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99990	1	3	1.938	5.30	4	1	-2.60	2	0	17			5.0
+7	256	27	8	1	25	ff347267	[2, 6, 7, 14]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99946	13	4	2.938	7.75	13	0	-511.12	13	0	25			5.1
+9	256	27	8	1	18	07114293	[0, 3, 11, 14]	12	10	8	8	8	7	6	6	6	4	8	9	7	1	0	0	0	0	0	0	0	0	0	0.99990	6	3	2.438	5.75	10	0	168.61	12	0	18			8.6
+12	256	27	8	1	16	2fda8d19	[0, 4, 6, 10]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99999	14	4	2.375	5.50	0	0	35.93	0	0	17			4.6
+11	256	27	8	1	29	b889f389	[3, 5, 9, 12]	12	10	8	8	8	7	6	6	6	4	3	4	2	1	0	0	0	0	0	0	0	0	0	0.99993	12	4	2.312	5.25	3	0	-5.30	1	1	0			4.7
+13	256	27	8	1	26	3340424b	[1, 7, 14, 15]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99994	3	4	2.562	6.25	8	0	144.94	8	0	26			4.7
+10	256	27	8	1	17	5732bb5d	[3, 5, 8, 13]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99626	8	3	6.266	42.12	8	2	-512.88	8	2	17			4.9
+16	256	27	8	1	1	a6ea3eb1	[1, 4, 7, 10]	12	10	8	8	8	7	6	6	6	4	3	4	3	0	0	0	0	0	0	0	0	0	0	0.99991	8	4	2.375	5.50	9	0	-128.13	4	0	1			4.7
+14	256	27	8	1	23	55fb3ce5	[2, 5, 8, 11]	12	10	8	8	8	7	6	6	6	4	4	5	4	0	0	0	0	0	0	0	0	0	0	0.99993	1	4	2.375	5.50	9	0	3.25	8	1	0			4.8
+15	256	27	8	1	26	26f35c81	[0, 2, 14, 15]	12	10	8	8	8	7	6	6	6	4	4	5	2	0	1	0	0	0	0	0	0	1	0	0.99997	10	3	2.500	6.00	1	0	-58.21	0	1	0		c2:943	7.3
+21	256	27	8	1	22	48825377	[3, 5, 6, 8]	12	10	8	8	8	7	6	6	6	4	3	4	3	0	0	0	0	0	0	0	0	0	0	0.99995	11	4	2.438	5.75	0	0	2.63	0	0	4			4.4
+20	256	27	8	1	28	aa378839	[0, 3, 7, 10]	12	10	8	8	8	7	6	6	6	4	2	3	2	0	0	0	0	0	0	0	0	0	0	0.99988	14	4	1.906	5.13	7	1	3.34	12	2	24			4.5
+22	256	27	8	1	25	1d32e3af	[7, 9, 11, 15]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99991	7	3	1.938	5.30	1	1	-129.05	6	2	25			4.9
+23	256	27	8	1	26	d88595c9	[0, 1, 4, 14]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99996	5	4	2.312	5.25	0	0	2.97	5	0	13			5.1
+19	256	27	8	1	9	69944467	[6, 9, 11, 12]	12	10	8	8	8	7	6	6	6	4	2	3	0	1	0	0	0	0	0	0	0	1	0	0.99995	13	3	1.906	5.13	6	1	4.40	2	0	25		c2:933	6.9
+18	256	27	8	1	22	83e41bc3	[1, 8, 10, 13]	12	10	8	8	8	7	6	6	6	4	9	10	7	1	0	0	0	0	0	0	0	1	0	0.99992	10	4	2.375	5.50	1	0	3.38	6	1	23		c2:927	7.0
+17	256	27	8	1	7	fb71a153	[2, 5, 8, 10]	12	10	8	8	8	7	6	6	6	4	3	4	3	0	0	0	0	0	0	0	0	0	0	0.99992	15	4	2.375	5.50	11	0	2.93	14	0	4			8.4
+24	256	27	8	1	3	44fd49d7	[0, 4, 7, 9]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99985	14	4	1.906	5.13	3	1	-9.35	15	0	6			4.9
+26	256	27	8	1	19	9f5f0bdb	[2, 7, 11, 12]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99910	2	3	6.938	23.75	2	0	-511.00	2	0	19			4.7
+25	256	27	8	1	14	c619dd6b	[2, 6, 8, 15]	12	10	8	8	8	7	6	6	6	4	2	3	2	0	0	0	0	0	0	0	0	0	0	0.99984	2	3	2.250	5.00	11	0	-2.72	13	0	3			4.8
+27	256	27	8	1	15	ab4deb19	[4, 5, 11, 13]	12	10	8	8	8	7	6	6	6	4	6	7	6	0	0	0	0	0	0	0	0	0	0	0.99899	1	3	8.562	30.25	1	0	-512.72	1	0	15			4.5
+28	256	27	8	1	26	347b218b	[0, 6, 12, 13]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99995	0	3	2.312	5.25	3	0	3.66	1	0	10			4.7
+29	256	27	8	1	25	e597ba17	[4, 5, 9, 14]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99991	2	4	2.500	6.00	10	0	84.65	12	1	25			4.6
+30	256	27	8	1	18	ac8f1c1b	[3, 4, 11, 12]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99990	4	3	2.375	5.50	4	0	-2.95	4	0	13			4.6
+32	256	27	8	1	17	dc6791fd	[6, 11, 14, 15]	12	10	8	8	8	7	6	6	6	4	6	7	3	1	0	0	0	2	0	0	0	0	0	0.99878	12	4	8.188	28.75	12	0	-513.54	12	0	17			4.7
+33	256	27	8	1	6	3e376d09	[0, 3, 4, 9]	12	10	8	8	8	7	6	6	6	4	4	5	4	0	0	0	0	0	0	0	0	0	0	0.99711	0	4	2.312	5.25	10	0	-512.62	7	1	6			4.6
+35	256	27	8	1	24	96fc94e3	[0, 7, 12, 15]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99986	11	4	1.656	5.25	14	2	-4.81	1	0	25			4.8
+36	256	27	8	1	27	4c7a6647	[9, 10, 12, 15]	12	10	8	8	8	7	6	6	6	4	12	13	11	1	0	0	0	0	0	0	0	0	0	0.99993	11	4	2.375	5.50	14	0	2.74	6	0	3			4.9
+34	256	27	8	1	7	2e362f07	[3, 6, 11, 14]	12	10	8	8	8	7	6	6	6	4	1	2	0	0	0	0	0	0	0	0	0	0	1	0.99992	3	4	2.375	5.50	14	0	10.78	4	1	7		c3:992	6.7
+38	256	27	8	1	19	06d774fb	[3, 4, 12, 15]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99993	6	4	2.312	5.25	1	0	-63.05	5	2	19			4.5
