igneum/docs/analysis/horizon-2026-10.md

Horizon, October 2026: the ranked research across every system

Written 6 to 7 October 2026 by the Horizon coordinator (branch horizon, worktree igneum-wt-horizon) on the project lead's ask of 6 October 2026, 20:4x UK: "deep backward and forward predictive research and modelling for our algo and all of our systems: is there room for improvement, room for more coin utility, algo improvements, security improvements, anything we can do to make a 51% attack impossible, basically creating a level of polish that has not been seen before." Later the same evening: "research anything else that we can research too, predictions, forward thinking, what we can actually do that has not been done or applied, think outside the box", "be revolutionary", and "if we create a new way of hashing or a new way of proof of work to revolutionise the space then that's absolutely fine, I want you to deploy everything to create something that has not ever been done before."

The bar main set, and the bar this document holds every claim to: "impossible" is not available to any proof-of-work chain. The bar is that a majority of hash buys nothing: it cannot reverse what finality locked, cannot forge a proof the nodes re-execute, cannot change a rule without 95 percent signalling, and loses more than it earns. Every claim here is a number with a model or a simulation behind it, priced in rented hash at the measured 6 October rate (docs/bench-log.md, "Rental cost of hash, 6 October 2026": USD 0.0117 per MH/s-hour, so USD 11.7 per GH/s-hour and about USD 281 per GH/s-day; the live devnet at 1.16 GH/s), with its consequence per tier and what to build. Hours are agent hours (the project lead's rule: Claude-side work takes hours, never weeks).

The lanes

Lane	File	State
1 consensus-security	docs/analysis/horizon/consensus-security.md, and the paper docs/analysis/51-percent.md; models sim/horizon/consensus-security/	landed, commit eec2cd7
2 algorithm	docs/analysis/horizon/algorithm.md; model sim/horizon/algorithm/model.py	landed, commit 5ff7393
3 finality-and-weight	docs/analysis/horizon/finality-and-weight.md; models sim/horizon/finality-and-weight/	landed, commit c3aa502
4 economy-and-utility	docs/analysis/horizon/economy-and-utility.md; models sim/horizon/economy-and-utility/	landed, commit 4617a01
5 network	docs/analysis/horizon/network.md; the experiment docs/analysis/block-rate-devnet2.md	landed, commits 3777014 and b965b64 (runs A and B; A2 dropped)
6 polish	docs/analysis/horizon/polish.md	landed, commit ac4cc93 (95 ledger rows)
7 frontier	docs/analysis/horizon/frontier.md; model sim/horizon/frontier/frontier_model.py	landed, commit 5ff7393
8 new-proof-of-work	docs/analysis/horizon/new-pow.md; prototypes proto-newpow/	landed, commits cbcff47 and 92db7c5 (two prototypes measured on rented 4090s, verdicts in section 6)

1. One page for the project lead

Closed 6 October 2026, 22:3x UK, every lane landed.

Standing decisions (the project lead, 6 October 2026, 22:3x UK). Verbatim: "Fees cannot fund security for a decade" and "Miners need to be the security". Meaning: miners are the security always, no time bound, no stake, no outside checkpoints or committee. The chain pays its own miners from emission plus fees; nobody pays upkeep, not the founder, not a treasury, not a dev fund. Self-sustaining means the emission curve keeps mining worth doing on its own for as long as fees are small, which lane 4 measures as a decade or more, so emission never decays on a schedule that assumes fees take over. Fee revenue is never assumed as the security budget in any model or public sentence. His third line, "Wrong constants and claims in our own text", acknowledges lane 6's finding; the nine ledger rows in item 9 are the fix.

1. The proving pool was the one line where a majority earned more than it spent: 11,636 IGN an hour at 51 percent of blocks. Fix: the proof verified in consensus. In 0.3.16 (fork exec-sync-0313 da2d17ec), switch proving_consensus_verify_daa off by default. Decision owed: when it activates.

2. Tonight's two-hour finality pause. 42.7 percent of the frozen voter table left in three minutes and the frozen table held the pause a window (30 days on mainnet). Fix: a signed leave item, the key out of every denominator an hour later. In 0.3.16 (finality-pause-node 766e70ca, finality-pause 27f82ec). The slices-under-10-percent rule: done.

3. A 51 percent attacker buys a 90-second reorder window and nothing past a certificate. Done: the peer-driven unwrap class (8e2f5cbe, 0.3.16), the receive-side version gate (f1ea7a38, 0.3.15). Next: weight-gated deep fork choice and vote-or-burn.

