What survives the handoff?

Why the form of an AI answer matters to what we learn from it

Luuk Joseph Soons · CounterProof Research
Essay · 1 November 2026

Can a report preserve a warning and still make it easy to overlook?

Imagine a software review that reaches a modest conclusion: a change passed the tests run with a small set of sample files. Larger files remain untested. One report puts that limitation beside the result. Another gives the result a prominent place and leaves the limitation in an appendix. Both contain the same factual statements.

We might expect readers to take different things from them. Whether they actually do, and whether that changes their decisions, would need to be checked. But the example makes the question clear. Preserving a statement and making its significance apparent are different achievements.

This is an invented example, not a finding from a CounterProof engagement. It concerns an ordinary problem: what happens when work passes from an investigator to a report, from a report to a library, and from that library to someone beginning a new task?

AI makes these passages easy to multiply. An assistant can read a review, produce a summary and supply the next assistant's background. Each step can help. Each also gives us another occasion to ask what has been selected, emphasised or left behind.

The score and the performance

Music offers a way into the question. A score makes some features of a work available to a performer. A recording preserves a particular performance. A teacher can demonstrate something that a pupil has struggled to hear or understand on the page. These forms can support one another, although receiving one is not the same as receiving all three.

Our essays on musical tradition began with a listener's experience: hearing something familiar across different composers and performances. That experience invites investigation. It does not tell us, by itself, who learned from whom. For that we turn to manuscripts, teaching records and other evidence of transmission.

There is a further question. What changes when music moves between notation, performance, recording and teaching? A record may preserve something with great fidelity while leaving a later listener uncertain how to interpret it. The same difficulty can arise with a perfectly preserved AI output.

This is why music has a place in our research. It gives us familiar experiences through which to approach less familiar questions. The comparison has limits: a model is not thereby a musician, and a training process is not thereby a school. But we can ask, in both settings, what a successor receives and what is needed to make good use of it. Our musical inquiry.

A correction needs somewhere to arrive

The history of printing offers another useful perspective. Elizabeth Eisenstein described how readers in different places could send corrections and additions to authors, editors and printers for later editions. The book could become part of a continuing exchange. Adrian Johns emphasised the work, skills and social relations through which dependable printed objects were produced. Their disagreement helps us attend to both the medium and the practices around it. 1

We should resist a simpler story in which printing suddenly made correction possible. Ann Blair describes manuscript textbooks used at Harvard between 1680 and 1730 whose corrections could be entered in the text or in an errata list. These manuscripts belonged to a society already using print, but the example still shows that a printed copy was not a prerequisite for attaching corrections to a text. 2

The useful question for an AI institution is therefore quite practical. Can we identify the account being corrected? Can someone find the correction? Does it explain what changed and which uses are affected?

Generating a fresh answer does not erase an earlier answer that was saved. Nor is a chat transcript necessarily an unstable record. A transcript can be preserved and cited. Conversely, a polished report can circulate after its conclusions have been withdrawn. The relevant properties must be examined in the actual records and practices.

An institution may have a large library and still have poor means of correcting it. It may also have modest tools and a dependable habit of linking an old conclusion to its replacement. What matters is how those arrangements work when a successor needs them.

What a table invites us to see

Suppose a review contains several findings. A table can help readers compare them, notice missing information and locate supporting evidence. It can also make separate rows look more independent than the findings really are. Two rows may describe consequences of the same underlying problem.

A narrative has different possibilities. It can explain that relationship, but it can also bury a qualification or make a weak connection sound persuasive. Neither format earns trust merely by being a table or a story.

Media theory and the history of writing invite us to examine such choices. Here, a medium is a means of expression and communication: speech, writing, print, recording or a digital interface. Its form is part of the conditions under which a reader encounters an account. What receives a heading? What becomes a footnote? Which questions does a form require its author to answer, and which does it never ask?

There is also experimental work on related questions. In a small experiment reported in 1994, Jiajie Zhang and Donald Norman asked eighteen students to solve versions of a puzzle with the same abstract rules but different arrangements of what participants had to remember and what the physical materials constrained. They found differences in performance, although not every comparison between conditions was statistically significant. This was a study of particular tasks, not of AI reports, but it provides a reason to investigate representational arrangements rather than assume they are incidental. 3

A useful comparison for our work would give readers the same evidence, reasons, limitations and dependencies in different formats. If one version supplies reasons that the other omits, we have changed the information as well as its presentation. We should also ask whether any improvement is worth the extra time required to produce and read it.

