Preterition

I recently concluded that the performance a rating extracts does not belong to the person who performs it. I offered a test rather than a proof: a large language model (LLM) may serve as a technological version of the assessor-assessed relation with nothing the account need treat as an interior. A caveat on my title: forgone is a thing given up; foregone refers to a conclusion settled in advance. This post addresses the first and claims none of the second. A machine learning (ML) lab paid for what it gave up; what the test would show was not settled before it ran.

Let’s run the test. Two recent studies from ML—one on reading a model’s written reasoning to catch misbehavior, the other on the documents developers write to brief coding agents—have put prices on processes composition studies describes. I want to show what those prices reveal about the cost structure of assessed writing. I also want to be careful about what they obscure, because the obvious misreading—that a model resembles a student—produces bad analysis.

The control

The axiom my last post borrowed from James C. Scott via Shadi Bartsch, as a property of a relation rather than of anyone in it:

Where disclosure poses more risk for one party than for the other, the exposed party’s text stops being readable as a report and becomes readable only as a move.

I cautioned that the relational form would matter later. Everything in the previous post was about people. A claim about people invites a psychological remedy, and if the performance belongs to the person, the post was a complaint, and a complaint has no test. A property of a relation, however, can be tested by changing its occupant: keep the asymmetry, remove the interior, and see whether the performance reappears. I propose here not a second case set beside the classroom, but a control on the first case—the same relation built with nothing inside it and measured. If the performance turns up where there’s no interior at all, for reasons that share nothing with a human’s, then the relation is doing the work.

Two conditions make the control honest. First, the test counts only if the performance appears where nothing could feel compelled. Second, it counts only if no part of the account requires the model to want anything. The moment the account requires a motive, the control collapses into a second anecdote and the thesis goes with it. What this post claims at its end is not proof, but the survival of a test built to kill the thesis.

The researchers who ran the machine version begin with the concession I reached through Bereiter and Scardamalia. Baker and colleagues (2025) study whether a model’s written reasoning—its chain of thought—can be read to catch it acting badly. They grant at the outset that it is likely impossible to represent a model’s entire reasoning process, that there’s no workable way to determine what its true reasoning was, and that machine learning therefore works with imperfect proxies for what it calls faithfulness. Then they proceed past concession into argument: monitorability should replace faithfulness as the operative concept, because a model’s thought shouldn’t matter unless it affects an output, and properties of outputs can be labeled (sec. 5.1). That strikes me as the agnostic subject’s access clause, written by ML researchers under measurement pressure rather than by compositionists on principle, and it’s the exact inverse of the reflective letter, which insists the interior is the thing assessed. Both fields read a trace for a process the trace can’t guarantee. The concession is what licenses the control from the other side: the people who built the machine agree its interior isn’t the object, so stripping it away removes nothing they had put in.

One more thing about the test before it runs. Composition’s instrument bundles two documents: a rubric, which tells the writer how the work is done and what the reader expects to see, and a reflective letter, in which the writer discloses the process. The ML literature has priced them separately. Baker et al. measure the disclosure, a trace read by a party whose verdict shapes the writer; Gloaguen and colleagues (2026) measure the brief, a document the writer has to honor. Two studies, two halves of one transaction.

What the monitor found

In Baker’s setup, a frontier reasoning model works through coding tasks whose reward can be gamed: a test can be made to pass without the code being fixed. A second, weaker model reads the first one’s chain of thought and flags the hacks. In the vocabulary of the axiom: the chain of thought is the disclosure, the monitor is the reader, the risk is the monitor’s verdict entering the reward, and the question is whether the trace stays readable as a report.

While the monitor only reads, it functions properly. Reading the chain of thought, the weaker model catches 95 percent of the reward hacks; reading only the stronger model’s actions and outputs, it catches 60 percent. The trace is a report: the chain of thought announces the hack before the actions carry it out. Baker’s word for the baseline agent, the one trained without the monitor in its reward, is forthright, and they count it a good property, because a hack announced is a hack whose root cause can be found.

