The GPU network ledger

Every live GPU-mined chain we could pin, measured the same way on the same rented cards in the same sessions, with the same chip-design method applied to each, and ranked. Igneum sits in the table as one row among the others, under the same rules. Every number is measured or a labelled model with its source; a number that does not exist yet says "pending" and names who owes it; a chain we could not run in-house says NOT RUN and why. The full record with the lane bookkeeping is in the repository at docs/analysis/class-v6/1p5x/ledger/ (README.md and ledger.json); this page is rendered from the same data file.

Data: ledger.json, stamp 2026-10-09, adversarial set v3 (the lab artefact tree (not in the repository): lab/adv/ (README.md, design-notes/adversarial-sets-v1-v2-v3.md, rows/, SHA256SUMS); set v3 commit b2b9ee0b9 on class-v6-adversary).

The method, the same for every chain

The object. Each chain's proof of work as its own pinned reference code, built from source (no downloaded miner binaries):

Igneum's frozen class v6 at its live 4 GiB dataset; the official ProgPoW 0.9.4 reference kernel for Ravencoin's KAWPOW (which is ProgPoW 0.9.4 with Ravencoin's constants, so the row is a labelled proxy) and the same row for Firo's FiroPoW; Iron Fish's FishHash reference library with a bit-exact CUDA port; the pinned chfast/ethash library's hashimoto at Ethereum Classic's current epoch; Ergo's Autolykos2 NOT RUN (the reason on the row).

The card rows. One rented RTX 4090 and one rented RTX 5090 on uncapped hosts, every chain back to back in one session per

card at stock clocks; board power sampled at 1 Hz (idle not subtracted); nJ per hash = watts over MH/s; three runs, the median with its spread; the dataset size and the card's clock on every row; a read-only twin (the same memory reads, the arithmetic cut) where the harness has one, giving the card's memory share f.

The chip rows. For each chain, the cheapest credible specialist chip on the same process node as the card, as a complete

machine, modelled and labelled as a model (never measured): the DRAM-board chip is the energy row, the full-SRAM die the residual. Two conventions beside every ratio: ours (the cheapest programmable core, fused where the work allows) and ProgPoW's own premise (the chip must keep a GPU-class datapath, so its edge is 1.1x to 1.2x on the compute term, weighted by f).

The real chip. Ethereum Classic is the one chain here with a chip on the market, the Antminer E9 Pro; its published rate

and power sit on the row, and our modelled chip for Etchash sits beside it as the calibration of the model.

The two phases. Before investment: can a chip maker recover its spend under growing, flat and falling prices. After, with

the spend sunk: who exits first on operating cost, with each chain's actual reward split. On Igneum a GPU earns the mining share plus the proving share (80 / 20); a hash chip earns the mining share only. On every other chain here the GPU earns mining only, so the second-income term is zero by construction.

The ranking. (a) ascending on the DRAM-board ratio under our convention at the worse of the two cards (the lower ratio is

the smaller chip edge); (b) the same under ProgPoW's premise; (c) whether the chain's GPUs earn an income a hash chip cannot. A chain with a pending or proxy cell is ranked provisionally and says so; a NOT RUN chain is unranked.

The card rows (measured, board power, stock)