An archive cannot decide what deserves belief

Preserving a record protects the opportunity to examine it. It does not establish that the record is true.

The Homer Multitext provides a helpful documentary example. It makes distinct textual witnesses available with manuscript images and historical context. Readers can inspect the materials from which an interpretation is made. We can draw on that practice when preserving different model outputs and the circumstances in which they were produced, without assuming that models and oral poets generate language in the same way. 4

There is a lesson here for apparent agreement. Several accounts may repeat a common source. They may share an ordinary expression. They may also agree because each reaches the correct answer. Similarity invites questions about these possibilities; it does not answer them.

Once an account enters an institutional library, another question arises: what happens when someone challenges it? A correction that remains only in its author's notes may never reach those relying on the original. A durable institution needs ways to preserve objections, reconsider conclusions and make changes visible to later users.

Neil Postman's reflections on information and technological change help frame the question of attention. More material does not settle which material is relevant to a particular decision. His argument is a prompt for inquiry, not evidence that reducing the number of findings in a code review would improve it. An omitted finding might be the one that matters. 5

Three parts of the same journey

These questions meet the three strands of CounterProof's wider programme.

Agentic Stemmatics examines origins, transmission and dependence. If three reports agree, what does that agreement add? Preserving their separate records makes investigation possible; it does not establish that they are independent.

Agentic Representation examines the account presented to a reader. Does it convey the relevant grounds and limitations adequately? A report can preserve accurate sentences while encouraging a conclusion those sentences do not support.

Agentic Institutions examines continuity across changing people, models and sessions. Can a correction reach the next task, be understood there and change what happens? An archive is one part of that arrangement. Responsibility for maintaining and using it is another.

Media theory and the history of writing direct attention to the forms through which the work travels. Together, they bring questions and historical experience to programmes that must still establish the value of their own proposals. An introduction to Agentic Epistemics.

What this means for the work

For someone building software with AI, the immediate concern is straightforward. What can the next developer reasonably conclude from this review? Can they see what was examined, what remains uncertain and what would require another check?

Sometimes a short limitation beside the result may be enough. Sometimes a linked correction or a better table may help. Sometimes the existing practice will already be adequate. The research should be able to recognise all three possibilities.

Historical inquiry has a broader value too. It helps us see that familiar arrangements for preserving knowledge were made by people, under particular conditions. We can examine those arrangements, learn from them and question the authority they acquire. That remains worthwhile even when it does not yield a new reporting template.

The ambition is to help people and AI produce work that others can understand, challenge and develop. A useful record carries enough of its grounds for the next person to know what can be relied upon—and where inquiry should begin again.

Sources and further reading

[1] Print and correction. Elizabeth Eisenstein, “An Unacknowledged Revolution Revisited”, American Historical Review 107 (2002), especially p. 95; Adrian Johns, “How to Acknowledge a Revolution”, especially p. 117. The cited passages were checked in the original articles; this is not a new reading of both full books.

[2] Manuscript correction. Ann Blair, “Textbooks and Methods of Note-Taking in Early Modern Europe”, section “Student manuscripts at Harvard 1680–1730,” author-manuscript PDF p. 17.

[3] External representation. Jiajie Zhang and Donald A. Norman, “Representations in Distributed Cognitive Tasks”, Cognitive Science 18 (1994), 87–122, especially Experiment 2. Related reading: Jack Goody, The Domestication of the Savage Mind (1977), chapters 4–7, and Brian Street, Literacy in Theory and Practice (1984). These books are further reading, not evidence that a proposed AI reporting format works.

[4] Preserving witnesses. Homer Multitext: About and Mary Ebbott's interview on editors, scribes and readers.

[5] Information and attention. Neil Postman, “Informing Ourselves to Death” (1990) and “Five Things We Need to Know About Technological Change” (1998).

CounterProof Research essay. By Luuk Joseph Soons. Prepared with Codex (OpenAI) assistance, with a GLM (Z.ai) review of the drafts and supplied source notes. That review did not independently verify the underlying scholarship. The essay reports no new experiment or measured improvement in CounterProof's services. The author has a commercial interest in CounterProof's adversarial code-review practice.