4. The chip is settled in kind, open in degree: 5.7x per joule at class v3, 2.1x at v4. The lever is the latency-shadow size N as a genesis ladder, each step by 90 percent signal. Text (M32, M33): done. Decision owed: the N ladder at genesis.

5. Fees stay tiny for about ten years; emission carries security, with no end date. USD 450 a day at launch and 4,200 in year 5, against 54,800 and 13,700 of emission. Text (E19, E20, E21, P24, P25): done. Next: unproven credit rolled forward, the job price decoupled from f_p.

6. The signalling window could be bought for a day. Seven consecutive daily windows at 95 percent, the floor a week past publish: in 0.3.16 (ca3-v4-0316 0760b844). The thresholds in one sentence (G15): done.

7. New proof of work, measured. Scheme A (mining is proving): never. Scheme B (tensor shadow): never as class content; kept as reserve R8. Scheme C (dataset from stored state): the class v5 candidate; hash rate and watts unchanged, build +1.4 ms, verifier +0.11 to 0.21 ms per unit.

8. Block rate: 1 a second for the testnet and launch. 10 a second on Devnet 2 ran 77.6 percent red on node cost. 10 waits behind three gates: per-block CPU under 50 ms on a laptop core, finality constants in DAA seconds, vote aggregation. Done.

9. Polish. Pause wording in Ember and the API: in 0.3.16 (ember-tune b726ce4). fork_is_close and the publisher's activation guard: in 0.3.16 (1357d28, 08f5276). Export-disk cap: done (gpu-fleet f9ad70e). Nine text corrections (M32, M33, F26, E19, E20, G15, P24, E21, P25) plus X31 to X33: done. Unsigned installers: decision owed. Relay fixes (28c028b) on fud-close: merge owed.

Decisions owed from the project lead: the cryptanalysis spend (USD 80,000 to 160,000); the testnet date word; the N ladder at genesis; the verification switch activation.

2. The ranked list: top 25 across every lane

Rank is payoff over cost across lanes, with safety first, then liveness, then money, then text. "L1 r3" means lane 1's own rank 3; the lane file holds the full evidence row.