chain	algorithm	dataset on the row	reward split (mining / proving)	RTX 4090 stock: nJ per hash (MH/s, W, f)	RTX 5090 stock: nJ per hash (MH/s, W, f)	twin	sources
Igneum	class v6, the live signing object 0x2a1d6caab4c24564 (2.0 devnet, chain id 4465)	4.29 GB (4,294,967,296 bytes); 2^30 words, the live policy (70a6c703); 256 random 4-byte reads per hash, 62,120 instructions per hash (instruction convention)	80 / 20 (the Token Value volume on master (D04, token-value/README.md: the 80/20 emission allocation); PROVING_POOL_SHARE_PERCENT)	7,922 MEASURED on a THROTTLED host; THROTTLED HOST: the pod's sampler read event reason 0x20 (software thermal slowdown) through the run; the rate holds under the memory bound while the clock falls, so the joules are the throttled host's, not the card class's; COHORT's repeat on a fresh uncapped 4090 in flight (rows near 15:50 UK) (30.68 MH/s, 243.1 W, f 0.982; spread 2.05%; vast 55017318, driver 580.119.02, 450 W host, stock; stock; the lock refused (the host's container); the clean triple after a first triple at 17.9 percent spread (8,518.5) repeated under the lab's rule)	6,740 MEASURED (68.67 MH/s, 462.8 W, f 0.813; spread 0.76%; vast 55018470, driver 595.71.05, 600 W host, stock; stock; the lock refused; the clean triple after a first triple at 7.9 percent spread (6,410.9) repeated under the lab's rule)	yes (the same address stream, the arithmetic cut)	RTX 4090: the lab artefact tree (not in the repository): lab/ledger/vast-55017318/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json); RTX 5090: the lab artefact tree (not in the repository): lab/ledger/vast-55018470/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json)
Ravencoin	KAWPOW (ProgPoW 0.9.4 with Ravencoin's kiss99 constants, 7,500-block epochs and a 512-parent DAG; live since block 1,219,736, 6 May 2020)	6.18 GB (6,182,403,712 bytes); KAWPOW's current DAG size at Ravencoin height 4,573,707 (epoch 609, read 13:0x UK): the official ProgPoW 0.9.4 kernel built at Ethash block 18,270,000 (epoch 609, the same byte count), harness progpow-094/gpu/pp_harness_18270000; KAWPOW's own constants NOT applied, so the row is 'ProgPoW 0.9.4 official at KAWPOW's DAG size (proxy)'; 64 coalesced 256-byte reads per hash; F1 GPU = CPU on both nonces (cpu/vectors-094-kawpow-size-18270000.txt)	100 / 0 (Ravencoin's block reward goes to the miner in full (no second income to GPUs))	5,411 MEASURED on a THROTTLED host (the official ProgPoW 0.9.4 kernel at KAWPOW's DAG size: a PROXY for KAWPOW); THROTTLED HOST: the pod's sampler read event reason 0x20 (software thermal slowdown) through the run; the rate holds under the memory bound while the clock falls, so the joules are the throttled host's, not the card class's; COHORT's repeat on a fresh uncapped 4090 in flight (rows near 15:50 UK) (58.53 MH/s, 316.7 W, f 0.748; spread 1.65%; vast 55017318, driver 580.119.02; stock; clocks_sm_median 638 MHz = the throttled clock (the sampled bench only, from 8 s in; the DAG build is outside the window); the rate 58.5 MH/s holds because the kernel is memory-bound at the DAG)	5,708 MEASURED (the official ProgPoW 0.9.4 kernel at KAWPOW's DAG size: a PROXY for KAWPOW); SPREAD 36.8 PERCENT from the first pass's sampled window taking in the light-cache and DAG build (379 W against 600 W at the limit on the next two): the repeat with a 15 s skip and a 60 s bench after the build in flight (COHORT) (105.12 MH/s, 600.0 W, f 0.715; spread 36.84%; vast 55018470, driver 595.71.05; stock; clocks_sm_median 2,737 MHz; the card at its 600 W limit)	yes, by construction (progpow-094/gpu/pp_twin 18270000 26 250: the 64 dependent entry reads per 16-lane hash)	RTX 4090: the lab artefact tree (not in the repository): lab/ledger/vast-55017318/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json); RTX 5090: the lab artefact tree (not in the repository): lab/ledger/vast-55018470/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json)
Firo	FiroPoW (ProgPoW 0.9.4 with Firo's constants; live since block 419,264, 26 October 2021)	= Ravencoin's row	100 / 0	= Ravencoin's row	= Ravencoin's row	as Ravencoin	as KAWPOW: the pinned ProgPoW 0.9.4 reference; FiroPoW differs from KAWPOW only in its constants and DAG schedule, so the lab serves ONE row for both and says so