+31	256	27	8	1	14	d4aea2c7	[2, 5, 12, 14]	12	10	8	8	8	7	6	6	6	4	4	5	4	0	0	0	0	0	0	0	0	0	0	0.99994	15	3	2.375	5.50	1	0	7.27	12	0	15			8.4
+37	256	27	8	1	18	d1f25c1f	[1, 7, 8, 13]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99994	13	4	1.938	5.30	8	1	3.82	3	0	18			4.7
+39	256	27	8	1	3	a51e7c6d	[3, 6, 12, 13]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99905	1	4	2.031	5.83	7	1	-513.32	1	0	3			4.9
+40	256	27	8	1	16	4bc62819	[1, 9, 10, 13]	12	10	8	8	8	7	6	6	6	4	3	4	2	0	1	0	0	0	0	0	0	0	0	0.99995	4	4	2.375	5.50	3	0	-3.37	15	0	2			4.6
+41	256	27	8	1	16	7a5295a9	[0, 3, 7, 11]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99982	11	4	2.000	5.66	0	1	-3.16	14	1	4			4.8
+42	256	27	8	1	16	e3baafe9	[2, 7, 10, 14]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99989	9	4	2.375	5.50	13	0	3.63	9	2	0			4.9
+44	256	27	8	1	14	2d1bd537	[2, 7, 13, 15]	12	10	8	8	8	7	6	6	6	4	1	2	0	1	0	0	0	0	0	0	0	0	0	0.99616	8	4	7.281	50.25	8	2	-512.09	8	2	14			4.7
+43	256	27	8	1	19	cd014ac5	[11, 12, 13, 14]	12	10	8	8	8	7	6	6	6	4	5	6	4	1	0	0	0	0	0	0	0	0	0	0.99919	7	4	2.594	9.02	7	1	-127.25	7	1	20			5.1
+46	256	27	8	1	24	6472aa9d	[3, 9, 10, 13]	12	10	8	8	8	7	6	6	6	4	5	6	4	1	0	0	0	0	0	0	0	0	0	0.99722	4	4	2.703	13.62	4	2	-512.03	4	2	24			5.0
+47	256	27	8	1	25	00a713dd	[3, 7, 8, 9]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99989	11	4	2.375	5.50	6	0	3.36	7	2	18			5.1
+49	256	27	8	1	30	b1cbeac1	[6, 8, 9, 14]	12	10	8	8	8	7	6	6	6	4	6	7	5	1	0	0	0	0	0	0	0	0	0	0.99991	0	4	2.438	5.75	5	0	-2.86	5	0	21			5.1
+45	256	27	8	1	18	d9c46e27	[3, 4, 6, 13]	12	10	8	8	8	7	6	6	6	4	3	4	2	1	0	0	0	0	0	0	0	0	0	0.99963	13	3	2.812	10.25	13	1	255.79	13	1	18			8.7
+50	256	27	8	1	10	2b3dbe09	[7, 13, 14, 15]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99807	5	4	3.500	14.14	5	1	-512.40	5	1	10			6.1
+48	256	27	8	1	13	5e051e51	[9, 10, 12, 13]	12	10	8	8	8	7	6	6	6	4	6	7	5	1	0	0	0	0	0	0	0	0	0	0.99986	14	3	2.500	6.00	11	0	-2.94	5	0	21			7.6
+54	256	27	8	1	31	24b4a1e5	[5, 10, 13, 15]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99994	12	4	2.375	5.50	10	0	2.77	11	2	14			5.3
+51	256	27	8	1	3	d94b2ac7	[0, 2, 4, 9]	12	10	8	8	8	7	6	6	6	4	4	5	4	0	0	0	0	0	0	0	0	0	0	0.99993	12	4	2.312	5.25	3	0	-3.88	6	1	15			8.8
+55	256	27	8	1	7	0e0c23b9	[1, 5, 13, 14]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99625	7	4	2.312	5.25	8	0	-511.26	7	2	7			5.3
+53	256	27	8	1	9	7ec0e095	[1, 8, 9, 14]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99949	15	3	2.500	6.00	13	0	159.62	15	2	9			8.8
+52	256	27	8	1	12	dd60e49d	[2, 4, 12, 14]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99618	14	4	8.078	56.62	14	2	-512.06	14	2	12			9.4
+57	256	27	8	1	25	2bfb510f	[0, 3, 5, 9]	12	10	8	8	8	7	6	6	6	4	4	5	4	0	0	0	0	0	0	0	0	0	0	0.99997	11	4	1.938	5.30	10	1	-3.40	2	0	19			6.1
+59	256	27	8	1	7	ab70a8c7	[2, 3, 8, 10]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99993	14	3	2.094	6.19	15	1	-11.71	13	0	7			5.1
+61	256	27	8	1	4	99ec1a13	[3, 8, 13, 14]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99909	1	4	2.375	5.50	9	0	-512.35	1	0	4			5.0
+58	256	27	8	1	30	2faae5d7	[1, 5, 14, 15]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99987	7	3	1.969	5.48	14	1	3.57	8	1	3			9.0
+56	256	27	8	1	24	dc5e362b	[6, 7, 9, 10]	12	10	8	8	8	7	6	6	6	4	4	5	3	0	0	0	0	0	0	0	0	1	0	0.99995	2	4	2.312	5.25	3	0	3.51	8	1	6		c2:965	11.7
+65	256	27	8	1	23	8d7fa8fb	[1, 10, 11, 13]	12	10	8	8	8	7	6	6	6	4	3	4	2	0	1	0	0	0	0	0	0	0	0	0.99811	4	4	2.562	12.50	4	2	-449.07	4	2	23			5.0
+64	256	27	8	1	8	f7401b9f	[4, 7, 10, 15]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99990	1	4	2.562	6.25	0	0	-68.29	0	0	8			6.1
+60	256	27	8	1	4	cf5e9ceb	[0, 2, 5, 8]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99989	8	3	2.250	5.00	15	0	-108.06	2	2	5			9.0
+62	256	27	8	1	11	ae89e4ab	[3, 5, 8, 13]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99815	9	3	2.203	9.62	9	2	-343.68	9	2	11			8.9
+63	256	27	8	1	20	eda85b67	[0, 3, 9, 13]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99988	2	4	2.375	5.50	4	0	-64.53	14	0	21			9.3