Rank	Item	Lane	Evidence	Model	Hours	Consequence per tier	What to build	Gate
1	Verify the aggregated segment proof in consensus; a record whose proof fails against the pinned aggregator key is invalid	L1 r1	spec 07 7.7 items 3 and 4; P21; a producer captures its block share of the 20 percent pool with fake records	capture = pool x (H inside the window + most outside): 11,636 IGN/h at 51 percent; SP1 light-verifier cost per record 1.8 to 2.4 s measured (bench-log)	8 to 12	prover: paid only for proofs; holder: the pool is real; node operator: one verify per record on the proof pool thread; miners: nothing	the verify call in the record check of consensus/core proving, the v0 per-shard record payout-only and capped at the exclusive window	fast-time: a fake-proof record is rejected by every node and the producer loses the block; a true record pays
2	The departure announcement: a leave item (key, DAA score, signature) in blocks; D = 1 h later the key is in no denominator, sliding or frozen, and its votes are invalid; app Stop and the fleet library send it	L3 r1, L1 r8	tonight's pause (L3 3.1): 42.7 percent left in 3 min, the frozen table held a window; sim T: first lock 1 h after a 34, 45 or 50 percent departure, 0 conflicts in every partition, eclipse and equivocator row	finality_horizon.py seeds 7, 11, 13	6 (+4 for F5's trusted certificate)	holder: a 30-day mainnet pause becomes 1 h when leavers are honest; a silent leaver still costs a window; pool and rig: one message on a clean stop; attacker: buying keys to leave them gains nothing (w + L must still reach 2/3)	the item in the coinbase finality section, the denominator rule in finality.rs, the send in Ember and tools/fleet/lib	harness: 45 percent leaves with leaves, lock within D + 1 checkpoint; 0 conflicts in 50/50, 60/40 and the 34 percent eclipse
3	Signalling over 7 consecutive daily windows at 95 percent, the floor no nearer than 7 days past the publish, the stale-box list empty before the floor	L1 r6, L4 4.5	the P2 one-day window is buyable: 19 N of hash for 24 h (USD 534k at 100 GH/s) forces a flip onto an unready fleet; a 6 percent holdout buys delay for nothing	signalling.py, signal_game.py	3	pool and rig: a week more before a class change and a week of visible share; home miner: a week to update; attacker: the bill x7 in public	the window count in the P2 rule, spec 5.7 text, one sentence carrying 60 / 90 / 95 in spec, CLAUDE.md and the litepaper	fast-time: 6 of 7 days at 95 percent does not flip; 7 does; the harness's failed case
4	Weight-gated deep fork choice: a tip whose fork point is older than D (10 min of past-median time) is a candidate only if the blocks on it since the fork were produced by keys holding at least 1/3 of the weight table at the fork point	L1 r2	during a pause or the first 20 days a renter with fresh keys double-spends at the 12-h depth for USD 146 at 1 GH/s	cost_model.py; deterministic over the block's past	10 to 16	holder and exchange: rented hash cannot reorg past 10 min even while nothing locks; honest miners: nothing (their keys hold the weight); new chain: the first 20 days gain a bound they lack today	the candidate filter in GHOSTDAG tip selection over the certified weight table	fast-time: a 51 percent fresh-key fork from 15 min back is refused by every honest node; a 1/3-weight fork is accepted; partition heal unchanged
5	Vote-or-burn: a block whose producer key has participation under 0.5 over the presence window in its own past burns 20 percent of its producer share	L1 r3	the pause as a liveness attack has zero marginal cost once the veto is held	0.2 x A x 0.8 x subsidy: 7,757 IGN/h at 34 percent (USD 155/h at 0.02)	6 to 8	attacker: a pause costs; honest home miner: nothing while signing (the official client signs every checkpoint); pool user: the pool's participation; partition-safe because the test is the block's own past	the participation read in coinbase validation, the burn in the subsidy split	fast-time: a 34 percent silent set's blocks pay 80 percent; a 50/50 partition burns nothing on either side
6	Roll unproven shard credit into the next proven segment's pool instead of the escrow for ever	L4 r2	a ten-day prover refusal strands 547,570 IGN a day (5.5 M IGN) with no rule; spec 5.3, 7.7 item 3, 7.8 item 7 are silent	stress.py refuse scenario, stranded_share	8	prover: credit is never lost to the market's slow days; holder: no undecided burn; rollup customer: nothing	the roll-forward in the pool credit split, spec 5.3 text	fast-time: 100 unproven then 10 proven segments return the escrow to zero
7	Decouple the job price from f_p: reserve = measured proving electricity per pgas at the published settlement rate (spec 5.10.3), the 1.5 premium becomes a bid	L4 r1	at the adopted floor a billion-cycle job is 15 IGN = USD 0.075 / 0.30 / 1.50 against Boundless's 0.21 (approximate); above USD 0.014 per IGN the chain overprices by rule	utility.py sections 1 to 2	16	rollup customer: a price that can clear; prover: still paid above electricity; holder: more jobs, more burn	the reserve rule in spec 5.4 and 5.11, the bid field on the job	a simulated job book clears within 20 percent of Boundless's median at all three prices
8	Operational rule, no protocol change: a standing box never leaves the live chain for an experiment; any orchestrated departure over 10 percent of weight is staged in slices under 10 percent an hour; the fleet library refuses to swap a standing box's chain	L3 r2	the rehearsal took 36.5 percent at once; sim L2 and M4: a gradual departure costs nothing because every lock re-freezes the table	spec 3.7 item 2	1	fleet operator: finality stays on; every tier: no pause from our own experiments	a refusal in tools/fleet/lib plus a --staged departure	the next rehearsal keeps finality_active true