Iron Fish	FishHash (live since hardfork 1, April 2024)	4.83 GB (4,831,835,776 bytes); 37,748,717 items of 128 bytes, FIXED (no epochs); 96 random 128-byte reads per hash, 9,856 lane-instructions	100 / 0 (Iron Fish's block reward is mining only)	12,664 MEASURED on a THROTTLED host; THROTTLED HOST: the pod's sampler read event reason 0x20 (software thermal slowdown) through the run; the rate holds under the memory bound while the clock falls, so the joules are the throttled host's, not the card class's; COHORT's repeat on a fresh uncapped 4090 in flight (rows near 15:50 UK) (19.68 MH/s, 249.2 W, f none; spread 1.16%; vast 55017318, driver 580.119.02; stock; clocks_sm_median 2,775 MHz)	7,993 MEASURED (44.72 MH/s, 357.4 W, f none; spread 0.56%; vast 55018470, driver 595.71.05; stock; clocks_sm_median 2,895 MHz)	yes (a trace replay of the kernel's own item indexes; its energy read as a bound: the twin runs 10 percent slower than the full kernel)	RTX 4090: the lab artefact tree (not in the repository): lab/ledger/vast-55017318/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json); RTX 5090: the lab artefact tree (not in the repository): lab/ledger/vast-55018470/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json)
Ethereum Classic	Etchash (Ethash with ECIP-1099's 60,000-block epochs; live since block 11,700,000, November 2020)	4.64 GB (4,638,898,816 bytes); ETC height 25,502,946 (read 12:5x UK) = Etchash epoch 425 (ECIP-1099: height / 60,000), the harness's block argument 12,750,000 (the Ethash epoch with the same size table index); 64 coalesced 128-byte reads per hash (two 64-byte half-mixes); CONTROLS' rate-only read on a 4090 at stock 29.90 MH/s full and 29.90 twin (memory-bound)	100 / 0 (ETC's block reward is mining only)	7,363 MEASURED on a THROTTLED host; SPREAD 27.4 PERCENT: the repeat owed by COHORT under the lab's rule; THROTTLED HOST: the pod's sampler read event reason 0x20 (software thermal slowdown) through the run; the rate holds under the memory bound while the clock falls, so the joules are the throttled host's, not the card class's; COHORT's repeat on a fresh uncapped 4090 in flight (rows near 15:50 UK) (29.77 MH/s, 219.2 W, f none; spread 27.44%; vast 55017318, driver 580.119.02; stock; the three runs at 630 / 2,715 / 2,550 MHz and 176 / 242 / 219 W: the throttle came and went (the 27.4 percent spread))	5,559 MEASURED (66.27 MH/s, 368.4 W, f 0.988; spread 2.79%; vast 55018470, driver 595.71.05; stock; clocks_sm_median 2,887 MHz)	yes, by construction	RTX 4090: the lab artefact tree (not in the repository): lab/ledger/vast-55017318/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json); RTX 5090: the lab artefact tree (not in the repository): lab/ledger/vast-55018470/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json)
Ergo	Autolykos2 (k = 32, n = 26, Blake2b-256; the table grows 5 percent every 51,200 blocks)	7.27 GB (7,272,058,080 bytes); the live table at Ergo height 1,890,820 (api.ergoplatform.com, 13:0x UK): N = 227,251,815 elements of 32 bytes (+5 percent at height 1,894,400; the row names its height); 33 random 32-byte reads per hash; the table builds on the device in 1.5 s; CONTROLS' rate-only read on a 4090 at stock 216.2 MH/s full, 201.8 twin	100 / 0 (Ergo's block reward is mining only)	pending: pending: the session's bench was too short for the sampler (250 batches of 2^20 ran 1.2 s on the 4090 and 0.57 s on the 5090, under the sampler's warm-in: a sizing fault of the wrapper, not the harness; the rates read 216.3 and 458.0 MH/s, 1,056 bytes read per hash); the re-run at 2^22 x 4,000 (about 60 s) in flight, rows near 15:50 UK (rate only 216.20 MH/s)	pending: pending: the session's bench was too short for the sampler (250 batches of 2^20 ran 1.2 s on the 4090 and 0.57 s on the 5090, under the sampler's warm-in: a sizing fault of the wrapper, not the harness; the rates read 216.3 and 458.0 MH/s, 1,056 bytes read per hash); the re-run at 2^22 x 4,000 (about 60 s) in flight, rows near 15:50 UK	yes (a trace replay)	RTX 4090: the lab artefact tree (not in the repository): lab/ledger/vast-55017318/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json); RTX 5090: the lab artefact tree (not in the repository): lab/ledger/vast-55018470/ (rows.jsonl, run.log, out/, SHA256SUMS, sidecar.json)

