By Luuk Joseph Soons · CounterProof Research · 29 October 2026
Why would an institution concerned with reviewing software spend time with Sweelinck, Bach and the people who preserve their music?
For me, the question began in listening. Wim Winters playing Sweelinck's Fortuna on the clavichord set the inquiry in motion. Hearing Xavi Torres play the Fantasia cromatica, I was reminded of Bach, particularly Daniil Trifonov's performance of the unfinished fugue from The Art of Fugue, with his own completion.
Something seemed familiar across the distance between these composers. What had travelled? A way of handling musical ideas? A teaching tradition? Or was part of the association something I brought to the performances?
The experience did not answer those questions. It made me want to ask them. That is a good beginning for research, provided we resist turning the feeling into its conclusion.
Music also makes certain problems easier to encounter. We can enjoy a performance, then become curious about its inheritance. We can value a tradition while asking how it changed. These questions have practical counterparts in our work with AI: what has been received, what has been added, and what should another person trust?
There is actual manuscript criticism behind this connection. Henle's commentary on Bach's Chromatic Fantasy and Fugue identifies marks indicating loudness in a later copy that were passed into another manuscript. The source from which the first copy descended lacked those marks. The editors identify them as additions by the copyist and note their appearance in modern editions. A later contribution could acquire the appearance of inherited authority through repetition. Henle's critical commentary, pp. 137–138.
The same commentary cautions that differences between manuscripts can preserve stages in Bach's own revisions. An editor cannot assume that every difference is a mistake. Copying, revising and interpreting are different activities, even when they leave similar-looking traces.
Now imagine a small software team. Its current policy allows a reservation up to and including a deadline. An older design note says that reservations must arrive before it. Three AI reviewers receive that old note and approve code which rejects a reservation made at the deadline.
Three approvals sound reassuring. Yet they may rest on one mistaken premise. A check against the current policy and an executed boundary test could expose the problem. Counting approvals would not.
This example is invented, not drawn from a client engagement, but the distinction is concrete. We want to know what each reviewer received and what each actually checked. Shared wording may suggest a relationship; an exposure record can establish that material was supplied. Neither alone proves what caused every subsequent conclusion.
Agentic Stemmatics investigates related questions about model outputs and their possible dependencies. Its harder proposals concerning model lineages require repeated observations and tests against alternatives. A shared phrase does not reveal a family tree. The musical case helps explain why the question matters; it does not establish that a proposed AI method can answer it.
Trifonov offers a different lesson. His recording includes his completion of Bach's unfinished final fugue, credited as his contribution. A listener is being offered a musician's continuation, rather than a newly recovered ending by Bach. Deutsche Grammophon's account of the recording.
That distinction leaves us free to appreciate the result. Attribution does not diminish a creative act. It lets us understand what we are hearing.
The corresponding duty in a research report is to distinguish the source from what we have made of it. An AI model can propose an explanation, assemble a useful comparison or suggest a repair. Those contributions should be identified and tested according to their purpose. An invented test result is a different matter. Its fluent presentation cannot give it the standing of an observation.
Agentic Representation asks how an account shapes what its reader understands. Attribution is part of this, but the question goes further. Return to our fictional reservation checker. A summary might truthfully say that three reviewers approved it and every test in the existing suite passed. If it leaves out the missing deadline case and the outdated policy, its accurate sentences still fail the person deciding whether to release the change.
We therefore have to examine both the claims and the account they form. What has been selected? What has been omitted? Which qualification would change the reader's decision?
Music also challenges our framing. Nicholas Cook argues that performance cannot be understood simply as the reproduction of a finished text. Meaning arises through the activity of performance and its relationships with other performances. Cook, “Between Process and Product”.
We should take that challenge seriously. Tracing a report's sources cannot tell us everything about the work it describes. Instructions, actions and results need attention too. Nor should we label every transformation a corruption: two implementations can satisfy the same requirement, just as different interpretations can be musically legitimate. We must first establish what the task asks someone to preserve or accomplish.
The third example concerns continuity. The New Bach Edition pairs musical texts with critical reports. Its revised series responds to newly discovered sources and fresh assessments of existing evidence. Preservation includes the possibility of changing an earlier conclusion. Bärenreiter's account of the New Bach Edition and its revision.
That gives us a useful question for our own institution. When a session ends, what should the next person or agent inherit?
For the reservation checker, “fixed the bug” is a poor answer. A successor needs the applicable policy, the boundary case, the correction and a warning about the obsolete note. We must then examine a later task to see whether the guidance was received and used. A beautifully organised archive can preserve an error very efficiently.
Agentic Institutions studies this broader problem of continuity through changing participants. The ambition includes a harder requirement than remembering: successors must sometimes reject what they inherit. A rule may have been wrong, or circumstances may have changed. Its presence in our library, even under the founder's name, should not protect it from a justified correction.
These programmes share concerns, but none guarantees the others. A source can be well traced and wrong. A report can describe it faithfully and still mislead through omission. An institution can teach that account for years. Each programme needs evidence of its own usefulness.
A sceptical engineer may ask whether this is ordinary quality assurance given elaborate names. Much of the foundation is familiar: testing, careful reporting, version control and learning from failure. We should acknowledge that debt. New terminology earns its place only if it helps specify an unresolved problem or produce an improvement that simpler practice does not already provide.
Better models may make some of today's procedures unnecessary. Our methods should face that comparison. If an ordinary checklist or one capable reviewer gives an equally useful result at lower cost, the research should help us recognise it. Institutional durability ought to include the ability to retire our own methods.
Music cannot establish those practical benefits. A client deserves evidence from the assessment of their software, a clear statement of what was checked, and an honest account of what remains unresolved. Nobody should need to like Bach to judge our service. Nor have we shown that a musical lesson improves technical understanding; that needs its own comparison on unfamiliar tasks.
But music can give us another way to begin. It invites us to care about how something valuable reaches another generation, and to notice the labour of those who preserve, interpret and renew it. That is part of the institution I want to build: one whose successors receive useful work and remain free to question it. The freedom matters as much as the inheritance.
Source note: This essay draws on published editorial commentary and performance scholarship. It does not reproduce a manuscript collation or an analysis of the recordings. The listening experience is the author's; the software example is fictional. Prepared with Codex (OpenAI).
Review note: A draft received an adversarial critique from Kimi (Moonshot), followed by small clarifications. Kimi could not retrieve the cited sources; the drafting assistant checked them separately. Requested Grok and Claude reviews produced no substantive returns. The review records are retained and available on request. This was model-assisted critique, not scholarly peer review or validation of the proposed AI methods.
CounterProof Research is part of a commercial adversarial-review practice of Clavestra Capital Limited (Malta, C 113987). The author has a financial interest in methods that could benefit that practice.
Read alongside Counterpoint in music, CounterProof in method, Different writers, shared traditions, and Earned, not inherited.
Cite as: Soons, L. J. (2026). Why music belongs in our research. CounterProof Research, 29 October 2026.