docs/analysis/class-v6/logs/f8space-p212-225-family_gate_era_census-212-14.log

@@ -0,0 +1,13 @@
+binary bc21542a5b421170 commit fg7 test family_gate_era_census from 212 seeds 14 threads 4 start 2026-10-08T13:05:16Z
+lease: 0 of 4 pool cores free (88 leased); waiting
+lease: holding 4 pool cores (28,29,30,31, waited 1217 s, class measure): family gate: family_gate_era_census f8space-p212-225 seeds 212+14
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 14 eras 212..226 in 35 s on 4 threads -> /srv/builds/_adv-family-gate/logs/f8space-p212-225-family_gate_era_census-212-14.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 34.69s
+
+Terminated
+lease: released 4 pool cores after 1252 s, exit 0
+end 2026-10-08T13:26:08Z rc 0

docs/analysis/class-v6/logs/f8space-p212-225-family_gate_era_census-212-14.tsv

@@ -0,0 +1,15 @@
+k	shape	reps	mixer	width	R	M	pos	w_add	w_xor	w_mul	w_mad	w_shfl	w_rotl	w_sub	w_mulhi	w_rotr	w_or	attempt	candidates	a_prime	a	b	c_const	c_lane	c_sat	c_bias	c_distinct	c1_sat_source	c2_low_entropy	c3_hot_item	min_ratio	min_site	top_count_max	bucket_ratio_max	bucket_z_max	bucket_site	bucket_win	bit_z_max	bit_site	bit_win	bit	last_resort_k	c3_ratios	secs
+215	256	27	8	1	7	9f95aef9	[3, 4, 7, 14]	12	10	8	8	8	7	6	6	6	4	5	6	4	1	0	0	0	0	0	0	0	0	0	0.99969	6	3	2.000	5.66	10	1	-256.22	6	1	7			8.7
+213	256	27	8	1	31	04aeeee3	[0, 1, 3, 9]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99995	13	3	1.938	5.30	5	1	3.02	3	1	0			8.8
+214	256	27	8	1	1	04870f7f	[1, 2, 6, 15]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99990	13	3	2.438	5.75	14	0	-10.66	13	0	2			8.9
+212	256	27	8	1	8	d56f9b19	[5, 7, 13, 15]	12	10	8	8	8	7	6	6	6	4	4	5	3	0	0	0	0	0	0	0	0	0	1	0.99577	9	4	1.719	5.75	9	2	-512.14	9	2	8		c3:991	13.0
+216	256	27	8	1	30	ddd46ce1	[0, 1, 7, 9]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99987	8	3	2.312	5.25	3	0	-3.32	12	0	20			8.5
+217	256	27	8	1	15	2bb44f89	[10, 11, 14, 15]	12	10	8	8	8	7	6	6	6	4	1	2	0	1	0	0	0	0	0	0	0	0	0	0.99987	8	3	2.312	5.25	3	0	-64.88	0	1	15			8.7
+218	256	27	8	1	10	210c1f81	[3, 6, 12, 13]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99934	5	4	2.000	5.66	5	1	-353.27	5	1	11			9.2
+219	256	27	8	1	2	fbc1458b	[2, 8, 13, 14]	12	10	8	8	8	7	6	6	6	4	2	3	1	1	0	0	0	0	0	0	0	0	0	0.99993	12	3	2.312	5.25	3	0	2.83	9	2	4			9.1
+221	256	27	8	1	16	147a60c9	[2, 6, 9, 11]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99900	14	3	7.688	26.75	14	0	-512.53	14	0	16			8.0
+220	256	27	8	1	16	3b974b5b	[0, 1, 9, 10]	12	10	8	8	8	7	6	6	6	4	2	3	2	0	0	0	0	0	0	0	0	0	0	0.99872	1	4	6.812	32.88	1	1	-438.52	1	1	16			8.6
+222	256	27	8	1	27	38ebb1bd	[0, 5, 8, 10]	12	10	8	8	8	7	6	6	6	4	1	2	1	0	0	0	0	0	0	0	0	0	0	0.99987	0	4	2.375	5.50	0	0	-122.37	1	0	27			8.9
+223	256	27	8	1	17	195d676f	[0, 6, 8, 13]	12	10	8	8	8	7	6	6	6	4	4	5	3	1	0	0	0	0	0	0	0	0	0	0.99993	6	3	2.031	5.83	1	1	3.29	15	1	4			8.5
+224	256	27	8	1	18	8a4c7517	[1, 2, 3, 6]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99987	12	4	1.656	5.25	13	2	3.29	2	0	17			8.7
+225	256	27	8	1	2	2ef225c7	[0, 12, 14, 15]	12	10	8	8	8	7	6	6	6	4	0	1	0	0	0	0	0	0	0	0	0	0	0	0.99740	4	3	2.375	5.50	5	0	-567.19	4	1	2			8.9

