Buffered togetherfee decides
Pairs sent before the window that landed in its first slot: the higher fee came first.
Solana mainnetmeasured 2026-10-31 to 2026-11-04, after Alpenglow
In all ten validators the higher-fee transaction came first in more than half of the usable pairs, p from 0.59 to 1.00, but never strictly. Four reached the preregistered N of 100: AgaveBam-1 "fee rule holds, partial", JitoLabs-1 "mixed", FireBAM-1 "fee rule holds, partial", HarmonicAgave-1 "fee rule holds, strong". The other six are descriptive, N below 100.
Three regimes
Pooled over all validators, from results/secondary.json. p is the share of pairs in which the higher fee came first; q the share in which the first-sent came first.
Pairs sent before the window that landed in its first slot: the higher fee came first.
Two requests in random order: the first-sent came first.
A randomly labelled pseudo-high transaction came first about half the time.
Results by validator
Dots are point estimates, bars are exact 95% intervals. The reference line is 0.5, the faint band is the preregistered "no rule" zone, 0.35 to 0.65. Sorted by N. Hover a dot for the count.
Analysis: final: phase main only, real data; one Holm family of 20 tests; family alpha 0.025. Hollow ticks mark the first analysis of 2026-11-01 (2,649 pairs), kept unchanged in results/first_analysis_2026-11-01/. Verdict rules (N of at least 100; strict, strong, partial, mixed, no rule) are fixed in docs/PREREGISTRATION.md. Source: results/verdicts.json.
| Validator | Client, version | N | p [95% CI] | q [95% CI] | Preregistered verdict |
|---|---|---|---|---|---|
| AgaveBam-1 | AgaveBam 4.4.0-rc.1 | 196 | 0.81 [0.74, 0.86] | 0.61 [0.54, 0.68] | fee rule holds, partial |
| JitoLabs-1 | JitoLabs 4.4.0 | 138 | 0.64 [0.56, 0.72] | 0.78 [0.70, 0.84] | mixed: fee rule partial and request-position rule partial |
| FireBAM-1 | FireBAM 26.10.5 | 119 | 0.77 [0.69, 0.84] | 0.52 [0.43, 0.61] | fee rule holds, partial |
| HarmonicAgave-1 | HarmonicAgave 4.4.0 | 110 | 0.94 [0.87, 0.97] | 0.61 [0.51, 0.70] | fee rule holds, strong |
| Rakurai-1 | Rakurai 4.3.2, 4.4.0 | 37 | 0.59 [0.42, 0.75] | 0.70 [0.53, 0.84] | insufficient data (n=37) |
| Firedancer-1 | Firedancer 26.10.5 | 27 | 0.93 [0.76, 0.99] | 0.41 [0.22, 0.61] | insufficient data (n=27) |
| HarmonicFiredancer-1 | HarmonicFiredancer 26.10.5 | 27 | 0.89 [0.71, 0.98] | 0.63 [0.42, 0.81] | insufficient data (n=27) |
| Agave-1 | Agave 4.3.2, 4.4.0 | 23 | 0.61 [0.39, 0.80] | 0.83 [0.61, 0.95] | insufficient data (n=23) |
| HarmonicFiredancer-1 | HarmonicFiredancer 26.10.5 | 5 | 0.80 [0.28, 0.99] | 1.00 [0.48, 1.00] | insufficient data (n=5) |
| HarmonicFiredancer-1 | HarmonicFiredancer 26.10.5 | 1 | 1.00 [0.03, 1.00] | 1.00 [0.03, 1.00] | insufficient data (n=1) |
The delivery path
Share of probes whose pair landed together in the target's own leader window, all arms (results/runs.md). Firedancer-family and Harmonic leaders (Firedancer-1, HarmonicFiredancer-1) included our RPC-sent pair in their own window in only 7% to 9% of probes, against 16% to 49% for the others. Most of their pairs landed one or more windows later, in other leaders' blocks.
Landing itself was not the problem: 4,493 of 6,572 main-run transactions landed (68%), high and low fee alike. Source: results/runs.md.
Depth of the work
Nothing on this page is typed by hand. The numbers come from the result files, the documents and the code at build time.
Method
Two of our own transactions, from two payers, both writing one fresh shared account. The scheduler must put one before the other. The compute unit price differs 5,500x: 5,500,000 vs 1,000 micro-lamports per CU.
Which transaction is high-fee and which goes first in the request are drawn independently. Sent as one JSON-RPC request about 270 ms before the target's leader window; a separate arm sends two requests 200 ms apart.
A pair is usable only if both transactions sit in one block produced by the target, read from the block's Fee reward and never assumed from the schedule, inside the targeted window.
Estimands, thresholds, N of at least 100 and the equivalence margin were fixed before the data. Hashes of the plan and of the data were written to mainnet as memo transactions.
How the design handles each point of the bounty, shortened from docs/DESIGN.md.
Windows are 4 slots and 4-aligned (measured). Targets are frozen identities, one per client type, chosen by a stated rule that requires a stable gossip claim (same clientId and version over at least 18 hours) because claims flap: 33 of about 3,900 nodes changed claim within 80 minutes on 2026-10-30.
Our VM cannot send UDP or QUIC, so every transaction goes as an HTTPS JSON-RPC request to one provider.
Within a pair the interaction is the design (shared W in arms P, E, S; none in arm D). Across pairs there is none by construction: one probe per window, a fresh W per probe, distinct payers, a payer reused only after confirmation or expiry, at least 2 s between probes.
Other transactions compete with ours. Ours sit high in the priority queue (about 1.1 and 3.4 lamports per cost unit, while 18% to 50% of transactions pay no priority fee), so the leader can schedule both promptly; the order between them is still the leader's decision.
The position of the high-fee transaction in the request is randomized independently of the fee assignment, giving q as a second estimand. The spaced arm sends two requests 200 ms apart in random order, the only arm with a known send order, to test for an arrival rule and for overtaking.
Pre-specified estimands and verdict thresholds; an equal-fee arm as the noise floor; an equivalence test (90% intervals of both p and q inside [0.35, 0.65]) as the only route to "no rule detected"; Holm correction over all tests; synthetic data with planted rules and two localnet known answers as checks on the analysis code.
Timeline
Every anchor is a memo transaction on mainnet carrying the SHA-256 of a file in this repository. Compare it with sha256sum of the file at the named commit.
Verify in 60 seconds
Pick a random sample of our landed transactions, fetch the blocks from a public RPC and confirm the recorded positions and producers. No wallet, no key, nothing is sent.
git clone https://github.com/recruitabx/Inside-The-Solana-Block.git
cd Inside-The-Solana-Block
docker compose run --rm verify
python3 -m venv .venv && . .venv/bin/activate && pip install -e . && itb verify
Last sample: 30 of 30 random landed transactions PASS, results/verify_sample.txt.
The cap is enforced in code before every send; fees settle only for transactions that land. Audit: PASS, every fee-paying transaction of our accounts is in the spend ledger.
Honest limits
The full list of problems, in order, is in What went wrong in the report.