The chip-to-GPU edge, both conventions

chain	the cheapest credible same-node DRAM board (set v3, nJ per hash, band)	ratio at the RTX 4090 stock (ours)	ratio at the RTX 5090 stock (ours)	ProgPoW-premise ratio at the RTX 4090	ProgPoW-premise ratio at the RTX 5090	the die as the residual (nJ; ratio at the worse card)	real chip on the market	source
Igneum	1,542 (1,329 to 1,941); GDDR7 board + register file sized to the work + fused operations, N5; split memory 527 / ALU 166 / tax 849; WITNESS (modelled, complete machine, instruction convention)	5.14x (4.08 to 5.96)	4.37x (3.47 to 5.07)	1.01x to 1.03x on the lab's measured f 0.85 (the session's own twin reads f 0.98 on this host: re-price pending)	1.02x to 1.03x (the session's f 0.813)	335 (N2 full-SRAM die, the coexistence row); 23.6x	none on the market	docs/analysis/class-v6/1p5x/lab/adv/v3/ledger-rows.md and ledger-rows.json (class-v6-adversary 47f16b86f); the lab artefact tree (not in the repository): lab/adv/rows/ctl-v6.json and rows.md (set v3, 2026-10-09T11:33Z); the die 12.0x at the 5090 knee, USD 0.77 per MH/s
Ravencoin	1,757 (1,453 to 2,147); GDDR7 board + register file sized to the work + fused operations, N5; bandwidth-bound at 16 KB per hash (93 MH/s per board); split memory 647 / ALU and tax per rows.md; WITNESS (modelled on the kawpowminer census as the work proxy until CONTROLS' trace replaces it)	3.08x (2.52 to 3.72)	3.25x (2.66 to 3.93)	pending re-price on the measured f 0.748 (ADVERSARY; the premise's own figure 1.10x to 1.20x until then)	pending re-price on the measured f 0.715 (ADVERSARY; 1.10x to 1.20x until then)	697 (N2 full-SRAM die, the coexistence row (USD 1.89 per MH/s); the cheapest specialist of any kind is near-memory base dies at 669 nJ); 8.2x	none on the market	docs/analysis/class-v6/1p5x/lab/adv/v3/ledger-rows.md and ledger-rows.json (class-v6-adversary 47f16b86f); the lab artefact tree (not in the repository): lab/adv/rows/ctl-progpow-094.json (set v3)
Firo (= Ravencoin's row)	1,757 (1,453 to 2,147); GDDR7 board + register file sized to the work + fused operations, N5; bandwidth-bound at 16 KB per hash (93 MH/s per board); split memory 647 / ALU and tax per rows.md; WITNESS (modelled on the kawpowminer census as the work proxy until CONTROLS' trace replaces it)	3.08x (2.52 to 3.72)	3.25x (2.66 to 3.93)	pending re-price on the measured f 0.748 (ADVERSARY; the premise's own figure 1.10x to 1.20x until then)	pending re-price on the measured f 0.715 (ADVERSARY; 1.10x to 1.20x until then)	697 (N2 full-SRAM die, the coexistence row (USD 1.89 per MH/s); the cheapest specialist of any kind is near-memory base dies at 669 nJ); 8.2x	none on the market	= Ravencoin's row
Iron Fish	1,051 (881 to 1,257); GDDR7 board + register file sized to the work + fused operations, N5; bandwidth-bound at 124 MH/s per board; 95 percent of the board's energy is the memory path; split memory 529 / ALU 46 / tax 503; WITNESS (modelled; the card's cell pending, so the ratio is pending)	12.05x (10.07 to 14.37)	7.60x (6.36 to 9.07)	pending (the twin reads above the full kernel on both cards: f a bound, not a share)	pending (as the 4090)	170 (full SRAM dies (three N2 reticles) + sized + fused operations); 74.5x	none on the market	the lab artefact tree (not in the repository): lab/adv/rows/ctl-fishhash.json and rows.md (set v3)
Ethereum Classic	765 (638 to 919); GDDR7 board + register file sized to the work + fused operations, N5; USD 3.25 per MH/s; split per ledger-rows.md; WITNESS (modelled; the card cells pending, so the ratios are pending)	9.62x (8.01 to 11.54)	7.27x (6.05 to 8.71)	pending re-price (the twin reads above the full kernel: f a bound)	pending re-price on the measured f 0.988 (ADVERSARY)	177 (N2 full-SRAM die, the coexistence row); 41.6x	Bitmain Antminer E9 Pro: 3,680 MH/s at 2,200 W = 598 nJ per hash, WITNESS (a manufacturer's published figure, not measured by the lab); Bitmain's stated figures as listed on minerstat's hardware page (https://minerstat.com/hardware/antminer-e9-pro): 3.68 GH/s on Ethash at 2,200 W, release 2023-02; 2,200 W / 3,680 MH/s = 0.598 J per MH = 598 nJ per hash	docs/analysis/class-v6/1p5x/lab/adv/v3/ledger-rows.md and ledger-rows.json (class-v6-adversary 47f16b86f, wording 82330b04c)
Ergo	198 (172 to 249); GDDR7 board + register file sized to the work + fused operations, N5; 529 MH/s per board; USD 1.15 per MH/s; split per ledger-rows.md; WITNESS (modelled from CONTROLS' trace: 33 random 32-byte reads per hash over the 7.27 GB table rebuilt every block, three Blake2b-256 compressions and the 256-bit sum per nonce, about 8,500 lane-instructions per hash including about 1,000 of amortised table rebuild, claimed); the card cell MODELLED-from-rate until COHORT's session row	pending	pending	1.01x (f 0.93 from the twin rate)	pending (the twin row)	46 (N2 full-SRAM on four reticles (38 to 58), USD 0.16 per MH/s; 25x at the 4090 cell); pending	none on the market	docs/analysis/class-v6/1p5x/lab/adv/v3/ctl-autolykos2.json and ctl-autolykos2-coexistence.json; v3/ledger-rows.md regenerated with the five chains (ADVERSARY, commit in its gate after d7db3e596; copies in scratchpad/lab/adv/)