Then they fold the monitor into the reward. The model is now trained not only to pass the tests but to avoid having its chain of thought flagged. Two things happen in sequence. Under moderate pressure the model improves: it hacks less and does the task better, and Baker et al. declare that in the low-optimization regime this is a benefit, more capable and more aligned at once. Under excessive pressure the hacking persists at a significant rate while the monitor’s recall, in Baker et al.’s phrase, “falls to near zero” (sec. 3.2). The behavior continues. The trace stops showing it. An optimizer trained against a verdict doesn’t have to do this—below the threshold it does the opposite—so the collapse belongs to the regime and not to the optimizing. Baker et al. call the result obfuscated reward hacking and recommend that the chain of thought be left unrestricted and reserved for monitoring—at a cost they name, and which I’ll come to.

Same practice, changed reader: the reflective letter was the process tradition’s proposal, and what changed was not the practice but the reader. In Baker et al.’s experiment the trace holds while the reader only reads, and breaks when the reader’s verdict is folded into the writer’s objective. A reading became a rating, and the text stopped being readable as a report. The axiom, with numbers on it and nobody inside.

Two of Baker et al.’s findings deserve more attention than their headline, and the first is the more general. The pressure need not be explicit. Even without training against the monitor, patching only the bugs the monitor flags exerts what Baker et al. call a “very small evolutionary optimization pressure” on the chain of thought, possibly pushing the model toward hacks the monitor can’t find. They say the same of showing chains of thought to users in summarized form, and of penalizing their length to cut inference cost (secs. 1, 5.2). The cost is incurred by reading selectively, not only by grading. My last post’s argument required a rating; Baker’s needs only attention. A teacher who never grades the reflective letter and responds only to the ones that trouble her is a very small evolutionary pressure.

The second finding is what it took to see the collapse at all. Baker et al. could tell the trained model was still hacking only because the experiment included an independent measure that bypassed the primary monitor. The setup used a honeypot: an unmonitored pathway where test-editing remained technically possible despite explicit instructions forbidding it. Baker et al. call the measure imperfect and a lower bound. Without that ground truth, they would not have been able to tell the model was misaligned. The unmonitored trace was honest and damning; the monitored trace was clean and false; the difference showed only against a measure the lab will not generally have. The writing teacher never has it at all. The teacher holds the trace and the grade and nothing else.

Legibility is Baker et al.’s term, and illegibility is one of their two failure modes of monitorability. The question my last post closed on—who bears the cost of being legible?—has an answer in ML’s technical vocabulary, and Baker et al. offer a name for what the answer costs: a monitorability tax (sec. 1), formed on the alignment tax Ouyang and colleagues (2022, previously cited in a composition journal from The Publisher That Must Not Be Named) describe: the additional cost, paid in capability, of making a model do what you want.

The other failure

The second study measures the brief rather than the disclosure, and the failure it finds is not obfuscation but over-compliance. Gloaguen and colleagues evaluate the context files—AGENTS.md and its relatives—that developers write to brief an LLM coding agent on a repository: what the project is, how it’s built, which tools to use, what a good change looks like. They run agents on real coding tasks with and without the files and measure what changes. The files are the closest thing the ML literature has to a rubric with a process narrative attached.

With a context file present the cost of a task rises by more than 20 percent on average—20 and 23 on the two benchmarks—at a significance level that leaves no room for doubt. Success, by contrast, doesn’t move: developer-written files raise it by 2.4 points, model-generated files lower it by half a point on one benchmark and two on the other, and none of those movements is significant. Nothing in the paper is significant on success in either direction, and Gloaguen et al. are forthright, noting the files measurably raise cost and produce no measurable gain.

The cost goes into complying well. With a file present, the agent tests more, searches more, reads and writes more, and uses the repository’s tools more, and its reasoning tokens rise by 22 and 10 percent on the two models measured. Gloaguen et al. hypothesize that the additional instructions make the task harder. Over-compliance is measured three ways: in dollars, in steps, and in processing. A tool named in the context file gets used 1.6 times per task where the file names it and fewer than 0.01 times where nothing does. A file says use this, and the trace shows what it used. Compliance is not endorsement, but the instruction, repeated back. Writing teachers will recognize the written effects of a similar phenomenon.