docs/analysis/class-v6/logs/fg-records.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
-"""family-gate lane: one gate-record JSON per drawn era from the census TSV rows (the format of family-gate.md section 4.3).
-usage: fg-records.py <outdir> <stratum>=<tsv>[,<tsv>] ... ; the harness binary sha and box from the matching .log header."""
+"""family-gate lane: one gate-record JSON per drawn era (the format of family-gate.md section 4.3) from the census TSV rows, written as ONE JSON-lines file per stratum under <outdir> (25,000 files would not belong in a repository).
+usage: fg-records.py <outdir> <stratum>::<tsv>[,<tsv>] ... ; the harness binary sha from the matching .log header."""
 import sys, csv, json, os, pathlib, hashlib
 FAMILY = {"id": "class-v6-draft-2026-10-08", "spec": "docs/spec/01-lottery-hash.md 1.4.6, 1.13.1 (the family's bands in docs/analysis/class-v6/family-gate.md section 1)",
           "object": "igneum-pow class-v5 8f481459 plus the harness (branch family-gate-v5), the acceptance keyed on the family's shapes behind IGNEUM_FAMILY_GATE"}
@@ -13,14 +13,14 @@ def header(logpath):
 out = pathlib.Path(sys.argv[1]); out.mkdir(parents=True, exist_ok=True)
 n = 0
 for arg in sys.argv[2:]:
-    stratum, paths = arg.split('=', 1)
+    stratum, paths = arg.split('::', 1)
     for p in paths.split(','):
         if not os.path.exists(p): continue
         h = header(p[:-4] + '.log')
         with open(p) as f:
             for r in csv.DictReader(f, delimiter='\t'):
                 acc = r['attempt'] not in ('', '-1')
-                rec = {"family": FAMILY, "label": "internal research, not an independent review",
+                rec = {"family": FAMILY["id"], "label": "internal research, not an independent review",
                        "era": {"label": f"igneum-family-gate/era/{r['k']}", "stratum": stratum,
                                "draw": {"m": int(r['mixer']), "weights": [int(r[k]) for k in ['w_add','w_xor','w_mul','w_mad','w_shfl','w_rotl','w_sub','w_mulhi','w_rotr','w_or']],
                                         "width_words": int(r['width']), "shape": [int(r['shape']), int(r['reps'])], "R": int(r['R']), "M": "0x" + r['M'], "pos": json.loads(r['pos'])}},
@@ -31,8 +31,9 @@ for arg in sys.argv[2:]:
                             "exhausted": not acc, "verdict": "PASS" if acc else "EXHAUSTED"},
                            {"name": "c2-c3-ratio", "floor": [0.98, 0.995], "min_ratio": float(r['min_ratio']) if acc else None, "min_site": int(r['min_site']) if acc else None, "verdict": "PASS" if acc else "n/a"},
                            {"name": "bucket-256-items", "max_ratio": float(r['bucket_ratio_max']) if acc else None, "max_sigma": float(r.get('bucket_z_max') or 'nan') if acc else None, "site": int(r['bucket_site']) if acc else None, "window": int(r.get('bucket_win') or -1) if acc else None, "bound": "not yet set (the clean spread is this census)", "verdict": "RECORD"},
+                           {"name": "last-resort-scan", "first_passing_k": (r.get('last_resort_k') or None) if not acc else None, "verdict": "n/a" if acc else ("PASS" if (r.get('last_resort_k') or '') not in ('', 'fallback') else "FALLBACK")},
                            {"name": "index-bit-bias", "max_sigma": float(r['bit_z_max']) if acc else None, "site": int(r['bit_site']) if acc else None, "bit": int(r['bit']) if acc else None, "window": int(r.get('bit_win') or -1) if acc else None, "band_sigma": 6, "over_band": (abs(float(r['bit_z_max'])) > 6) if acc else None, "verdict": "RECORD"}]}
-                (out / stratum).mkdir(exist_ok=True)
-                (out / stratum / f"era-{r['k']}.json").write_text(json.dumps(rec, indent=1))
+                with open(out / f"{stratum}.jsonl", "a") as fh:
+                    fh.write(json.dumps(rec, separators=(',', ':')) + "\n")
                 n += 1
-print(f"{n} records under {out}")
+print(f"{n} records under {out} (the family block: {json.dumps(FAMILY)})")