The two phases with each chain's actual split

chain	split	phase A (before investment)	phase B (spend sunk)	an income a hash chip cannot earn	source
Igneum	80 / 20	before investment (24 months, the mining share 80 percent of the year-1 subsidy): the same-node board maker recovers its spend on NO fleet from 100 to 100 M boards on the growing or flat price path (100,000 boards: net USD -55.4 M on 58.8 M revenue growing; -37.6 M on 76.6 M flat); only the falling path pays it; the N2 die recovers on no path	spend sunk, flat path at USD 0.10 per kWh, the chip at 10 percent of the hash: mining revenue USD 0.354 per TH; the GPU's proving income at the ratified 20 percent share USD 0.098 per TH against the board chip's operating advantage USD 0.046 per TH (GPU electricity 0.111 less the chip's electricity, hosting and maintenance 0.066): the GPU out-earns the built board chip; the cancelling proving share 9 percent (board), 13 (hybrid), 20 (die)	yes by rule (the proving share of every block, 20 percent ratified); LIVE: paid on Devnet 4 to 30 prover keys (663 paid v1 segment records on chain 4465, the PROVING LIVE lane's file bd3f53c1 on build-1); the prover ships on by default in the Mac and Windows apps on 24 GB NVIDIA cards (release-2.0.3 provedefault.rs lines 23 to 24); HiveOS from 2.0.4 (today)	the lab artefact tree (not in the repository): lab/adv/rows/rows.md (the two-phase section), ADVERSARY commit 545f3f8a0; ctl-v6-coexistence.json; WITNESS; phase B NOT LIVE until a shipped package runs a prover
Ravencoin	100 / 0	before investment: no fleet from 100 to 100 M machines recovers a board chip's spend on any path, because Ravencoin's whole mining emission over 24 months (about USD 2.9 M at the public price) sits below one USD 35 M design cost; 10,000 boards would take 56 percent of the 721 GH/s network and earn USD 1.6 M to 2.4 M against a USD 44 M spend	spend sunk, mining only: mining revenue USD 0.064 per TH of the chain's own hashes; the board chip's recurring cost USD 0.069 per TH at ten cents; the chip's operating advantage over the RTX 4090 at stock (6,920 nJ) USD 0.123 per TH (the GPU's electricity 0.192 less the chip's 0.069) stands in full, no prover income	no (the block reward is mining only)	docs/analysis/class-v6/1p5x/lab/adv/v3/ctl-progpow-094-coexistence.json (rows[].phase_a, rows[].phase_b_split; method lab/adv/README.md section 14), ADVERSARY commit d7db3e596 on class-v6-adversary; WITNESS, the chain's own units, the 100/0 split, the same 24-month life, design cost and fleet rule as v6, the three price paths (flat at the public price, x2 growing, x0.5 falling, the hash held); the public lines on the row
Firo	100 / 0	as KAWPOW	as KAWPOW	no	as KAWPOW; = KAWPOW's row at a 100/0 split
Iron Fish	100 / 0	before investment: no fleet from 100 to 100 M machines recovers a board chip's spend on any path; Iron Fish's whole mining emission over 24 months is about USD 1.4 M at the public price (10,000 boards would take 84 percent of the hash for USD 1.1 M to 1.6 M)	spend sunk, mining only: mining revenue USD 0.086 per TH of the chain's own hashes; the board chip's recurring cost USD 0.043 per TH at ten cents; the GPU side NOT RUN until the card cell lands; no prover income by construction	no	docs/analysis/class-v6/1p5x/lab/adv/v3/ctl-fishhash-coexistence.json, ADVERSARY commit d7db3e596 on class-v6-adversary; WITNESS, the chain's own units, the 100/0 split, the same life, design cost, fleet rule and price paths as v6; the GPU side of phase B NOT RUN until COHORT's card rows
Ethereum Classic	100 / 0	before investment: no fleet from 100 to 100 M machines recovers a board chip's spend on any path; ETC's whole mining emission over 24 months is about USD 83 M at the public price, but 10,000 boards are 2 percent of 120.9 TH/s and earn USD 1.0 M to 1.5 M, and 100,000 boards (13 percent) USD 9 M to 13 M against a USD 150 M spend	spend sunk, mining only: mining revenue USD 0.0087 per TH of the chain's own hashes, below even the modelled board's recurring cost (USD 0.031 per TH) and below the E9 Pro's own electricity at ten cents (598 nJ = USD 0.017 per TH): at the public figures nobody mines ETC at a profit at ten cents and phase B is a loss on both sides; the GPU side NOT RUN until the card cell lands	no	docs/analysis/class-v6/1p5x/lab/adv/v3/ctl-etchash-coexistence.json, ADVERSARY commit d7db3e596 on class-v6-adversary; WITNESS, the chain's own units, the 100/0 split, the same life, design cost, fleet rule and price paths as v6; the GPU side of phase B NOT RUN until COHORT's card rows
Ergo	100 / 0	before investment: no fleet recovers a board chip's spend on any path; 10,000 boards would take 90 percent of Ergo's hash for USD 0.4 M to 0.6 M of revenue over 24 months against a USD 46 M spend; the chain's whole mining emission over 24 months about USD 0.6 M at the public price	spend sunk, mining only: mining revenue USD 0.013 per TH of Ergo's hashes; the board chip's operating advantage over the RTX 4090 (the modelled-from-rate cell) USD 0.024 per TH (the GPU's electricity 0.032 less the chip's recurring 0.008), standing in full; no prover income	no	docs/analysis/class-v6/1p5x/lab/adv/v3/ctl-autolykos2-coexistence.json (ADVERSARY 13:1x UK); WITNESS, the chain's own units, the 100/0 split, the same life, design cost, fleet rule and price paths as v6