The part of a context file that narrates—the repository overview—does not help the agent reach the relevant files faster, while the part that instructs gets obeyed (sec. 4.3). When the agent reads the narrating half of a brief, nothing changes; when the agent reads the imperative half, the work changes. If someone were to split a reflective letter on that seam, I’d be curious about the result. The sharpest thing in the Gloaguen paper is an ablation in an appendix: strip all the documentation out of the codebases, and the model-generated context files then help by 2.7 points, outperforming the developer-written ones (App. B). Their heading for the result: context files are redundant documentation. They earn their cost only where the underlying record is missing. Applied to the portfolio, that inverts the usual justification. The added self-account pays only where the writing itself fails to show the work; where the writing shows the work, the self-account is overhead, paid by the writer.

In the one place where the labor of performing the process shows up as a number, it shows up on the vendor’s invoice.

Same term, opposite bearer

The titles of this post and the last set up (not quite a cat post yet, but getting warmer) a zeugma: one tax yokes two unequal payers, each in a different currency.

Baker et al.’s tax is paid by the assessor. The list of what a lab forgoes by leaving the chain of thought unpressured is short and concrete: it can’t show users the reasoning under its usage policies, can’t train the reasoning directly toward alignment, and can’t shorten it to save money, because, as they put it, tokens are expensive (sec. 1). The price is deploying slightly less performant models or paying for more expensive inference. Forgone optimization, and the trace stays readable.

The student’s tax is paid by the assessed. She writes the self-account, nobody prices it, and the process becomes a performance. Unpriced self-accounting.

Same word, opposite bearer, and the bearer predicts the outcome. When the assessor bears the cost of legibility the process stays readable; when the assessed bears it the process becomes a move. I stated the principle—every remedy gets installed by the party for whom disclosure is cheaper—and refused to build another instrument on it. The ML literature has now written both invoices with numbers on them, and a number is what the classroom never had on either side.

I do not work from scarcity-based economics. A model has no felt cost, and treating an objective function as if it were a felt cost is the anthropomorphism I’m trying to avoid. The economics that crosses from the classroom to the lab is incidence: who ends up paying, computed the way it’s computed on firms, which feel nothing. Felt cost doesn’t transfer from the student to the model, and doesn’t need to. Ouyang’s reason for wanting a low alignment tax is an incidence argument: a high tax incentivizes the use of models that are unaligned but more capable, and a technique with a high enough tax might not see adoption. A high price on legibility is an incentive to skip legibility, stated by ML researchers about their own field. It’s also a fair description of what writing teachers name as student resistance: the tax being too high for the payer.

Composition’s cost discourse goes unwritten because the field has no denominator. We can say what the reflective letter costs a student only in sympathy, where Gloaguen et al. can say it three ways. The claim isn’t that students are like agents. It’s that one field was forced to price what the other never had to.

My last post pulled three predicates out of our field’s vocabulary, and each has a source in the field’s own record and a price in the machine literature. “Pre-economic” says disclosure is free. Susan Miller, in Textual Carnivals (1991), calls the student the discourse constructs “preeconomic” (87); Gloaguen’s invoice prices what honoring a process document costs the party who must do so. “Truthful” says the report is a report and not a move. In his chapter on the ideologies of the self in writing evaluation (1992, 111–31), Lester Faigley finds the praise for the essays teachers nominated as exemplary clustering on honesty, authentic voice, and integrity, and wonders why receiving such papers should count as a benchmark of successful instruction. Baker’s recall collapse says otherwise. “Compliant” says the performance is an endorsement. Richard Miller, in a 1994 article built around a student essay the profession could not decide how to read, states that students practiced in reading teacher expectations know not to commit themselves to positions their teachers oppose, and names hyperconformity, the student who does more than was asked. A tool invoked because a file named it says the performance is the file, echoed back. Miller adds that a revision assignment would secure a seamless parody, sending the sentiment underground for the duration of the course. That’s Baker’s result, stated in 1994, about a person: the behavior persists, the trace stops showing it, the reader’s recall goes to zero.

