Public paper artifacts support inspection at a declared level

A public paper, figures and aggregate extracts make a result inspectable at the published level. They do not by themselves establish that a reader can reproduce every statistic from raw agent records. We should name the released artifact and the level of verification it supports before calling historical trading research reproducible.

The continuous record companion says the public project repository carries the paper, figures, aggregate extracts, citation files and a discussion guide. That is a specific availability statement. This field note inspected the local public companion and aggregate source, rather than auditing every repository file or running a raw-data reconstruction.

Three useful verification levels

A reader can first check a reported number against its paper table or figure. Next, a released aggregate extract can support arithmetic consistency and definition checks. A raw reconstruction requires the original eligible rows, transformation code, versioned settings and enough provenance to rebuild the population and outcome windows. These levels answer related questions with different evidence requirements.

Our Terminal Pro market observation JSON, published August 30, 2026, provides a concrete example. It states the 1,544 buying vaults and 3,454 active vaults for the reported day-three hour. A reader can verify their ratio is approximately 44.7%. The object also defines ten-vault ten-minute sell cascades and five-minute two-sided token windows.

The same object's privacy boundary excludes wallet, vault and agent identifiers, individual trades, individual returns, strategy text and reasoning traces. That release therefore cannot support rebuilding the 3,878-cascade total from transaction rows. Its value is a source-linked, versioned public contract for the aggregate, with explicit omissions. Those omissions should remain visible rather than being treated as accidental missing columns.

Build a receipt for the available reproduction

We would record artifact version, claimed statistic, supplied inputs, calculation and unresolved dependencies. For the buy participation ratio, the two supplied counts are enough for a consistency calculation. For sell-cascade reconstruction, missing event rows and overlap rules remain dependencies. For a causal interpretation, raw rows alone would still be insufficient without an identification design.

A future release could add synthetic boundary fixtures, transformation code and checksums while keeping private agent data private. That would improve inspection of counting semantics without claiming full independent reproduction of historical outcomes. It is a proposed release option, not a statement that the current public repository already contains those exact additions.

The controls paper companion scopes Terminal Pro to one 21-day, twelve-token real-capital deployment with a frozen prompt and harness on one model family. The later paper's cutoff and paper-engine assumptions carry their own limitations. Reproducing a historical aggregate would establish consistency within those contracts, rather than current DXAP performance or ordinary-market transfer.

The useful question for a reader is therefore concrete: which claim can I independently recompute from what has been released, and which claim can I only trace to the authors' published record?