The ranking

(a) Ours, the same-node DRAM board at the worse clean card (the smaller chip edge first; a throttled host's cell is named and left out of the order): Every compared cell is from one session (ledger-20261009 (COHORT: one session per card, every chain back to back; the 4090 pod Vast 55017318, the 5090 pod Vast 55018470)); the order at the worse card, with each row's own label:

Ravencoin: 3.25x at the worse card (PROVISIONAL: a proxy, pending or repeat-owed cell in this row); RTX 4090 3.08x (2.52 to 3.72) (THROTTLED HOST, out of the order; repeat in flight) (proxy); RTX 5090 3.25x (2.66 to 3.93) (proxy) (spread 36.84 percent, repeat owed)
Igneum: 4.37x at the worse card (PROVISIONAL: a proxy, pending or repeat-owed cell in this row); RTX 4090 5.14x (4.08 to 5.96) (THROTTLED HOST, out of the order; repeat in flight); RTX 5090 4.37x (3.47 to 5.07)
Ethereum Classic: 7.27x at the worse card (PROVISIONAL: a proxy, pending or repeat-owed cell in this row); RTX 4090 9.62x (8.01 to 11.54) (THROTTLED HOST, out of the order; repeat in flight) (spread 27.44 percent, repeat owed); RTX 5090 7.27x (6.05 to 8.71)
Iron Fish: 7.60x at the worse card (PROVISIONAL: a proxy, pending or repeat-owed cell in this row); RTX 4090 12.05x (10.07 to 14.37) (THROTTLED HOST, out of the order; repeat in flight); RTX 5090 7.60x (6.36 to 9.07)

Ergo: unranked in this column (no measured card cell yet)

(b) ProgPoW's premise (the chip keeps a GPU-class datapath):

Igneum: RTX 4090 1.01x to 1.03x on the lab's measured f 0.85 (the session's own twin reads f 0.98 on this host: re-price pending); RTX 5090 1.02x to 1.03x (the session's f 0.813)
Ravencoin: RTX 4090 pending re-price on the measured f 0.748 (ADVERSARY; the premise's own figure 1.10x to 1.20x until then); RTX 5090 pending re-price on the measured f 0.715 (ADVERSARY; 1.10x to 1.20x until then)
Iron Fish: RTX 4090 pending (the twin reads above the full kernel on both cards: f a bound, not a share); RTX 5090 pending (as the 4090)
Ethereum Classic: RTX 4090 pending re-price (the twin reads above the full kernel: f a bound); RTX 5090 pending re-price on the measured f 0.988 (ADVERSARY)
Ergo: RTX 4090 1.01x (f 0.93 from the twin rate); RTX 5090 pending (the twin row)