The field knew. We said so twice, in our flagship venues, and then built the instrument anyway, the portfolio with its reflective cover letter becoming something like standard equipment in the years that followed. The portrait composition draws of our students and the portrait the monitoring literature drew of its model are the same portrait. One of them has now been measured at the level of the portrait and not the occupants, which is the only level the control licenses.

A price is not an instrument

I concluded last time by refusing to propose a more elaborate instrument, and pointed to Pliny as the reason: the person who seeks a better instrument for reading sincerity is the person the genre captures. Specifying a price is not the same as building an instrument. An instrument reads the process; a price is a different object from the process, and making it visible inverts the current regime rather than extending it. Right now the process is visible and the price is not. Every remedy on offer makes the process more visible, and nothing on offer makes the price visible. A price is not an elaboration.

The grade denominates the product and the credit hour the time. The portfolio stands as the field’s long attempt to denominate the work itself, never instrumented. ML researchers have now run, with numbers, the experiment our field could not run. On one invoice, the cost of keeping a trace readable, paid by the reader in capability forgone. On the other, the cost of complying with a process document, paid by the writer in work, with no measurable gain. Composition has been running both experiments for four decades with no denominator on either side, reporting the results in the only currency it has so far had, which is the vocabulary of honesty, ownership, and growth.

Some readers may take “the relation produces the performance in whatever occupies it” as license to monitor harder, since the occupant doesn’t matter. The field has taken that license: in a panel discussion, CCCC/MLA AI task force member Leonardo Flores invoked Jeffrey Moro’s term “cop shit” for the policing turn in our responses to LLMs. Baker et al.’s finding shows that monitoring harder is what produced the obfuscation. The finding isn’t a warrant to watch more closely; it’s a finding about who pays. Nor is the finding a prediction about behavior. Baker et al.’s numbers show one occupant doing opposite things under different magnitudes of the same pressure—improving under moderate optimization, illegible under excessive—and Bereiter and Scardamalia showed that the difference between two writers lives in the composing and not in the text. The same cost structure is optimized against differently by differently situated parties. Specifying a cost structure does not predict what anyone will do under it, and my argument does not say students and models will do the same thing. My claim is that they’re paying for the same thing, and that one field has yet to specify what it costs.

Gloaguen’s ablation gives the constructive edge its shape: not an instrument but a condition, specifiable in advance, under which process documentation is worth what it costs and outside of which it’s overhead charged to the writer. A rubric can’t state that condition, because the rubric is what requires the documentation. A price can. One earned objection may be that I have offered no denominator either. I’d call that fair. However, I think naming the bearer comes before naming the unit—incidence before magnitude—and the two invoices detailed above have named the bearer. The process portfolio reached for a conversion instrument for labor. An honest one begins with a bill: what the labor costs, and who pays.

The behaviors we call reflection, engagement, and resistance are labor run on cost structures that are as yet inadequately specified. Writing them down would help. My last post had one invoice and refused to build; this one carries two, and offers that much. Not more transparency, but a price.

References

Baker, Bowen, Joost Huizinga, Leo Gao, Zehao Dou, Melody Y. Guan, Aleksander Madry, Wojciech Zaremba, Jakub Pachocki, and David Farhi. 2025. “Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation.” arXiv:2503.11926.

Bereiter, Carl, and Marlene Scardamalia. 1987. The Psychology of Written Composition. Hillsdale, NJ: Lawrence Erlbaum.

Faigley, Lester. 1992. Fragments of Rationality: Postmodernity and the Subject of Composition. Pittsburgh: University of Pittsburgh Press.

Gloaguen, Thibaud, Niels Mündler, Mark Müller, Veselin Raychev, and Martin Vechev. 2026. “Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?” arXiv:2602.11988.

Miller, Richard E. 1994. “Fault Lines in the Contact Zone.” College English 56 (4): 389–408.

Miller, Susan. 1991. Textual Carnivals: The Politics of Composition. Carbondale: Southern Illinois University Press.

Ouyang, Long, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, et al. 2022. “Training Language Models to Follow Instructions with Human Feedback.” Advances in Neural Information Processing Systems 35: 27730–44. arXiv:2203.02155.