9	Peer floor and mesh, and no unwrap on a peer-driven path with a sync-request fuzz gate (tonight's pruned-node crash: consensus/src/processes/sync/mod.rs:87, plus 27 sibling sites listed in L1)	L1 r4	run A's star (321 tips, 77 percent red); the hub crash at 19:57Z: any peer can crash any pruned node at zero hash cost	a star with its hub down is n islands; the sibling list by file:line	9 to 11	node operator: no crash from a peer's request; home miner and rig: the app alarms under 3 outbound peers or 2 checkpoints without a vote; fleet: no star	the fleet lib dials 3 boxes beside the hands; the node alarm and app line; the unwrap sweep; the fuzz in tools/ci	the fuzz runs the sync request space against a pruned node with no panic; the alarm fires on a known-cut case and stays quiet on a healthy one
10	N (the latency-shadow size) as a genesis ladder per era {100k, 130k, 200k, 330k, 650k, 1.0M}, floor and ceiling fixed at genesis (10x verifier headroom on the 2.5x rule; O-1.14 sets the ceiling), one era draw consumed as for epoch_len, each step by 90 percent miner signal over 7 days, never unconditional	L2 r5, L7 finding 1	HBM4 raises the stored-dataset chip's bare edge (4.9x to 12x depending on tFAW, L2 and L7 disagree and both are labelled); N = 100k holds the chip at 2.1x on GDDR7 at k = 1, N = 330k at 1.3x; an unconditional doubling retires the M5 Max at era 1 (-10.5 percent at 200k)	model.py --section ladder (L2 5.3a, the reconciled table)	6 to 8	100k to 130k: the M5 Max -3.3 points, nobody else; 130k to 200k: M5 Max -6 more, the 5090 -2.7 at its cap, the 4070 +21 W; 200k to 330k: compute-bound on every capped NVIDIA card; pool users nothing at any step; verifier +0.17 to 0.56 ms per warp	the ladder in the era-draw table of spec 1.13, the step rule on P2's mechanism	per step: every public-benchmark card within 5 percent of its previous-step rate, bit-exact on three vendors, verifier under 10 ms cold on the O-1.14 core
11	Aggregate-first vote verification and aggregated in-block carriage as the mainnet default (spec 3.4.2 item 2, decided)	L3 r3	per-vote verification is 1.6 s per checkpoint per node at 1,000 voters and about 13 s at 8,192 (approximate); measured lock delay p50 0.86 to 1.26 s at 4 to 93 voters on node 1	L3 4.3 table; blst 0.3.17 fast_aggregate_verify already in the fork	8	node operator on a laptop: a checkpoint costs one pairing, not a thousand; every tier: lock delay stays under 3 s at 8,192 voters	batch votes per (index, hash), verify once, bisect on failure; the in-block aggregate	fast-time with 1,000 and 8,000 synthetic voters: p50 lock delay under 3 s, CPU under 25 percent of a core, 0 conflicts
12	The finality weight table carried inside the recursive segment proof, updated one mergeset per segment: a consensus proof at mergeset cost	L7 r1	phase two's hardest item (P4) becomes incremental on code that exists; about one BLS verify per 30 s (one shard's budget, approximate, unmeasured)	frontier_model.py	60 (first 10: measure)	phone wallet and bridge: finality from one proof, no node trusted for the voter set; rollup customer: one object; prover: one more public-value block per segment	the W2 transition in the aggregator guest, the GHOSTDAG colouring check in-guest	first: BLS verify and colouring cycle counts inside the SP1 guest on a 24 GB card within 2x of the estimate; then a proof per checkpoint under 30 s on a proving-only 5090
13	Reproducible-build attestations in a registry contract; Ember refuses a release under N of M attestations	L7 r3	ledger G7, G13: the update key is an operator channel; Bitcoin's guix.sigs is the precedent with the chain as the sigs repo	text and contract	12	every tier: an update needs N independent builders, not one key; fleet: the box's reproducible Windows and Linux builds (CLAUDE.md 6 Oct) are the inputs	the contract, the builder tool, the Ember check	a release with 1 of 3 attestations is refused by Ember; 3 of 3 installs; the known-failed case
14	Close O-1.14 with a real 2019 laptop run; adopt the half-core proxy as the standing stand-in; dr736 out, dr368 in as R0	L2 r1	box proxy: dr736 10.51 ms cold and 15.49 half-core; class v4 5.06 and 8.23	L2 5.5	2	miners 0; pool verifier cores x1.3 at dr368 instead of x2.4; a 2019 node keeps 1.8 ms under the gate on class v4	Windows cross build on the box, relay to the laptop, bench, log	ms per warp under 10 cold on the real core for class v4 and dr368
15	The vote signs the execution root too (chain id, index, block hash, post_root); a certificate pins the state; snapshots check against the last certificate	L1 r5	snapshot poisoning: a wrong state above the pin is undetectable today	every voter executes natively (spec 07)	8 to 12	holder: a lock is a state lock; node joining from a snapshot: cannot be poisoned; cost: exec lag enters lock latency	the vote message, the pin rule, the snapshot check	fast-time: a poisoned snapshot node cannot join the quorum; honest lock delay rises by the measured exec lag only
16	finality_provisional reported beside finality_active, never as a lock, plus a detector-driven security_alert (correlated group over 1/3 of a day's blue blocks; pause; conflict) shown by wallets and the explorer as "confirm at 12 h"	L3 r4, L1 r7	sim P: provisional conflicts in every partition (516 to 1,062 per 360 min), final 0; the exchange guidance exists only as text today	L3 4.1; the detector of counter-asic-3	3 + 4 to 6	exchanges: a third row in the guidance; holder: told when to wait; home miner: a line in Ember	the RPC fields, the explorer and wallet copy, spec 3.9 row	forced pause: explorer and wallet show provisional and the alert; a healthy day shows neither
