Dated observation note. Recorded 2026-08-06, ~5.5 hours after activation. Method and caveats below; raw per-block JSON alongside the numbers.

Status: preliminary (94 post-fork blocks). Re-run at +1,000 and +5,000 blocks before publishing anything externally.

What was measured

Certificate public data for heights 96,151–96,344 (194 blocks: 100 pre-fork, 94 post-fork; 96,345 excluded for a fetch race at the then-tip), parsed from raw blocks served by our own Blockbook (/api/v2/rawblock/), layout per Keshi's internal protocol notes (Certificate internals). Pool attribution via coinbase payout address against the known wallets in Keshi's internal protocol notes (Known pools); the unattributed remainder is shown, never redistributed. Reproduce with scripts/fork-census.py <lo> <hi> <out.json>.

Consensus rule under test (v1.3.0, activation 96,251): reject rank < 128; difficulty bound multiplier 128/rank (neutral at 128, 8× tighter at 1024).

Findings

1. The network snapped to the penalty-neutral point within the first blocks.

k collapsed with it: pre-fork mode 32,768; post-fork k = 2,048 in 59/94, which is exactly 16·rank, the minimum legal k for rank 128. Miners moved to the cheapest-valid configuration, not merely a valid one.

2. Zero invalid-under-new-rules blocks on the canonical chain. No post-fork certificate shows rank < 128. The transition was clean chain-side. (Whether any miner produced such a block and had it rejected is unrecorded: the v1.2.1 fork-observer was declined; see ROADMAP §P0b.)

3. Pool block share shifted sharply across the boundary (windows of ~5.4 h pre / ~5.2 h post; small n, so treat as indicative):

Kryptex mined rank 1024 in 43 of its 57 pre-fork blocks and was still submitting 4–8×-penalized ranks (512/1024) in ~35% of its post-fork blocks, consistent with slow fleet reconfiguration, and consistent with (but not proof of) the share collapse being self-inflicted by the penalty. PearlHash, already mostly at low rank pre-fork, absorbed the share. This is the per-pool migration latency Phase 9.2 predicted would be measurable.

4. Corpus-hygiene invariants held on all 194 blocks: reserved 28-byte trailer all-zero everywhere (no covert channel), mma_type = 0, zero MoE certificates (e = 0 throughout, consistent with the MoE-was-never-used hypothesis), cert_version = 2 throughout.

Errata 2026-08-06 (OBS-003): the "consistent with MoE-was-never-used" inference above was unsound: this note's window is entirely after the dense-only fork, where zero MoE is consensus-required and supports no usage inference. The full-corpus census found 1,929 MoE certificates inside the 71,935–91,630 window: the capability WAS used in production.

5. Shape-provenance signals (feeds Phase 8, not conclusions):

• A tile pattern h·w = 144 appears (2 pre, 2 post), a third kernel variant, outside both the official fragment (128) and the 256 seen in the fixture blocks. The n=3 fixture sample undercounted kernel diversity.
• m/n values are dominated by exact powers of two and, notably, 2^t + 1 values (513, 1025, 4097, 8193, 32769, 65537, 131073, …, 8388609 = 2^23+1). Live-batch token counts do not look like this. No (n, k) pair in either window matches any published pearl-ai model layer.
• Blocks passing the loose official-stack screen (rank ∈ {64,128} ∧ h·w = 128 ∧ m ≥ 1024): 3/100 pre-fork → 17/94 post-fork. Caveat: post-fork rank 128 is forced, so this screen weakens after the fork by construction; the exact [0,8]/stride-8 fragment check (T1 proper) is what matters and needs the full-pattern comparison, not just the span.

Caveats

• 94 post-fork blocks ≈ 5 hours. Shares at this n have wide variance; the rank crosstab is hard data, the share-shift interpretation is not settled.
• Attribution is single-known-wallet per pool; proxy-payout arrangements would misattribute.
• Chain-side only: no orphan, reject, or peer-version data was captured for the transition window (the live watch did not run; observer declined).
• Parsed via the prototype decoder path, not yet the dual-decoder discipline Phase 6 requires. Treat as provisional until the Go decoder + golden vectors reproduce it.