docs/analysis/class-v6/logs/floor-read-A.log

@@ -0,0 +1,45 @@
+lease: 0 of 4 pool cores free (24 leased); waiting
+lease: holding 8 pool cores (8,9,10,11,12,13,14,15, waited 693 s, class measure): family gate: the (c3) floor read on point A 64 seeds (v5_hot_census under the band weights)
+
+running 1 test
+test accept::tests::v5_hot_census ... v5_hot_census: 64 seeds 2..66 in 79 s on 8 threads
+class v4 sub-version 3 accepted programs' minimum site ratio: min 0.9932 p0.1 0.9932 p1 0.9933 p5 0.9962 median 0.9999 max 1.0000
+floor 0.980: 0 of 64 class v4 accepted programs under it (0.000 percent)
+floor 0.990: 0 of 64 class v4 accepted programs under it (0.000 percent)
+floor 0.995: 3 of 64 class v4 accepted programs under it (4.688 percent)
+floor 0.998: 6 of 64 class v4 accepted programs under it (9.375 percent)
+floor 0.999: 11 of 64 class v4 accepted programs under it (17.188 percent)
+low: seed 38 v4 attempt 0 min site 0 ratio 0.9932; v5 attempt 1 min ratio 0.9999
+low: seed 11 v4 attempt 1 min site 10 ratio 0.9933; v5 attempt 2 min ratio 0.9999
+low: seed 54 v4 attempt 0 min site 12 ratio 0.9945; v5 attempt 1 min ratio 0.9999
+low: seed 55 v4 attempt 0 min site 3 ratio 0.9962; v5 attempt 0 min ratio 0.9962
+low: seed 4 v4 attempt 0 min site 1 ratio 0.9963; v5 attempt 0 min ratio 0.9963
+low: seed 24 v4 attempt 1 min site 2 ratio 0.9969; v5 attempt 1 min ratio 0.9969
+low: seed 32 v4 attempt 0 min site 8 ratio 0.9980; v5 attempt 0 min ratio 0.9980
+low: seed 48 v4 attempt 2 min site 0 ratio 0.9981; v5 attempt 2 min ratio 0.9981
+low: seed 61 v4 attempt 0 min site 5 ratio 0.9987; v5 attempt 0 min ratio 0.9987
+low: seed 45 v4 attempt 0 min site 0 ratio 0.9988; v5 attempt 0 min ratio 0.9988
+low: seed 39 v4 attempt 0 min site 2 ratio 0.9990; v5 attempt 0 min ratio 0.9990
+low: seed 3 v4 attempt 1 min site 4 ratio 0.9991; v5 attempt 1 min ratio 0.9991
+low: seed 17 v4 attempt 2 min site 1 ratio 0.9991; v5 attempt 2 min ratio 0.9991
+low: seed 14 v4 attempt 0 min site 12 ratio 0.9992; v5 attempt 0 min ratio 0.9992
+low: seed 53 v4 attempt 0 min site 4 ratio 0.9992; v5 attempt 0 min ratio 0.9992
+low: seed 42 v4 attempt 0 min site 13 ratio 0.9994; v5 attempt 0 min ratio 0.9994
+low: seed 40 v4 attempt 1 min site 1 ratio 0.9995; v5 attempt 1 min ratio 0.9995
+low: seed 15 v4 attempt 0 min site 2 ratio 0.9995; v5 attempt 0 min ratio 0.9995
+low: seed 7 v4 attempt 0 min site 0 ratio 0.9995; v5 attempt 0 min ratio 0.9995
+low: seed 60 v4 attempt 0 min site 8 ratio 0.9997; v5 attempt 0 min ratio 0.9997
+low: seed 6 v4 attempt 0 min site 15 ratio 0.9997; v5 attempt 0 min ratio 0.9997
+low: seed 25 v4 attempt 1 min site 9 ratio 0.9997; v5 attempt 1 min ratio 0.9997
+low: seed 49 v4 attempt 0 min site 15 ratio 0.9997; v5 attempt 0 min ratio 0.9997
+low: seed 57 v4 attempt 0 min site 13 ratio 0.9998; v5 attempt 0 min ratio 0.9998
+low: seed 35 v4 attempt 0 min site 8 ratio 0.9998; v5 attempt 0 min ratio 0.9998
+attempts histogram class v4 (mean 0.297): {0: 49, 1: 12, 2: 2, 3: 1}
+attempts histogram class v5 (mean 0.344): {0: 47, 1: 13, 2: 3, 3: 1}
+seeds whose class v5 draw lands on another attempt than the class v4 draw: 3 of 64 (4.688 percent)
+class v5 accepted programs under the class v5 floor 0.995: 0 (must be 0)
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 78.69s
+
+lease: released 8 pool cores after 771 s, exit 0

docs/analysis/class-v6/logs/floor-read-B.log

@@ -0,0 +1,44 @@
+lease: holding 8 pool cores (8,9,10,11,12,13,14,15, waited 0 s, class measure): family gate: the (c3) floor read on point B 64 seeds (shape 64, band era 5 weights)
+
+running 1 test
+test accept::tests::v5_hot_census ... v5_hot_census: 64 seeds 2..66 in 77 s on 8 threads
+class v4 sub-version 3 accepted programs' minimum site ratio: min 0.9934 p0.1 0.9934 p1 0.9957 p5 0.9962 median 0.9999 max 1.0000
+floor 0.980: 0 of 64 class v4 accepted programs under it (0.000 percent)
+floor 0.990: 0 of 64 class v4 accepted programs under it (0.000 percent)
+floor 0.995: 1 of 64 class v4 accepted programs under it (1.562 percent)
+floor 0.998: 12 of 64 class v4 accepted programs under it (18.750 percent)
+floor 0.999: 14 of 64 class v4 accepted programs under it (21.875 percent)
+low: seed 15 v4 attempt 1 min site 12 ratio 0.9934; v5 attempt 4 min ratio 0.9993
+low: seed 64 v4 attempt 1 min site 4 ratio 0.9957; v5 attempt 1 min ratio 0.9957
+low: seed 54 v4 attempt 1 min site 14 ratio 0.9958; v5 attempt 1 min ratio 0.9958
+low: seed 39 v4 attempt 0 min site 4 ratio 0.9962; v5 attempt 0 min ratio 0.9962
+low: seed 42 v4 attempt 1 min site 4 ratio 0.9962; v5 attempt 1 min ratio 0.9962
+low: seed 52 v4 attempt 1 min site 14 ratio 0.9962; v5 attempt 1 min ratio 0.9962
+low: seed 55 v4 attempt 0 min site 3 ratio 0.9963; v5 attempt 0 min ratio 0.9963
+low: seed 10 v4 attempt 0 min site 8 ratio 0.9963; v5 attempt 0 min ratio 0.9963
+low: seed 18 v4 attempt 0 min site 14 ratio 0.9969; v5 attempt 0 min ratio 0.9969
+low: seed 24 v4 attempt 1 min site 2 ratio 0.9971; v5 attempt 1 min ratio 0.9971
+low: seed 41 v4 attempt 0 min site 14 ratio 0.9979; v5 attempt 0 min ratio 0.9979
+low: seed 47 v4 attempt 4 min site 11 ratio 0.9980; v5 attempt 4 min ratio 0.9980
+low: seed 32 v4 attempt 0 min site 9 ratio 0.9980; v5 attempt 0 min ratio 0.9980
+low: seed 17 v4 attempt 0 min site 3 ratio 0.9982; v5 attempt 0 min ratio 0.9982
+low: seed 34 v4 attempt 3 min site 4 ratio 0.9990; v5 attempt 3 min ratio 0.9990
+low: seed 53 v4 attempt 1 min site 15 ratio 0.9991; v5 attempt 1 min ratio 0.9991
+low: seed 5 v4 attempt 0 min site 11 ratio 0.9991; v5 attempt 0 min ratio 0.9991
+low: seed 3 v4 attempt 1 min site 10 ratio 0.9994; v5 attempt 1 min ratio 0.9994
+low: seed 7 v4 attempt 1 min site 13 ratio 0.9996; v5 attempt 1 min ratio 0.9996
+low: seed 49 v4 attempt 0 min site 4 ratio 0.9996; v5 attempt 0 min ratio 0.9996
+low: seed 40 v4 attempt 0 min site 0 ratio 0.9996; v5 attempt 0 min ratio 0.9996
+low: seed 37 v4 attempt 2 min site 14 ratio 0.9997; v5 attempt 2 min ratio 0.9997
+low: seed 43 v4 attempt 1 min site 3 ratio 0.9997; v5 attempt 1 min ratio 0.9997
+low: seed 61 v4 attempt 1 min site 13 ratio 0.9998; v5 attempt 1 min ratio 0.9998
+low: seed 20 v4 attempt 0 min site 14 ratio 0.9998; v5 attempt 0 min ratio 0.9998
+attempts histogram class v4 (mean 1.453): {0: 22, 1: 21, 2: 8, 3: 3, 4: 4, 5: 5, 6: 1}
+attempts histogram class v5 (mean 1.500): {0: 22, 1: 20, 2: 8, 3: 3, 4: 5, 5: 5, 6: 1}
+seeds whose class v5 draw lands on another attempt than the class v4 draw: 1 of 64 (1.562 percent)
+class v5 accepted programs under the class v5 floor 0.995: 0 (must be 0)
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 77.21s
+
+lease: released 8 pool cores after 77 s, exit 0