17	Every miner-signalled execution parameter enters the consensus digest the same release, with a tools/ci check that fails a Params field marked signalled and absent from the digest	L4 r5	signal-then-defect on an execution parameter is a silent state fork unless the handshake refuses the defector; Params.fees is in the digest, nothing guarantees the next	signal_game.py section 3	6	node operator: no quiet fork; every tier: a defector is refused at the handshake	the check, the digest rule in spec 5.8	the check fails on a planted field and passes on the live set
18	The 80/20 split as a 60 percent parameter inside a hard band [10, 30] percent	L4 r4	not load-bearing at launch traffic; 30 percent buys backlog relief at 100 shards a block (economy-2026-10-04 5.3)	stress.py	10	prover and miner: a split the market can move inside a band nobody can break; holder: the cap and the halving untouched	the band in spec 5.3 and 5.7	harness: a 60 percent vote reaches 30 percent; a 100 percent vote cannot pass the band
19	A WebAssembly verifier of the wrapped block proof in the tab, with the millisecond count shown	L7 r4	three working precedents (Helios WASM, ziren-wasm-verifier, wasm-groth16-verifier); the certificate half already runs at 58 to 68 ms warm, 139 to 155 ms cold (bench-log round 6, P3)	text	16	holder and rollup customer: verify in the browser; the P3 wrapper measurement comes forward	the Groth16 or Plonk wrapper, the WASM verifier on the site	a block proof verifies in the tab under 500 ms on a laptop; the measurement labelled on the page
20	Ember as node, wallet and light client for everyone	L7 r5	1,000 testnet miners become 1,000 verifying nodes; Monero shows about 5,000 peers (monero.fail), Ethereum 8,136 execution nodes (ethernodes), approximate	text	20	home miner: one app; node count = miner count; Sybil counts of nodes irrelevant by design; node cost per lane 5's tiers	the node and wallet surfaces inside Ember	a fresh Windows install mines, verifies and shows a balance in the first 60 s with no second download
21	Work-stake: vote weight as the external-job bond, with the spec sentence "no coin stake; the only thing at stake is 30 days of public work" written first	L7 r2	a 0.1 percent key has about 16,427 IGN of 30-day pool income plus its vote at risk against a designed 0.0015 IGN coin bond per job	frontier_model.py section 3	24 (prototype)	prover: a bond nobody can buy; rollup customer: a griefing cost that scales with the prover's standing; holder: "no stake" stays true as "no coin stake"	the bond rule in the job claim	a griefed job costs the griefer its weight for 30 days in the fast-time harness; the spec sentence lands before the prototype
22	Client-shipped certified checkpoint (index, hash, voter-table digest) refused if missing from the DAG; exec generations spaced geometrically to the finality depth (about 16, 1.8 GB today)	L1 r9, r10	long-range and seed attacks; a pause-time deep reorg needs a peer's snapshot today	assumevalid's shape; 114.8 MB per snapshot measured	3 + 3	every tier: a fresh install cannot be bootstrapped onto a private DAG; node operator: disk for generations, no blocked executor after a deep reorg	the checkpoint in the release, the generation schedule	a cold node refuses a DAG missing the checkpoint; a 5,000-block reorg re-executes from a generation with no snapshot request
23	Public-text corrections, one bundle: the prover's price as a formula with network hash as the input, never a number; the era draw and the reserve described as schedule changes against fixed datapaths, not unpredictability against a chip, with the chip's USD per MH/s-hour beside the honest cards'; F19's "day 19 to 20" is a v2 number (v3: a full window); funding.md section 4's dev-fee ceiling is 1 percent of the producer share (38,520 / 154,080 / 770,400); the dev fee is "default-on, switchable", not "optional"; "proofs at the cost of power" conditioned on hash; the three signalling thresholds in one sentence	L4 r3, r7; L2 r3; L3; L1	each row cites its lane section	text	1 to 3 each	holders and critics read claims that survive review	the litepaper, the customer brief, spec 5.7, funding.md, the ledger F19	the site's forbidden-strings check and a reviewer's read
24	Measure the FPGA lane on AWS F2 (one VU47P, HBM2, about USD 1.98 an hour) and replace the ceiling row	L2 r2	the measured 2.4 G reads/s equals the JEDEC tFAW ceiling; the old 1.9x row rests on a 12 ns tFAW the JEDEC HBM2 table does not give (28 ns); the soft overlay reads 0.30x to 0.47x of the 5090 per watt	L2 5.1	8 to 10 plus USD 2 to 8	none today; the public FPGA claim becomes a measured number	the overlay bench on F2	reads per second per watt at 1 GiB; the row replaced
25	Ember tune as the shipped default per card model, and the two fleet measurements every price rests on: the miner's hash loss while each tier proves, and a full 30 M-cycle shard beside the miner on 12 and 16 GB cards	L2 r7, L4 r8	the 4070 at 3.65 uJ untuned and 2.57 tuned (-30 percent); the hybrid row is the only one that undercuts the market and rests on one 5090 measurement; the 4.7 M fixture is 16 percent of a full shard (linear scaling says 240 s on a 3060, outside the 120-s claim timeout)	L2 5.1; utility.py hybrid_hash_loss	2 + 6 (fleet)	every NVIDIA tier gains 10 to 30 percent per joule; the 12 GB tier learns whether it can claim a full shard in time	the defaults table in the app from the fleet priors; eleven rows with both numbers in prover-tiers-real-cards.md	the rows land; the claim timeout is set from them

