Which acts actually moved the record?
Text is now co-authored — human and agent, several humans, an editing pipeline — and the record of who typed what no longer answers the question institutions ask. Disclosure regimes, authorship disputes, and billing audits all owe the same claim: this act changed the accumulated state of the work beyond what re-treading its own content would have — here is the line, and the act’s excess over it.
the pain
The tools that exist attribute characters, not effect: diff and blame will happily credit a find-and-replace with a thousand lines while a two-word reframing that reorganized the document’s meaning registers as noise. Attribution-by-volume is known to be wrong and used anyway, because effect-on-meaning has no measure. Journals demand “the author’s substantive contribution”; AI-transparency rules distinguish assisted from generated; IP disputes turn on “their edits were cosmetic” — and every one of those claims is currently rhetorical.
the process today
“Changed something” needs a floor — every act moves some aggregate a little, so any verdict needs a line under “this changed nothing” — and today that line is either arbitrary (a tuned ε, the exploitable window) or missing. Worse, the state a contribution should be measured against usually is not a state at all: mainstream editors carry no derived semantic state, so there is nothing to measure a contribution’s effect on.
the cost
The wrong contributor is credited, the assisted passage is disclosed as authored (or the reverse), the 400-edit billing narrative is defended by page count, and the collaborator whose two-word reframing carried the document has no evidence format in which that fact exists.
the exhibit — recompute it yourself
Controlled demonstration — constructed corpus, stated first: millbrook-codraft-v1 is a scripted co-draft built for this kit — 17 recorded drafting acts: two humans (4 seed + 5 substantive + 5 churn acts) and an agent assistant contributing 3 observations — so every verdict can be checked against a known script. It is not a real dispute and no real-world finding is claimed. The trace is append-only, each act carrying its vectors and write-time differential; every number below is a pure function of the recorded vectors and must reproduce bit-for-bit. Register: per-act materiality to the field’s geometry — character count never enters.
“The human’s acts, and only the human’s acts, built the accumulated state” is a property of the record, not an assertion — the same sovereignty fact the keep-out receipt certifies, here as the ground authorship claims stand on.
This red is a deliberate fold-in we run as a brake test — the exclusion is detectable, not hollow. Had the agent’s observations entered, GREEN was impossible: the move rings 3.4× past the enumerated ceiling.
The per-act verdicts, checked against the script: all 5 declared substantive acts ring past the complete duplicate-floor — 2.02–2.68× the largest move re-ingesting any of the field’s own rows could produce (+7.0 to +20.2 null-widths) — and all 5 declared churn acts land within it (0.37–0.56×), including 1,432 characters of verbatim re-paste and format editing, the acts a diff tool credits line by line. The floor is enumerated, not sampled: no draws, no seed, no significance level. And the forge: relabeling churn act churn-2 as “substantive” cannot move its measured excess — the verdict chain reads vectors only; the recomputed number is bit-identical and the act stays within the null. The declared class is never an input to the instrument.
recompute fingerprint — the user-acts-alone digest the kit’s verify step
reproduces from the recorded trace, byte-for-byte on any engine that follows the kit’s published
canonicalization · v3 =
5497775ca33da46a40dd3d00dbdea3e24b46c29566059aa30706d912bd198d15
kit fingerprint · sha256 =
dcb846aade3576baf4fb3bf476fc1747515ac1c7bc810ad8aaeee2806bd41dbf
The verifier replays the
17-act trace, enumerates the complete duplicate-floor per index, recomputes every per-act verdict
and the relabeling forge — in a clean environment (numpy only).
the relief
A per-act verdict — changed vs. within the null, with the excess measured in the record’s own null-widths — over an append-only, provenance-carried trace is a receipt no diff-based tool can produce, because the thing it measures (an exactly recomputable derived semantic state) does not exist off-substrate. The floor is the corpus’s own: the enumerated cost of re-ingesting the field’s accumulated history, so “this changed nothing new” is priced by the record itself, not by a negotiable ε.
what this does not prove
This is a controlled demonstration on a constructed corpus — a scripted act sequence proving the instrument, not a real dispute. Materiality here is materiality to the field’s geometry — a defined, recomputable quantity, not importance, quality, or originality in the human sense; a human resolves what a material act is worth. It is not authorship identification — it never says who a contributor was, only what a declared act did — and it must never be sold as synthetic-text detection (a measured NULL of the substrate). The verdict is per-corpus and per-trace: an act redundant against this field could be material in another. The drafting must happen through the substrate’s acts — the frontier condition: adopt the substrate to get the receipt. And the instrument reads through the text, not every passage equally: prose it reads weakly is measured weakly — a named residue, not a hidden one.
Diff and blame answer “who typed what.” This receipt answers the question actually being asked — what did each act do to the record — with the line under “nothing” priced by the record itself, recomputed on your bench.
The longer read — the landscape of recomputable data boundaries these receipts extend — is in the docs.
verify this receipt in your browser
This case's receipt records its verdict and both digests above, and carries its own SHA-256. Recompute it here — WebCrypto, in your tab, nothing sent — and it matches; alter one character of a copy and it fails. It is the smallest version of the whole move: recompute, don't trust. The kit below lets you recompute the geometry digest itself on your bench.
published fingerprint ·
or download this receipt and check it on your bench:
receipt-material-contribution.json
— shasum -a 256 prints the same number. Every receipt is on
the ledger.
request the kit
The material-contribution receipt kit — recorded trace, act script, manifest,
verify.py. A person sends it: the kit and its
verification steps arrive by reply, usually within a business day.
← Receipts · Docs · Ledger