docs/analysis/class-v6/logs/lossyplus1-family_gate_era_census-0-3000.log

@@ -0,0 +1,42 @@
+binary 42f2c77ff7e5216e commit fg6 test family_gate_era_census from 0 seeds 3000 threads 16 start 2026-10-08T11:26:56Z
+lease: holding 16 pool cores (21,22,23,24,25,27,28,29,30,31,32,33,52,53,54,55, waited 0 s, class measure): family gate: family_gate_era_census lossyplus1 seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 46 s
+family_gate_era_census: 200 eras in 89 s
+family_gate_era_census: 300 eras in 133 s
+family_gate_era_census: 400 eras in 178 s
+family_gate_era_census: 500 eras in 241 s
+family_gate_era_census: 600 eras in 305 s
+family_gate_era_census: 700 eras in 372 s
+family_gate_era_census: 800 eras in 436 s
+family_gate_era_census: 900 eras in 510 s
+family_gate_era_census: 1000 eras in 579 s
+family_gate_era_census: 1100 eras in 655 s
+family_gate_era_census: 1200 eras in 724 s
+family_gate_era_census: 1300 eras in 792 s
+family_gate_era_census: 1400 eras in 872 s
+family_gate_era_census: 1500 eras in 943 s
+family_gate_era_census: 1600 eras in 1009 s
+family_gate_era_census: 1700 eras in 1069 s
+family_gate_era_census: 1800 eras in 1133 s
+family_gate_era_census: 1900 eras in 1190 s
+family_gate_era_census: 2000 eras in 1246 s
+family_gate_era_census: 2100 eras in 1302 s
+family_gate_era_census: 2200 eras in 1358 s
+family_gate_era_census: 2300 eras in 1418 s
+family_gate_era_census: 2400 eras in 1476 s
+family_gate_era_census: 2500 eras in 1531 s
+family_gate_era_census: 2600 eras in 1593 s
+family_gate_era_census: 2700 eras in 1649 s
+family_gate_era_census: 2800 eras in 1711 s
+family_gate_era_census: 2900 eras in 1767 s
+family_gate_era_census: 3000 eras in 1820 s
+family_gate_era_census: 3000 eras 0..3000 in 1820 s on 16 threads -> /srv/builds/_adv-family-gate/logs/lossyplus1-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 1820.39s
+
+Terminated
+lease: released 16 pool cores after 1821 s, exit 0
+end 2026-10-08T11:57:17Z rc 0

docs/analysis/class-v6/logs/lossyplus2-family_gate_era_census-0-3000.log

@@ -0,0 +1,42 @@
+binary 42f2c77ff7e5216e commit fg6 test family_gate_era_census from 0 seeds 3000 threads 16 start 2026-10-08T11:26:56Z
+lease: holding 16 pool cores (34,35,36,37,38,39,40,41,44,45,46,47,48,49,50,51, waited 0 s, class measure): family gate: family_gate_era_census lossyplus2 seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 41 s
+family_gate_era_census: 200 eras in 79 s
+family_gate_era_census: 300 eras in 117 s
+family_gate_era_census: 400 eras in 161 s
+family_gate_era_census: 500 eras in 207 s
+family_gate_era_census: 600 eras in 269 s
+family_gate_era_census: 700 eras in 332 s
+family_gate_era_census: 800 eras in 392 s
+family_gate_era_census: 900 eras in 456 s
+family_gate_era_census: 1000 eras in 526 s
+family_gate_era_census: 1100 eras in 594 s
+family_gate_era_census: 1200 eras in 658 s
+family_gate_era_census: 1300 eras in 726 s
+family_gate_era_census: 1400 eras in 791 s
+family_gate_era_census: 1500 eras in 865 s
+family_gate_era_census: 1600 eras in 936 s
+family_gate_era_census: 1700 eras in 997 s
+family_gate_era_census: 1800 eras in 1051 s
+family_gate_era_census: 1900 eras in 1111 s
+family_gate_era_census: 2000 eras in 1163 s
+family_gate_era_census: 2100 eras in 1219 s
+family_gate_era_census: 2200 eras in 1268 s
+family_gate_era_census: 2300 eras in 1325 s
+family_gate_era_census: 2400 eras in 1385 s
+family_gate_era_census: 2500 eras in 1440 s
+family_gate_era_census: 2600 eras in 1496 s
+family_gate_era_census: 2700 eras in 1547 s
+family_gate_era_census: 2800 eras in 1608 s
+family_gate_era_census: 2900 eras in 1660 s
+family_gate_era_census: 3000 eras in 1722 s
+family_gate_era_census: 3000 eras 0..3000 in 1722 s on 16 threads -> /srv/builds/_adv-family-gate/logs/lossyplus2-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 1722.38s
+
+Terminated
+lease: released 16 pool cores after 1723 s, exit 0
+end 2026-10-08T11:55:39Z rc 0

