adv-accept-3: the 0.995-ratio seeds at 2^24 nonces (PASS)

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

docs/analysis/cryptanalysis/report-acceptance-rule-3.md

@@ -22,7 +22,7 @@ internal adversarial pass, not an independent review
 | Q1b | The last-resort program: weak, constant, predictable | `last_resort_v4(candidate(b, 256))` through the library on 3,000 seeds, the REAL rule run on each | `adv3 exhaust-mirror`: a reject-everything loop must hit the cap and print the last resort | last resort characterised; no predictable or weak accepted program | mirror fired (cap 256, byte-identical to the library's). 2,500 seeds: the real rule REJECTS 223 of 2,500 last-resort programs (8.9 percent): 209 by (a) cyclic stale load, 10 by (b), 4 by (c) distinct sum; 2,500 distinct programs; 0 lossy ops; chi-square against the weights 39 (mean), largest op share 37 percent | FINDING (correctness, unreachable at 4.6 x 10^-44): the "accepted as drawn" last resort fails the rule's own (a) in 1 of 12 seeds; 500 seeds to go |
 | Q2 | Steering with the full rule: tries per quantile of each property | `adv3 steer`, the real draw per seed, min per-site ratio at 256 units plus static properties; 2 x 10^4 seeds queued | `--plant hot`: two `or` writers before a load's source must put the planted set in the lowest quantile | tabulated; no property reaches a 1.1x chip gain under 10^5 tries | plant: 45 of 48 planted rows at ratio 0.207 to 0.993 against a population minimum of 0.999 at 256 units (3 did not fire: the scan found no ALU writer to change); the real rule rejects every effective plant by (a'). First 975 seeds (part 1): min ratio 0.998, median 0.999; 18 of 18 chain re-draws equal; 0 last-resort programs; per-attempt rejects (a') 1,820, (a) 257, (b) 87, (c'') 26, (c) 19 of 3,184 candidates | RUNNING (sweeps 93, 96 chunked) |
 | Q2b | Static steering at 10^6 seeds | `adv3 static`: the first candidate passing (a), (b), (a') per seed, static properties, quantiles to 10^-6 | the plant of Q2 (the same property code) | the 10^-5 and 10^-6 quantiles of each property, priced | DONE: table in section 3; the best property at 1 in 10^6 tries is a shadow block with 38 multiplies of 256 (mean 74), worth about 2 to 3 percent of the f = 1 chip's energy per hash (approximate); the load critical path at 40 of 128 (median 72) is worth nothing at the memory activate ceiling | BOUND |
-| Q2c | Live hot set of the lowest-ratio steered seeds | f8 `warps` on the lowest-ratio seeds of Q2 | f8's own const-item plant | the worst steered seed's gain priced | plant FLAGGED (X_0.1% +6.25, 24.9x); control clear; the five lowest of 975 (ratio 0.998): within the 1.2x gate (1.01x to 1.13x), no hot set, 6-sigma flagged (+32 to +164); the three lowest of 4,975 (ratio 0.995: seeds 2653, 3296, 4388): PASS with no test fired (1.0007x to 1.0008x, X_0.1%/f 0.002, 6-sigma +4.3 to +4.8, the control's level); worst case over the eight: 1.0004x | BOUND (10^6 nonces; 2^24 running) |
+| Q2c | Live hot set of the lowest-ratio steered seeds | f8 `warps` on the lowest-ratio seeds of Q2 | f8's own const-item plant | the worst steered seed's gain priced | plant FLAGGED (X_0.1% +6.25, 24.9x); control clear; the five lowest of 975 (ratio 0.998): within the 1.2x gate (1.01x to 1.13x), no hot set, 6-sigma flagged (+32 to +164); the three lowest of 4,975 (ratio 0.995: seeds 2653, 3296, 4388): PASS with no test fired (1.0007x to 1.0008x, X_0.1%/f 0.002, 6-sigma +4.3 to +4.8, the control's level); worst case over the eight: 1.0004x | BOUND (the 0.995 seeds also PASS at 2^24: 0.9998x to 1.0002x; seed 620 at 2^24 running) |
 | Q3 | Program id: collisions and derivation agreement | `adv3 ids` over 10^7 (seed, attempt) pairs; `adv3 idcheck` by three derivations | low-32 collision counts must match the birthday expectation; one attempt bit flipped must change the id | 0 collisions; the derivation that matches the packs named | 0 id collisions, 0 program-stream state collisions, 0 seed-word collisions over 10^7 pairs (control: 11,759 low-32 id collisions against 11,641 expected); FINDING: program.json's `program_id_derivation` string and spec 1.4.6 state a derivation WITHOUT the `"sub/" || 3_le16` suffix the code appends for generator 4; an implementation written from the text computes 30956569d8f3d8d7 for Devnet 3 where the pack carries fce15bf61030be57 | FINDING (interoperability, not consensus) |
 | Q4 | Determinism traps | `adv3 margins`; a second interpretation of the rule (`adv3 verdicts`, 2,000 seeds, every attempt); a spec-1.4.6-only interpretation (400 seeds); the read list | `--variant floor97` must change at least one verdict | 0 disagreements between code and the faithful second interpretation; the spec-only disagreement rate measured; every divergence listed | margins: the f64 ratio compare of (c'') never disagrees with the integer rule `50 d >= 49 E` on any of the 2^20 + 1 values of d for any window (margins 0.32 to 0.44 counts); second interpretation: 3,411 of 3,411 attempt verdicts of 1,056 seeds agree with the code, part by part (the same 1,445 (a'), 204 (a), 70 (b), 20 (c''), 7 (c') saturated-source, 7 constant-bit, 1 distinct-sum rejections and 792 accepts), chosen attempt equal on every seed; read list in section 6 (the spec text is behind the code on six points; the CLI `accept` command caps at 32 with no last resort) | RUNNING (sweeps 94, 97, 98 chunked) |
 | Q5 | Era steering | `adv3 era-steer` over 400 era seeds: stride, interleave, epoch 0's min ratio | none needed (a listing) | noted | queued | RUNNING (sweep 99) |
@@ -163,6 +163,14 @@ Part 2 (`992-adv-accept-3-f8-live-2.sh`, 21:33 BST, the three lowest-ratio seeds
 | 3296 (eb47ecbc0e86d847, attempt 2, site 1) | 1.0007x | +0.0002%, 0.002 | +4.77 within | PASS |
 | 4388 (5cd329c25bc6986b, attempt 0, site 6) | 1.0007x | +0.0002%, 0.002 | +4.78 within | PASS |
 
