Parasitic Load

Property managers have asked me whether CommunityPay's ledger and payment rails will sit underneath their accounting. In each case the accounting above the rails is model-driven. A large language model assigns the general ledger account, routes the invoice, classifies the expense, and proposes the entry. A person reviews some of the output. The request is that the ledger become the system of record for whatever the model produced.

There is a second version of the request, and it is more common: no ledger at all. A free consumer payment rail, a spreadsheet, and someone checking off who paid.

The answer to the first is no. The reasons are arithmetic, evidentiary, and — the part that gets discussed least — psychological. The second is the same argument with a person in the model's place, and it is the larger of the two.

What Parasitic Load Means

In mechanical and electrical engineering, parasitic load is power a system draws to run itself rather than to do the work it exists for. The alternator turning against the crankshaft. The pump circulating coolant for the plant that drives the pump. It is not waste in the sense of a leak. It is real power, correctly delivered, returning no useful external work.

A decision that is already determined has one correct output. This vendor, this fund, this account, under this policy. Deriving it once and recording the derivation costs what it costs. Deriving it again on every occurrence costs that again, and again, and returns nothing that was not already known.

That is the shape of the argument. It repeats wherever a determined thing gets derived again, and the rest of this essay follows it through four places: the arithmetic of error, the record that has to survive an examination, the review that is supposed to catch what the first two miss, and — where there is no ledger at all — the volunteer who does the deriving by hand. Each draws load. None returns work.

The Arithmetic

Take an association of 300 units. Twelve assessment postings per unit per year is 3,600. Receipts against them, roughly another 3,600. Forty vendor bills a month is 480, and the payments against them another 480. Late fees, NSF fees, reserve transfers, and the close add several hundred more. Call it 8,000 posting decisions a year, rounding down, for one association of ordinary size.

Assume a per-decision error rate of p, and assume errors are independent. They are not — a model's errors cluster on the vendors and categories that are ambiguous — but correlation changes where the errors land, not how many to expect: at 99.9 percent, eight expected errors are eight expected errors however they arrive. Clustering lowers the chance of at least one error in a year, so the last column of the table is the case least favorable to the model. Under independence, the probability of at least one error in the year is 1 − (1 − p)^8,000.

To hold the chance of even one misposting anywhere in the year below 5 percent, per-decision accuracy has to reach 99.999359 percent. No one claims that number for classifying invoices.

A human bookkeeper does not hit four nines, let alone five. So the argument is not that people are more accurate than models. It is that the two failure modes have different structure. A deterministic rule that is wrong is wrong the same way 8,000 times. It is one defect. Every entry it touched carries the name of the guard that passed it and a hash of the policy in force, so the affected set is one query, the fix is one place, and the correction runs across the set. A probabilistic system at 99.9 percent produces eight errors a year. Two of them may share a cause — the same ambiguous vendor, read the same wrong way in March and in June — and nothing in the record says so. Each is discovered on its own, explained on its own, and corrected on its own, because no entry names the rule that produced it. There is no rule to name.

The reason is structural. A rule is derived once and applied 8,000 times, so every application shares one derivation and every error shares one cause. A classifier derives its answer again on every input, so every error has its own cause, and nothing about the third connects it to the fourth.

A rule concentrates its errors. A classifier distributes them.

Concentrated error is a bug. Distributed error is a fog. The first is fixed in an afternoon. The second is what a forensic accountant is hired to walk out of, at hourly rates, two years later.

Materiality Governs the Opinion, Not the Identities

The first draft of this essay said institutional financial reporting requires 100 percent accuracy, not 99.9. That is wrong, and a CPA would have struck it in the margin. Audits apply materiality thresholds: the auditor accumulates the misstatements found, evaluates whether the ones left uncorrected are material, individually or in the aggregate, and issues an unmodified opinion over the ones that are not. That is the wrong frame for this problem.