Lanes 5, 6 and 8 landed after this table was ranked; their rows are in the lane files (network.md section 6, polish.md section 1, new-pow.md section 7) and in the one page above. The table itself is not re-ranked: rank 1 and rank 2 are in 0.3.16, rank 3 is in 0.3.16, rank 8 is done, rank 9's unwrap half is in 0.3.16, rank 23's text bundle is done.

Not recommended, with the reason (as valuable as the list above)

Item	Lanes	Why not
Prover attestations as a second finality leg	L1 r12, L3 r8	provers are the miners (same vote keys), so no new party; proof coverage is 2.4 to 4.7 percent of blocks tonight with lag p99 62 s, so every lock would wait on proofs; revisit only after 99 percent of blocks are proven within 60 s for 7 days with 3 provers per block, and even then it adds about a minute of lock delay
Time-locked (vesting) weight	L1 r13	a bought key transfers vested weight, so the acquired-keys bound is unchanged; honest new cohorts wait longer
Any automatic re-lock after an abrupt departure	L1 r14, L3 r7	departure and partition are the same observation in one view; the decaying denominator produces 467 to 473 conflicting locks in a 360-minute 50/50 honest split, the hysteresis floor brings the 13.3 percent equivocator bound back (565 to 597 conflicts at 20 percent); the two-tier rule is a report, never a lock
Mining is proving (the lottery's work as a proving step)	L8 scheme A, L7 3.10	dead on bytes (2.9 MB of openings per block), on the verifier (32 to 40 ms against 10), and on sampleability (Ball et al. 2017, Ofelimos 2022); re-opens Aleo's fastest-prover-wins
Proving others' chains as the main income	L7 3.11	all of Ethereum L1's proving is about USD 36 a day at the Sep 2026 tracker cost against USD 13,700 a day of year-1 emission at USD 0.005; demand must grow 1,000x against a falling cost curve
Burn-redirect audit bounties, a review escrow by 60 percent signal	L7 3.6, L4 task 3	a dev fund with a veto, the switch spec 5.5 removed; the honest payer for a second audit is the entity's own provers
Proof verification on a hardware wallet's secure element	L7 3.9	a bn254 pairing on a Cortex-M-class element is seconds to minutes (approximate); do the companion verify
A general compute market for unverifiable work (rendering, inference)	L7 3.12	an escrow without a verifier is a trust-me payment with lower fees; the verifiable subsets are named and kept

3. What a 51 percent attacker can and cannot do

The paper is docs/analysis/51-percent.md (lane 1). In one table, at the measured USD 11.7 per GH/s-hour:

Hash share	What it buys	For how long	Cost at 1 GH/s / 1 TH/s of network	What it earns	Net
20 percent	reorders only the last k = 18 blocks; nothing past a certificate	per attempt	subsidy forgone while withholding	its block share minus reds	loss
34 percent	wins the selected-chain race only inside 60 s (40 percent of the time); holds the veto (blocks every lock) after 20 days of mining at 51 percent, or 30 days at 52 percent of the network's hash	20 to 30 days to acquire, then free to hold while silent	about USD 6k / 5.8 M to acquire (half earned back as subsidy); a pause then costs nothing	subsidy; nothing from the pause itself unless it double-spends at the 12-h depth during the pause (USD 146 to 146k)	the one line vote-or-burn (rank 5) and weight-gated fork choice (rank 4) close
45 to 51 percent	wins the selected-chain race over a 90-s hold 70 to 85 percent of the time (32 to 46 blocks at 1 bps); turns 26 to 28 percent of honest blocks red; lifts its weight share to about 56 percent, never 2/3	the 63 to 93 s between a checkpoint block and its lock	USD 11.7 / 11,700 per hour of rented hash; the market could not supply a TH/s on 6 October (RunPod: 0 pods for 20 asks)	a quarter of honest subsidy while withholding; the pool capture of rank 1 (11,636 IGN/h) until the in-consensus verifier lands	loss on the chain; gain only through the pool, which rank 1 closes
67 percent of weight	locks alone	2.03 x N of hash for 30 days: USD 17,100 per GH/s of network		subsidy	at rental equilibrium about 33 percent of 30 days of subsidy net
any share	a rule change	95 percent of blue-block weight over the window, or the floor	19 N for 24 h today (USD 534k at 100 GH/s); x7 under rank 3

What no share buys: a block the nodes do not re-execute (the native-execution veto), a state the aggregated proof chain does not commit to, a lock past a certificate under two thirds of weight, a parameter change without the signal.