+At 2^24 nonces (21:34 to 21:37 BST, same log), where the gate is sharper and adv-accept read a base rate near 1 in 10 beyond 1.2x:
+
+| Seed | Top 0.1% over the window model | X_0.1%, X_f/f | Windowed 6-sigma | Verdict |
+|---|---|---|---|---|
+| 2653 | 0.9998x | -0.00003%, -0.0003 | +4.55 within | PASS |
+| 3296 | 1.0001x | +0.00001%, +0.0001 | +4.88 within | PASS |
+| 4388 | 1.0002x | +0.00003%, +0.0003 | +5.16 within | PASS |
+
 So the three programs with the LOWEST stand-in ratio of 4,975 are indistinguishable from the clean control live, while the five at 0.998 all flagged the 6-sigma test. At 256 units the ratio's own sampling noise (65,536 evaluations per site against a window of 2^26 to 2^28 words; the expected distinct count is within 0.1 percent of N, so a 0.5 percent shortfall is about 300 missing indices, 1.2 sigma of a Poisson count of that size) is the same order as the signal. The 256-unit proxy is a weak selector: it enriches the live-lumpy tail (adv-accept's 6 of 8 against a base rate near 1 in 10 at 2^24, its own correction) but its floor is noise. A grinder's real selector is the live census itself at about 3 s per seed on 32 cores, which this lane did not run at scale; its cost per seed is the number that prices the steering (10^4 seeds = 8 box-hours at this rate), and what it buys is bounded by adv-accept's worst live case, 1.002x.
 
 Reading. The selector (the stand-in ratio at 256 units) picks programs whose live item histogram is measurably lumpy (every one flags the windowed 6-sigma where the clean control sits at +5.5), and none of the five has a hot set: the best, seed 620, puts 0.326 percent of its reads into the top 0.1 percent of items against 0.289 percent expected, so a 1 MB on-die copy of those items removes 0.037 percent of DRAM reads, a gain of 1.0004x on the f = 1 chip of chip-model-v3 section 5. Steering cost for this property at this sample: 5 of 975 seeds (1 in 195) reach ratio 0.998; none of the 5 reaches the f8 gate at 10^6 nonces. Lane adv-accept's wider search (16,337 seeds, 2^24 nonces) found its worst at 2.05x over the window model and X_0.1% +0.155 percent, 1.002x for the chip; the two lanes' numbers agree on the order: a seed grinder with 10^4 to 10^5 tries buys a hot set worth under 0.2 percent of reads. At 2^24 nonces the gate is sharper (adv-accept: base rate near 1 in 10 beyond 1.2x among random accepted programs), so these five at 2^24 are owed and queued behind the sweeps.