docs/analysis/class-v6/logs/lossyplus3-family_gate_era_census-0-3000.log

@@ -0,0 +1,42 @@
+binary 42f2c77ff7e5216e commit fg6 test family_gate_era_census from 0 seeds 3000 threads 16 start 2026-10-08T11:26:56Z
+lease: holding 16 pool cores (56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71, waited 0 s, class measure): family gate: family_gate_era_census lossyplus3 seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 63 s
+family_gate_era_census: 200 eras in 122 s
+family_gate_era_census: 300 eras in 186 s
+family_gate_era_census: 400 eras in 258 s
+family_gate_era_census: 500 eras in 332 s
+family_gate_era_census: 600 eras in 407 s
+family_gate_era_census: 700 eras in 486 s
+family_gate_era_census: 800 eras in 564 s
+family_gate_era_census: 900 eras in 648 s
+family_gate_era_census: 1000 eras in 731 s
+family_gate_era_census: 1100 eras in 810 s
+family_gate_era_census: 1200 eras in 901 s
+family_gate_era_census: 1300 eras in 972 s
+family_gate_era_census: 1400 eras in 1042 s
+family_gate_era_census: 1500 eras in 1118 s
+family_gate_era_census: 1600 eras in 1186 s
+family_gate_era_census: 1700 eras in 1252 s
+family_gate_era_census: 1800 eras in 1315 s
+family_gate_era_census: 1900 eras in 1389 s
+family_gate_era_census: 2000 eras in 1457 s
+family_gate_era_census: 2100 eras in 1525 s
+family_gate_era_census: 2200 eras in 1592 s
+family_gate_era_census: 2300 eras in 1658 s
+family_gate_era_census: 2400 eras in 1727 s
+family_gate_era_census: 2500 eras in 1791 s
+family_gate_era_census: 2600 eras in 1848 s
+family_gate_era_census: 2700 eras in 1905 s
+family_gate_era_census: 2800 eras in 1955 s
+family_gate_era_census: 2900 eras in 2007 s
+family_gate_era_census: 3000 eras in 2058 s
+family_gate_era_census: 3000 eras 0..3000 in 2058 s on 16 threads -> /srv/builds/_adv-family-gate/logs/lossyplus3-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 2058.46s
+
+Terminated
+lease: released 16 pool cores after 2059 s, exit 0
+end 2026-10-08T12:01:15Z rc 0

docs/analysis/class-v6/logs/lossyplus4-family_gate_era_census-0-3000.log

@@ -0,0 +1,41 @@
+binary 42f2c77ff7e5216e commit fg6 test family_gate_era_census from 0 seeds 3000 threads 16 start 2026-10-08T11:26:56Z
+lease: holding 16 pool cores (8,9,10,11,12,13,14,15,16,17,18,19,20,26,42,43, waited 0 s, class measure): family gate: family_gate_era_census lossyplus4 seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 62 s
+family_gate_era_census: 200 eras in 121 s
+family_gate_era_census: 300 eras in 184 s
+family_gate_era_census: 400 eras in 259 s
+family_gate_era_census: 500 eras in 331 s
+family_gate_era_census: 600 eras in 410 s
+family_gate_era_census: 700 eras in 497 s
+family_gate_era_census: 800 eras in 578 s
+family_gate_era_census: 900 eras in 666 s
+family_gate_era_census: 1000 eras in 760 s
+family_gate_era_census: 1100 eras in 844 s
+family_gate_era_census: 1200 eras in 936 s
+family_gate_era_census: 1300 eras in 1012 s
+family_gate_era_census: 1400 eras in 1085 s
+family_gate_era_census: 1500 eras in 1156 s
+family_gate_era_census: 1600 eras in 1226 s
+family_gate_era_census: 1700 eras in 1291 s
+family_gate_era_census: 1800 eras in 1364 s
+family_gate_era_census: 1900 eras in 1433 s
+family_gate_era_census: 2000 eras in 1498 s
+family_gate_era_census: 2100 eras in 1565 s
+family_gate_era_census: 2200 eras in 1630 s
+family_gate_era_census: 2300 eras in 1696 s
+family_gate_era_census: 2400 eras in 1767 s
+family_gate_era_census: 2500 eras in 1830 s
+family_gate_era_census: 2600 eras in 1892 s
+family_gate_era_census: 2700 eras in 1953 s
+family_gate_era_census: 2800 eras in 2010 s
+family_gate_era_census: 2900 eras in 2063 s
+family_gate_era_census: 3000 eras in 2096 s
+family_gate_era_census: 3000 eras 0..3000 in 2096 s on 16 threads -> /srv/builds/_adv-family-gate/logs/lossyplus4-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 2096.29s
+
+lease: released 16 pool cores after 2097 s, exit 0
+end 2026-10-08T12:01:53Z rc 0

docs/analysis/class-v6/logs/w1band-family_gate_era_census-0-3000.log

@@ -0,0 +1,42 @@
+binary bc21542a5b421170 commit fg7 test family_gate_era_census from 0 seeds 3000 threads 16 start 2026-10-08T12:27:36Z
+lease: 0 of 8 pool cores free (80 leased), a higher class waits ahead; waiting
+lease: holding 16 pool cores (59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74, waited 381 s, class measure): family gate: family_gate_era_census w1band seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 65 s
+family_gate_era_census: 200 eras in 118 s
+family_gate_era_census: 300 eras in 156 s
+family_gate_era_census: 400 eras in 194 s
+family_gate_era_census: 500 eras in 229 s
+family_gate_era_census: 600 eras in 267 s
+family_gate_era_census: 700 eras in 301 s
+family_gate_era_census: 800 eras in 334 s
+family_gate_era_census: 900 eras in 366 s
+family_gate_era_census: 1000 eras in 400 s
+family_gate_era_census: 1100 eras in 435 s
+family_gate_era_census: 1200 eras in 472 s
+family_gate_era_census: 1300 eras in 518 s
+family_gate_era_census: 1400 eras in 573 s
+family_gate_era_census: 1500 eras in 624 s
+family_gate_era_census: 1600 eras in 672 s
+family_gate_era_census: 1700 eras in 725 s
+family_gate_era_census: 1800 eras in 775 s
+family_gate_era_census: 1900 eras in 824 s
+family_gate_era_census: 2000 eras in 870 s
+family_gate_era_census: 2100 eras in 915 s
+family_gate_era_census: 2200 eras in 963 s
+family_gate_era_census: 2300 eras in 1009 s
+family_gate_era_census: 2400 eras in 1061 s
+family_gate_era_census: 2500 eras in 1110 s
+family_gate_era_census: 2600 eras in 1159 s
+family_gate_era_census: 2700 eras in 1210 s
+family_gate_era_census: 2800 eras in 1261 s
+family_gate_era_census: 2900 eras in 1309 s
+family_gate_era_census: 3000 eras in 1356 s
+family_gate_era_census: 3000 eras 0..3000 in 1356 s on 16 threads -> /srv/builds/_adv-family-gate/logs/w1band-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 1356.41s
+
+lease: released 16 pool cores after 1737 s, exit 0
+end 2026-10-08T12:56:33Z rc 0