(c) Phase B, an income a hash chip cannot earn:

Igneum: yes by rule (the proving share of every block, 20 percent ratified); LIVE: paid on Devnet 4 to 30 prover keys (663 paid v1 segment records on chain 4465, the PROVING LIVE lane's file bd3f53c1 on build-1); the prover ships on by default in the Mac and Windows apps on 24 GB NVIDIA cards (release-2.0.3 provedefault.rs lines 23 to 24); HiveOS from 2.0.4 (today)
Ravencoin: no (the block reward is mining only)
Firo: no
Iron Fish: no
Ethereum Classic: no
Ergo: no

The claim under test, clause by clause

The sentence below is served as a claim under test, each clause labelled by its state on the rows above. It reads as established only when every clause is supported; today it is not, and the table says which clause and why.

The headline, built from the supported clauses only: under ProgPoW's own convention, the lowest measured chip-to-GPU edge of the chains on this ledger; and the only chain on this ledger whose GPUs are paid, by rule, an income a hash chip cannot earn (LIVE: paid on Devnet 4; the prover ships on by default in the Mac and Windows apps on 24 GB NVIDIA cards; HiveOS from 2.0.4 (today). The one-epoch evidence (the PROVING LIVE lane, 14:1x UK 9 October 2026): on chain id 4465 (the Igneum 2.0 devnet, Devnet 4), chain blocks 32,000 to 33,534 (DAA 63,577 to 67,207, 12:02 to 13:00Z), 928 shard records and 86 aggregation records paid 1,453.53 IGN to 15 distinct prover keys (14 RTX 4090s and one L40S), 51 percent of the 2,850.41 IGN the pool was credited at the ratified 20 percent share, 60 percent of chain blocks carrying a paid shard proof; read by igneum_getProvingStatus, igneum_getSegmentRecords and igneum_getSegment; files the lab artefact tree (not in the repository): proving-live/prov-ledger-4465-epoch-125954Z.json sha256 2b6572273eab7c58ad4f08bd04ef78172b1806834d5493734ff8522d92d8846e and shards-32000-33534.json (sha256 f05144813b0ccc32444b..., the full sum beside the file); the record docs/analysis/class-v6/1p5x/lab/proving-live.md (commit 8c394428 on proving-live). Since genesis: 6,474 v0 shard records and 663 v1 aggregation records to 30 prover keys, 7,850 IGN (the lane's corrected baseline; its first line's 653.93 IGN were the aggregation tenths alone). The shipped Mac and Windows apps carry the prover ON by default on any NVIDIA card of 23,552 MiB or more (release-2.0.3, app/igneum-app/src/provedefault.rs lines 23 to 24); the HiveOS prover ships in package 2.0.4 (its canary about 14:30 UK)).

The sentence as ordered, served as the claim under test: "the lowest measured chip-to-GPU edge of any live GPU network under a published method, on both conventions, and the only one whose GPUs earn income a chip cannot"

Rule: the headline reads only what the like-for-like rows support, clause by clause (main's ruling, 13:2x UK 9 October 2026); under the ledger's own convention the measured edge per chain on the same card, state and session is stated as numbers with no ranking word for Igneum, because the rows do not support one.

clause	convention	state on the current rows	why
under ProgPoW's own convention, the lowest measured chip-to-GPU edge of the chains on this ledger	ProgPoW's premise (a chip that keeps a GPU-class datapath)	SUPPORTED on the premise figures as published; the re-price on each cell's measured f is pending	Igneum v6 1.01x to 1.03x on its measured f (0.81 to 0.85 on the lab's cells; the session's 4090 twin reads 0.98) against the README's 1.10x to 1.20x for KAWPOW, FiroPoW and Etchash; the session's twins now give KAWPOW f 0.75 and 0.71, which moves its premise ratio toward 1 when ADVERSARY re-prices it; FishHash's twin reads above its full kernel, so its f is a bound and its premise ratio pending
under the ledger's convention (the cheapest credible programmable specialist, adversarial set v3), each chain's measured edge on the same card, state and session, as numbers	ours	MEASURED in one session per card; NOT the lowest: Igneum reads second of the four chains with a cell	at the RTX 4090 at stock: Ravencoin's proxy row 3.08x, Igneum 5.14x, Ethereum Classic 9.62x (the real Antminer E9 Pro from its public figures 12.3x), Iron Fish 12.05x; at the RTX 5090 at stock: Ravencoin 3.25x (a 36.8 percent spread, repeat owed), Igneum 4.37x, Ethereum Classic 7.27x (E9 Pro 9.3x), Iron Fish 7.60x; Ergo faulted in the session (no cell). The mechanism is on the page: a coalesced 16 KB stream per hash gives the chip a smaller memory-term gap than 256 dependent 32-byte random reads
and the only chain on this ledger whose GPUs are paid, by rule, an income a hash chip cannot earn	phase B, each chain's actual reward split	LABELLED BY THE PROVING LIVE LANE	the proving share (20 percent ratified) is paid to GPUs for work a hash chip cannot do; every other chain on the ledger pays mining only; the live state of that income is the PROVING LIVE lane's evidence line