The residual risks stated plainly: the first 20 days (no weight table yet); a pause after a sudden departure of a third of weight (30 days under v3 until rank 2 lands); the Sybil count of keys (X5's definition adopted, the measurement scheduled); the proving pool until rank 1; the 2/3-of-total rule under long churn (the window bound: a 50/50 honest partition locks on both sides from day 10).

4. Per lane: the three biggest findings

Lane 1, consensus-security

1. The lock bounds a majority; the k-cluster does not (45 to 51 percent wins a 90-s race 70 to 85 percent of the time; the lock at 63 to 93 s is the bound).
2. The veto is cheap (20 days of 51 percent: USD 6k at 1 GH/s, 5.8 M at 1 TH/s, half earned back) and the pause is then free.
3. The proving pool is capturable today (11,636 IGN/h at 51 percent) with a correct-statement fake-proof record; the only attack that earns more than it costs.

Lane 2, algorithm

1. Class v4's verifier measured on a 2022 server core: 4.90 / 5.06 / 8.23 ms (steady / cold / half-core); dr736 9.76 / 10.51 / 15.49, out; R0 is dr368.
2. The f = 1 GDDR7 chip: 5.7x per joule against the 5090 at v3, 2.1x at v4 with k = 1; USD 0.00021 per MH/s-hour against USD 0.0117 rented; the FPGA soft overlay 0.30x to 0.47x per watt.
3. The reserve and the era draw buy about nothing against a chip; the N ladder in the era draw at genesis is the lever, and lane 7's HBM4 column is reconciled with three named disagreements.

Lane 3, finality-and-weight

1. Tonight's pause: the 2/3 rule at checkpoint 6843 (53.1 percent of total after a 42.7 percent departure), then the frozen table (Q5) holding it for a window; locks formed without the hub; topology refuted.
2. Only the departure announcement passes the line (0 conflicting locks everywhere, the pause under an hour); the decaying denominator, the hysteresis floor and the two-tier rule fail or are reports.
3. Weight capture costs USD 8,424 x N x W/(1 - W) to rent for the window: the veto 0.52 N for 30 days (USD 4,300 per GH/s of network), a lock alone 2.03 N (USD 17,100); lock delay measured p50 0.86 to 1.26 s at 4 to 93 voters; the path breaks near 1,000 voters on per-vote verification.

Lane 4, economy-and-utility

1. The proving price is h/N: 100 to 300x Boundless at 1.16 GH/s, 0.2 to 0.4x at 100 GH/s beside the miner; the adopted floor overprices above USD 0.014 per IGN.
2. Fees are not a security budget for a decade (USD 450 a day at launch, 4,200 in year 5, against 54,800 and 13,700 of emission); sustained honest hash costs USD 24.8 per GH/s-day against USD 281 rented, so the 20-day veto costs 11.8x the honest fleet at every price and year.
3. The 80/20 survives every stress but a ten-day prover refusal, which strands 547,570 IGN a day of pool credit with no rule to return it.

Lane 7, frontier

1. HBM4 raises the stored-dataset chip's edge (lane 2 bounds the figure at 4.9x to 12x bare by tFAW); the N schedule belongs in the era draw at genesis.
2. Vote weight is already a slashable, non-purchasable bond: work-stake for external jobs, with "no coin stake" written into the spec first.
3. The consensus proof can be incremental: the weight table inside the recursive segment proof, one mergeset per segment; the first measurement is the in-guest BLS verify and colouring cycle counts.

Lane 5, network

1. Reds come from node cost, not topology: run A at 10 blocks a second on a star ran 77.6 percent red because the hub needed 61 to 345 ms per block; the propagation model predicts under 0.1 percent red in every bps x delay cell with an ideal hub.
2. The controller reads blue work only, so a star or a withholder eases difficulty; the whole-DAG estimator holds the rate in every cell.
3. The recommendation: 1 block a second for the testnet and launch; 10 behind three gates (per-block CPU under 50 ms on a laptop core at mergeset 248, finality constants in DAA seconds, vote aggregation). The experiment is docs/analysis/block-rate-devnet2.md.