Two things in a ledger are binary, admit no threshold, and can be broken by a wrong account. A control account agrees with its subledger or it does not. A bank reconciliation ties or leaves a difference someone has to explain. The trial balance is binary as well, and it is not on this list: a double-entry system balances it by construction, and an entry posted to the wrong account leaves it balanced. Materiality governs the auditor's opinion on the financial statements. It does not govern whether the ledger's own identities hold.

The vertical sharpens this further. A misclassification between two operating expense accounts might be immaterial to an opinion. Reserve money spent on an operating expense is not an immaterial presentation question. The statutes that permit it at all require the board to record the withdrawal, notify the owners, and adopt a repayment schedule, and a misclassification skips all three, at any dollar amount. The classification errors a general materiality threshold would wave through are the ones that carry consequence here.

The Fab

A semiconductor line does not run because any wafer is perfect. It runs because the process is characterized. Photolithography masks, etch chemistries, doping profiles, thermal budgets — every variable specified, every measurement logged, every tool identified. When a wafer fails inspection, the line walks the defect back to a step and a tool.

Traceability is the product. The wafer is the output.

A ledger has the same structure. Every journal entry has one authority that produced it. Every authority leaves a hashed record, chained to the one before it. The chain from invoice to bill to payment to entry to bank reconciliation is reproducible, and the entry at its center names the rule that governed it.

Introduce a probabilistic step upstream and the defect no longer walks back. The entry is traceable to the model, and the model is where the trail ends.

Reproducibility, Conceded

The standard reply is that temperature 0 makes inference deterministic. As deployed, that is false, for reasons that have nothing to do with sampling.

In September 2025 Thinking Machines Lab sent "Tell me about Richard Feynman" to a 235-billion-parameter open-weight model 1,000 times at temperature 0 and got 80 unique completions. All 1,000 agreed through token 102 and first diverged at token 103. The cause was not the one usually invoked, concurrent GPU threads adding in a different order on every run. Run the same batch twice and the kernels return the same bits. What changes is the batch size. Batch size changes the reduction strategy a kernel picks, and the strategy sets the order in which the same numbers are summed. Floating-point addition is not associative, so the sum can change with the order. That is the absence of batch invariance, and it lives in three places: normalization, matrix multiplication, and attention. Carry that into a posting pipeline and the classification of an invoice depends on who else was calling the endpoint that second.

They then fixed it. With batch-invariant kernels in place, all 1,000 completions came back bitwise identical.

So reproducible inference exists. Take the concession at full strength, and it still does not reach the requirement.

Bitwise reproducibility means the same input returns the same output on the same weights and the same kernel stack. Weights are updated. No one keeps a 2026 model's weights, its serving stack, and the exact kernel build so that a 2026 classification can be re-derived during a 2031 examination. Where all three are preserved, replay returns the output. It does not return the reason.

A reproducible black box is a recording, not an explanation.

What Someone Has to Sign

Federal Rule of Evidence 902(13) admits a record generated by an electronic process or system as self-authenticating on the certification of a qualified person that the process produces an accurate result. The certification is authentication only; the hearsay exception under Rule 803(6) — records of a regularly conducted activity, made by someone with knowledge, kept as a regular practice — has to be established separately.

Read the operative phrases against a model-driven posting pipeline. A qualified person certifies that the process produces an accurate result. A regular practice produces the record.

I do not know how a qualified person makes that certification about an LLM classification. The vendor documentation does not say the process produces an accurate result. Anthropic's commercial terms say factual assertions in outputs "should not be relied upon without independently checking their accuracy, as they may be false, incomplete, misleading or not reflective of recent events or information" (Commercial Terms of Service, D.3, effective June 17, 2025). Whether a qualified person can certify that process is the question I would want answered before putting a system of record behind one.

Why the Review Does Not Happen

Every operator describes a human in the loop. The loop is the control. Five well-documented effects act against it, and they are not failures of the individuals involved.