docs/analysis/class-v6/logs/w4band-family_gate_era_census-0-3000.log

@@ -0,0 +1,42 @@
+binary bc21542a5b421170 commit fg7 test family_gate_era_census from 0 seeds 3000 threads 8 start 2026-10-08T12:24:08Z
+lease: holding 8 pool cores (8,9,10,11,12,13,14,15, waited 0 s, class measure): family gate: family_gate_era_census w4band seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 107 s
+family_gate_era_census: 200 eras in 197 s
+family_gate_era_census: 300 eras in 249 s
+family_gate_era_census: 400 eras in 304 s
+family_gate_era_census: 500 eras in 358 s
+family_gate_era_census: 600 eras in 411 s
+family_gate_era_census: 700 eras in 464 s
+family_gate_era_census: 800 eras in 519 s
+family_gate_era_census: 900 eras in 574 s
+family_gate_era_census: 1000 eras in 634 s
+family_gate_era_census: 1100 eras in 694 s
+family_gate_era_census: 1200 eras in 754 s
+family_gate_era_census: 1300 eras in 819 s
+family_gate_era_census: 1400 eras in 910 s
+family_gate_era_census: 1500 eras in 1002 s
+family_gate_era_census: 1600 eras in 1086 s
+family_gate_era_census: 1700 eras in 1166 s
+family_gate_era_census: 1800 eras in 1244 s
+family_gate_era_census: 1900 eras in 1323 s
+family_gate_era_census: 2000 eras in 1416 s
+family_gate_era_census: 2100 eras in 1505 s
+family_gate_era_census: 2200 eras in 1603 s
+family_gate_era_census: 2300 eras in 1687 s
+family_gate_era_census: 2400 eras in 1785 s
+family_gate_era_census: 2500 eras in 1890 s
+family_gate_era_census: 2600 eras in 1980 s
+family_gate_era_census: 2700 eras in 2065 s
+family_gate_era_census: 2800 eras in 2158 s
+family_gate_era_census: 2900 eras in 2252 s
+family_gate_era_census: 3000 eras in 2328 s
+family_gate_era_census: 3000 eras 0..3000 in 2328 s on 8 threads -> /srv/builds/_adv-family-gate/logs/w4band-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 2328.38s
+
+Terminated
+lease: released 8 pool cores after 2328 s, exit 0
+end 2026-10-08T13:02:56Z rc 0

docs/analysis/class-v6/logs/w4shape64-family_gate_era_census-0-3000.log

@@ -0,0 +1,42 @@
+binary e7a50f8dcf57a14b commit fg5 test family_gate_era_census from 0 seeds 3000 threads 16 start 2026-10-08T12:27:03Z
+lease: 0 of 8 pool cores free (88 leased); waiting
+lease: holding 16 pool cores (16,17,21,22,23,24,25,26,80,81,82,83,84,85,86,95, waited 152 s, class measure): family gate: family_gate_era_census w4shape64 seeds 0+3000
+
+running 1 test
+test accept::tests::family_gate_era_census ... family_gate_era_census: 100 eras in 73 s
+family_gate_era_census: 200 eras in 131 s
+family_gate_era_census: 300 eras in 194 s
+family_gate_era_census: 400 eras in 258 s
+family_gate_era_census: 500 eras in 324 s
+family_gate_era_census: 600 eras in 383 s
+family_gate_era_census: 700 eras in 448 s
+family_gate_era_census: 800 eras in 517 s
+family_gate_era_census: 900 eras in 586 s
+family_gate_era_census: 1000 eras in 655 s
+family_gate_era_census: 1100 eras in 726 s
+family_gate_era_census: 1200 eras in 792 s
+family_gate_era_census: 1300 eras in 863 s
+family_gate_era_census: 1400 eras in 927 s
+family_gate_era_census: 1500 eras in 990 s
+family_gate_era_census: 1600 eras in 1058 s
+family_gate_era_census: 1700 eras in 1126 s
+family_gate_era_census: 1800 eras in 1200 s
+family_gate_era_census: 1900 eras in 1274 s
+family_gate_era_census: 2000 eras in 1356 s
+family_gate_era_census: 2100 eras in 1438 s
+family_gate_era_census: 2200 eras in 1521 s
+family_gate_era_census: 2300 eras in 1603 s
+family_gate_era_census: 2400 eras in 1678 s
+family_gate_era_census: 2500 eras in 1755 s
+family_gate_era_census: 2600 eras in 1825 s
+family_gate_era_census: 2700 eras in 1888 s
+family_gate_era_census: 2800 eras in 1948 s
+family_gate_era_census: 2900 eras in 2007 s
+family_gate_era_census: 3000 eras in 2073 s
+family_gate_era_census: 3000 eras 0..3000 in 2073 s on 16 threads -> /srv/builds/_adv-family-gate/logs/w4shape64-family_gate_era_census-0-3000.tsv
+ok
+
+test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 80 filtered out; finished in 2073.12s
+
+lease: released 16 pool cores after 2226 s, exit 0
+end 2026-10-08T13:04:09Z rc 0