Lane 6, polish

1. The pause had no cause on any surface: no finality_reason or frozen-table share in the node's report, the observer, the API, Ember or the hub (Q1, Q2, Q6); the 0.3.16 Ember carrier is on ember-tune.
2. Every update was urgent once an activation height was behind the node (fork_is_close, Q4), and nothing refused an activation height at or below the live DAA; both fixed on ember-tune for 0.3.16.
3. Unsigned installers on both desktops (Q3, the project lead's certificates) and the relay's security fixes still on fud-close (Q8).

Lane 8, new-proof-of-work

1. Scheme A (mining is proving) is a bound, not a design: 2.9 MB of openings per block or a 32 to 40 ms verify against the 10 ms gate; the useful fraction is 8 percent at 1 GH/s and 0.08 percent at 100 GH/s.
2. Scheme B (tensor shadow) is free for the honest card and so nearly free for the chip (0.056 to 0.70 pJ per multiply-add against 11 pJ per ALU op); never as class content, kept as reserve R8.
3. Scheme C (dataset from stored state) is the class v5 candidate: the 4090's rate and watts equal within noise (63.083 against 63.088 MH/s), build +1.4 ms, verifier +0.11 to 0.21 ms per unit, bit-exact on 1,024 items and 128 lanes.

5. What was not run, and why

Item	Lane	Why
The in-guest BLS verify and GHOSTDAG colouring cycle counts (rank 12's first gate)	7, 3	no 24 GB card free tonight: PC 2 and the fleet were on class v4 and Devnet 2
The real 2019-class core (O-1.14)	2	the US laptop was not on the relay; the box's half-core proxy stands in
The FPGA soft overlay on real HBM2	2	no FPGA in the fleet; AWS F2 plan written
The AMD watts at every N	2	the ADLX sampler row is owed on the runner's --cards-off mechanism (counter-asic-3-status 6a)
A cargo bench of fast_aggregate_verify at 93, 1,000 and 8,192 keys	3	the BLS figures are arithmetic on blst's published timings, approximate
The 10 bps mesh run A2	5	dropped: run B (the 1 bps control) landed clean and the model attributes run A's reds to node cost, which a mesh does not change
The full 30 M-cycle shard beside the miner on any tier; the hash loss while proving on Ampere and Ada	4	the fleet measured the 4.7 M fixture only
Market prices for Taiko-class batch proving and Bonsai	4	not published; marked approximate
The observed end of tonight's pause	3	expected at DAA 216,402, about 20:40Z; the timeline closes when the observer rows show the lock

6. Rules and corrections for main

#	Rule or correction	Lane	State
1	P21 is a priced economic hole until the in-consensus verifier lands; the 20 percent pool is not "paid to provers" without the caveat	1	relayed
2	The one-day P2 signalling window is buyable for a day; seven consecutive windows	1, 4	relayed
3	The DAG controller reads blue work only, so a withholder or a star eases difficulty 16 to 32 percent (run A's loop)	1, 5	relayed; lane 5 owns the fix
4	No unwrap on a peer-driven path; the sibling list; a sync-request fuzz in tools/ci	1	relayed
5	A standing box never leaves the live chain for an experiment; orchestrated departures over 10 percent staged under 10 percent an hour	3	relayed; the fleet library refusal is rank 8
6	Under rule v3 any sudden departure of a third of weight is a 30-day mainnet pause; the leave item is the one safe way to shorten it	3	relayed
7	F19's "stalls until day 19 to 20" is a v2 number; under v3 a full window	3	relayed, ledger text owed
8	"No coin stake; the only thing at stake is 30 days of public work" into spec 03/05 and the litepaper before any work-stake prototype	7	relayed; routed to the site-miner agent
9	The litepaper's "proving: a second income" carries the USD 36 a day arithmetic	7	relayed; routed
10	Ember's updater installs nothing while finality is paused	7	relayed; sent to the Ember agent for 0.3.16
11	The customer brief's "priced in dollars per proof" and the litepaper's "proofs at the cost of power" are conditioned on network hash	4	relayed
12	The three signalling thresholds (60 / 90 / 95) in one sentence across spec 5.7, CLAUDE.md and the litepaper	4	relayed
13	The spec is silent on pool credit nobody claims; today it is a burn nobody decided	4	relayed
14	funding.md section 4 overstates the dev-fee ceiling by a quarter	4	relayed
15	Any future miner-signalled execution parameter outside the consensus digest makes signal-then-defect a quiet state fork	4	relayed
16	The era draw and the reserve are not unpredictability against a chip; the public text should say what they buy	2	relayed