Data

• fork-census-96151-96345.json (per-block parsed certificates + attribution), produced 2026-08-06 from the local Blockbook; script scripts/fork-census.py.

Re-run at +1,000 blocks (2026-08-09)

Pre-committed follow-up per PLAN-001 Step 6, executed at tip 97,530 (the +1,000 point h97,251 was reached earlier than the plan's ≈08-11 estimate; measured post-fork block interval 3.80 min over 96,251→97,530). Same script, window 96,151–97,250 (100 pre-fork + 1,000 post-fork); data: fork-census-96151-97250.json. Reproducibility check: the re-parsed pre-fork window is identical to the original tables above (rank distribution and pool shares match exactly).

1. Convergence deepened; violations stayed at zero. Post-fork rank distribution (n = 1,000): 128 → 949 (94.9%), 256 → 21, 512 → 12, 1024 → 18, vs 84% at the +94 preliminary. rank < 128 (invalid): 0. Penalized ranks (>128) total 51/1,000 (5.1%), concentrated early (first 96,263, median 96,454) but not extinct; the last is at 97,249. Kryptex: 23 penalized of its 445 blocks, all but one before h96,851 (the exception: h97,218), refining the 2026-08-08 audit's "fully compliant by 96,851" to "compliant with one later exception in this window".

2. First per-pool boundary-hugging measurement (Phase 14.3). Share of post-fork blocks at the exact legal minimum configuration (rank = 128 ∧ k = 2,048 = 16·rank): 471/1,000 (47.1%) overall:

Two distinct legal optimization profiles coexist: PearlHash/PearlFortune sit at the k-minimum; Kryptex mines rank 128 with k up to the legal maximum for that rank (k ≤ 4r² = 65,536; its post-fork k mode is 4,096–8,192). Measured fact only; which profile is economically favored feeds the roadmap's open rank-economics question, and no mechanism is asserted here.

3. Pool shares over the 1,000-block window (~2.6 days): Kryptex 445 (44.5%, recovered from 28% in the +94 window), PearlHash 354 (35.4%), HeroMiners 64, LuckyPool 44, PearlFortune 35, unattributed 58 blocks across 33 distinct addresses (5.8%).

4. Shape and software fingerprints.

• Tile fragments post-fork: (16,16) 653 · (8,16) 241 · (2,64) 79 (the official wgmma fragment, 7.9%) · (8,32) 14 · (6,24) 11 · (12,12) 1 · (32,8) 1. The "h·w = 144" variant first seen in the preliminary resolves into two distinct patterns: (6,24) and (12,12), so the window holds at least seven fragment geometries ≈ a lower bound on distinct mining software (feeds Phase 14.4; tile patterns identify software, not honesty).
• The 70B fused gate/up shape (m 32,768 · n 57,344 · k 8,192) persists at 37/1,000, of which 31 carry the official (2,64) fragment, and all 37 are unattributed (30 distinct addresses; zero from labeled pools). The official-pattern fleet survives at a few percent, entirely in the diffuse unattributed stratum.
• Synthetic dimension signatures persist: post-fork m is 70.4% exact powers of two and 23.8% 2^t + 1 (94.2% combined); n is 45.0% + 24.4%. New dominant synthetic shape: n = 516,096 (m 4,096, k 2,048), 130/1,000 blocks: 124 from PearlHash, 6 from one unattributed address.
• Hygiene invariants held on all 1,100 blocks: reserved trailer all-zero, moe_e = 0 throughout (consensus-required post-91,630; usage inference lives in OBS-006, not here), zero post-fork rank < 128.

Numbers in this section were computed 2026-08-09 from fork-census-96151-97250.json; the official-fingerprint screen retains the preliminary's caveat (rank 128 is consensus-forced post-fork, so the loose screen weakens by construction; the exact-fragment comparison is the meaningful one). Next re-run: +5,000 (h101,251), ETA ≈ 2026-08-19 at the measured 3.80 min interval.