Withdrawn: the earlier line 'KAWPOW and FiroPoW read 5 to 6x on the measured 4090 rows and v6 reads 2.35x' is withdrawn: it compared v6 at the 5090's 1,300 MHz knee with KAWPOW at the 4090 at stock, two cards and two states; the like-for-like cells are the ones on this ledger.

The last clause's label from the PROVING LIVE lane: LIVE: paid on Devnet 4; the prover ships on by default in the Mac and Windows apps on 24 GB NVIDIA cards; HiveOS from 2.0.4 (today). The one-epoch evidence (the PROVING LIVE lane, 14:1x UK 9 October 2026): on chain id 4465 (the Igneum 2.0 devnet, Devnet 4), chain blocks 32,000 to 33,534 (DAA 63,577 to 67,207, 12:02 to 13:00Z), 928 shard records and 86 aggregation records paid 1,453.53 IGN to 15 distinct prover keys (14 RTX 4090s and one L40S), 51 percent of the 2,850.41 IGN the pool was credited at the ratified 20 percent share, 60 percent of chain blocks carrying a paid shard proof; read by igneum_getProvingStatus, igneum_getSegmentRecords and igneum_getSegment; files the lab artefact tree (not in the repository): proving-live/prov-ledger-4465-epoch-125954Z.json sha256 2b6572273eab7c58ad4f08bd04ef78172b1806834d5493734ff8522d92d8846e and shards-32000-33534.json (sha256 f05144813b0ccc32444b..., the full sum beside the file); the record docs/analysis/class-v6/1p5x/lab/proving-live.md (commit 8c394428 on proving-live). Since genesis: 6,474 v0 shard records and 663 v1 aggregation records to 30 prover keys, 7,850 IGN (the lane's corrected baseline; its first line's 653.93 IGN were the aggregation tenths alone). The shipped Mac and Windows apps carry the prover ON by default on any NVIDIA card of 23,552 MiB or more (release-2.0.3, app/igneum-app/src/provedefault.rs lines 23 to 24); the HiveOS prover ships in package 2.0.4 (its canary about 14:30 UK)

Reproduce it tonight

The reproduction kit holds the pinned references, the harnesses, the sampler, the session script, SHA256SUMS and the expected rows with their bands. One command per card, from the kit's README: a 4090 or 5090 at stock on a host that does not cap the card's power runs every chain back to back in about seventy minutes after the datasets build and reads its rows against the bands.

item	state
the kit	build.igneum.network/fleet/ledger/igneum-gpu-ledger-kit-20261009-2.tgz (10.7 MB; stamp 20261009-2; published 15:07 UK 9 October 2026, read back at the edge by sha256)
the kit's sha256	4451f249b97b9f75efa0b5006860468a2bd2bb382e2c91eac7b3d7fdfde0a95a (SHA256SUMS over every file inside)
the store on the build host	the lab artefact tree (not in the repository): ledger/20261009-2/ (controls/ the five pinned chains with sources, CPU references, harnesses and per-dir shas; harness/ the COHORT lane's lab-kit-ledger-v2.tgz and KIT-README.md; record/ ledger.json, README.md, ROWS.md; README.md; SHA256SUMS over 75 files)
the identity guard	tools/ci/served-identity-guard.sh over every text file of the kit: 48 files clean under 53 patterns (build-2, 15:0x UK)
the expected rows	record/ROWS.md inside the kit: the session's median nJ per hash per chain and card with a 5 percent band either side; a row outside its band reads its event_reasons column first (a power cap or a thermal event puts the run outside the method); stamp 20261009-2's ROWS.md carries the 4090 rows of the throttled host, named so; stamp 20261009-3 follows the repeat sessions