Roundtrip measures whether tokenized stocks on Robinhood Chain can actually be sold, and makes every number recomputable. At a pinned block, it buys with USDG and sells the whole position back throug
Executability Report — whether tokenised stocks can actually be sold, measured and published.
A collateral parameter depends on what comes back when you sell, not on the quoted price. So every 30 minutes this project pins a block and replays a full same-pool round trip — buy with USDG, sell the whole position straight back through the same pool — for the tokenised stocks on Robinhood Chain, at four sizes from 1,000to1,000to100,000. Each round is canonically serialised, hashed and published with a checksum, and the round hashes are anchored in an append-only
EvidenceLedgeron chain (0xc4f7…6fb9, chainId 46630), so a published number cannot be quietly rewritten.The finding is that failure is not one thing. Read down a token and the numbers do one of three things:
collapse — QQQ returns 1.93% at 1k,0.261k,0.2610k, 0.03% at $100k: one pool, and the same near-zero answer at every size.
switch — RDDT at $100k comes back either near 96% or near 0.04%, with nothing between.
slope — GME grades down as size rises; no cliff and no step.
Three shapes, three different answers to what size limit is safe.
Anyone can check it:
python code/verify.py 61129566recomputes the round hash from the published files and compares it with the record. Ten of this project's own errors are published too — numbered, with which were caught before publication and which were not.
<p><strong>Dated, because the measurement line predates the buildathon and it matters which is which.</strong></p><p><strong>2026-09-14</strong> — the experiment was pre-registered before the forward window opened. Four commits, all on the 14th: an absolute floor trigger, the five stable token names, canon pinned to a serialisation version, the sampling rate and drop policy pinned, and the block-selection method recorded as <em>deliberately</em> unpinned. The point was to write down what would prove the result wrong before any forward data existed. (<code data-inline="true" spellcheck="false">git log 68b2fad^..feb560a</code> in the <code data-inline="true" spellcheck="false">executability-report</code> repository.)</p><p><strong>2026-09-15</strong> — the forward window opened, and <code data-inline="true" spellcheck="false">EvidenceLedger</code> was deployed to Robinhood Chain testnet at <code data-inline="true" spellcheck="false">0xc4f7c2ed489d9f521d65b43cc4929d3c642c6fb9</code> (chainId 46630), together with its commit script.</p><p><strong>2026-09-19</strong> — the backfill was broadcast: 610 rounds committed on chain in 10 batches; the ledger watermark moved from 0 to <code data-inline="true" spellcheck="false">64,907,249</code>.</p><p><strong>Also during the window</strong>: the verifier was rewritten so its user-facing output is English and so that it <em>refuses explicitly</em> rather than hanging when pointed at a non-archive RPC endpoint (the refusal is measured, not asserted); a canonical serialisation specification (<code data-inline="true" spellcheck="false">rhdepth-v2</code>) was published so a second implementation need not read the reference code; the error archive grew to ten numbered entries, each with what it was and how it was caught; and the demo material was produced.</p><p><strong>The measurement line itself started 2026-09-04</strong>, ten days before the buildathon, and has run continuously since: 706 rounds over 15 days through 2026-09-18, plus three quarantined directories kept rather than deleted.</p>
<p>No external funding. Self-funded.</p>
