The first 16 GB card and the first Thunderbolt card. The app's first poll listed it (vendor nvidia, 16,311 MiB,
driver 610.47 shared with the 5090, kind discrete: the app does not know it is an eGPU) and started a miner on
it by itself. Job run-5060ti-bench-20261007-b (tools/bench-5060ti/pc2-5060ti-bench.ps1, the card alone under
the runner's --cards-off, the 5090 mining, no power limit written: 180 W default read back unchanged, PC 2 lost
power twice that day): G1 PASS with 867dbc45cfb36b4d (sub-version 1) and 90f794dd556f7a3b equal to the Mac's;
the window 1,105 dispatches of 2^24 in 602 s at 30.882 MH/s, 114.8 W mean at utilisation 90 percent and over,
0.269 MH/W, PCIe 4.0 x4 through the enclosure (the hash is latency-bound, the link costs nothing measurable).
Owed: the efficient point (the app's Ember Tune; PC 2's Power Helper refuses every request since the restart)
and the prove-beside read (the shipped WSL2 host has no IGNEUM_CUDA_DEVICE selector; the floor server is on
PC 2; the prover-floor host is the cut). The kit (tools/bench-5060ti/make-kit.sh) takes the sub-version 1 packs
from ca3-v4-amend aa31c396, since master's packs-ca3-v4 is the 6 October stream.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The node pin moves to 4c6b129d (release-0.3.14-node), which is what the live payload inputs carry, so windows-ci's payload-inputs step is green again (red since e9eca23 on master's 0.3.13 pin against the 0.3.14 inputs).
Conflicts: infra/fast-time/override-60x.json keeps master's side (the fresh-rule field was already there at u64::MAX; the release line would be a duplicate key); tools/ci/playbook-quit-check.sh keeps master's rule 2 and pre-rule list with the release side's Ember allow entry and the 0.3.15 expiry check.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PC 2 job run-ca3-shadow-pc2-20261006 (card empty, every pack bit-exact against the Mac): the 5090 holds its rate to
150,800 ops per hash and loses 2.7 percent at 199,600 under the app's 431 W cap, which binds from 102,100 ops up and
takes the clock from 3,037 to 1,834 MHz (86 MH/s at 330,700 ops); 350 W at the control, 2.65 to 3.27 microjoules per
hash; marginal ALU energy 10 to 13 pJ per counted op. Clock rows OWED (nvidia-smi refused -lgc without rights). Chip
side at N = 100,000 and k = 1: 2.1x over the 5090 on GDDR7, 0.9x over the M5 Max. GO at mx8+sh256x27.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The M5 Max ladder (Metal, packbench, IOReport GPU and DRAM watts without root): latency-bound to about 100,000 ops per
hash, the 5 percent point about 130,000, 11 to 27 W GPU at 100,000 ops, 0.78 to 1.40 microjoules per hash; the
verifier's law 2.06 ms + 3.2 us per 1,000 shadow instructions per warp on one core; every pack bit-exact. The
analysis file with the knob, the chip side (k = 1, 1.5, 0.3), the gates and the consequences; the bench-log entry;
the 5090 rows pending the PC 2 job (the playbook now carries the core-clock rows and the sh256x40 rung).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
docs/plans/counter-asic-3-derivation.md (the design, the acceptance test, the interpreter, the allowance argument,
the measurements, the PROPOSED reserve entry R0 for 1.13.2, what is owed), docs/analysis/chip-model-v3.md section 6
(the per-day derivation rows at 1.0x to 3x allowances), the bench-log entry, relay/playbooks/ca3-derive-pc2.ps1
(one PC 2 job: self-fetched packs zip, the installed worker through NVRTC, the card off only under test with its
key from settings.json). Verifier 4.875 / 4.944 ms per unit on one M5 Max core under the measure lock against
x8's 2.061 / 2.063; Metal build 29 ms against 22; hash rate equal; bit-exact on Metal and Apple OpenCL.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The statement and the pinned guests are unchanged; every fixture proof verifies as before. The defaults stay (batch-log2 22, SP1 defaults): the one knob that moves a mining card's prover costs a fifth of the hash rate; the plan carries the trade for the project lead and the batch fold for the next pin. Measured: docs/bench-log.md "aggregation cost on the RTX 5090"; the plan line: docs/plans/proving-v1.md "Aggregation cost (5 October, night)". Also: make-package's gate skips the exporter's .node-plan.json side files and takes the run lock for its execute step; the state-reply class (/api/state answering {} once paid_wei passes u64::MAX) found on the way and fixed on the app branch at 8e07279.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
(cherry picked from commit ea38ece9eaa5949dd657cbfea5308c4948177ff8)
Source: the scratch engine's own log in collect ember-c35-collect-1 (06:59Z): 22:31:02Z '0.3.10 is available: downloading',
22:31:05Z 'update: starting the installer first ... ota-apply.ps1', and the installed app's 'quit:' at 22:31:06Z.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
serde_json's to_value refuses a u128 over u64::MAX (18.45 IGN) and state_json turned the error into json!({}). A paid shard is 1.23 IGN on average, so a proving machine's dashboard went blank about 15 paid shards after every app start. Unit test over the boundary; 114 app tests pass.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