Fluency is read as accuracy. Processing fluency raises judged truth and judged competence independent of content (Reber and Schwarz 1999; Alter and Oppenheimer 2009). A language model's output is fluent by construction. It is trained to predict text people wrote, then tuned toward the responses people preferred, and fluent text scores well on both. The cue a reviewer uses to allocate scrutiny is the one cue this system emits whether or not the answer is right. A fluent wrong answer is rated truer than a disfluent one. The reviewer is not predicted to look harder at it. The reviewer is predicted to rate it right.

Automation bias, in both directions. Skitka, Mosier, and Burdick (1999) separate errors of omission (failing to respond to a problem the automation did not flag) from errors of commission (following the automated directive against contradictory information the operator already had). The remedy came in their next study: participants made accountable for their overall performance, or for the accuracy of their decisions, made fewer errors of both kinds (Skitka, Mosier, and Burdick 2000). Accountability of the person doing the checking, not reliability of the aid. In a community association the person legally accountable is a volunteer treasurer who did not make the decision, was not present for it, and cannot inspect it.

The vigilance decrement. Sustained detection of rare signals degrades within the first half hour of a watch, a result old enough to have been established on radar operators (Mackworth 1948). "We review the output" is a claim about invoice 4. It is not a claim about invoice 400. The pipeline's entire value proposition is volume, and review burden scales with exactly the quantity being automated.

The illusion of explanatory depth. People rate their understanding of a mechanism far above what they can produce when asked to explain it, and the gap is largest for explanatory knowledge — how devices and natural phenomena work — and small or absent for facts, narratives, and procedures (Rozenblit and Keil 2002). One source they name is a confusion of levels: knowing what a thing does is taken for knowing how it does it. Input in, classification out, is that confusion with nothing in the middle to correct it. I have been told a system was perfect. In hindsight I should have asked how it was measured, and against what. I did not ask.

The conjunction is overestimated. "It is right 99 percent of the time" is heard as a statement about the model. It is a statement about a year of postings, and the table above is what it says. Asked for the chance that every step in a long chain succeeds, people anchor on the per-step figure and adjust too little. Offered a bet on drawing a red marble seven times running from a bag that is 90 percent red, most took it over an even single draw, though the seven-draw chance is 48 percent (Bar-Hillel 1973). Tversky and Kahneman's example was a system that fails if any component fails: "Even when the likelihood of failure in each component is slight, the probability of an overall failure can be high if many components are involved" (Tversky and Kahneman 1974). Per-decision accuracy is salient; the count of decisions is not.

The control is the person. All five effects act on the person. Nothing in the pipeline acts on the five effects, and the four thousandth entry arrives whether or not the person is ready for it.

The Other Substitution

Now the second request. Not a model under the ledger — no ledger at all.

Start with the word "free." Zelle is owned by Early Warning Services, a company held by seven large banks. Funds move bank to bank; Zelle holds no balance and earns nothing on one. What the owners get is the network and the deposits that stay inside it. A stored-balance wallet works the other way, holding the balance and earning on it. Neither arrangement is free. The price is real, and it is paid somewhere that does not appear on the association's invoice.

CommunityPay holds no association money. Resident funds land in the association's own Stripe Connect balance, and the platform does not sit in the flow. There is no float to pay the cost out of, which is why the ACH cost appears as a line item instead of disappearing into a spread. A visible fee and an invisible one are not the same fee at a different price. They are the same cost, itemized or not.

Then the work the rail does not do.

Verification is a loop, not a task. A consumer payment carries a person's name, not a unit number. Someone logs in, reads a transaction list, matches names to units, marks a spreadsheet, and chases what did not arrive. Then the late fees, by hand, under whatever the governing documents actually say. One set of rules on this desk gives grace through the 20th, posts a $10 fee the day after, moves the last day of grace forward when it lands on a weekend, and moves it again when it lands on a holiday. That rule runs twelve times a year, and the weekend and holiday shifts land on different dates each time.

Segregation of duties is gone by construction. The control is that whoever authorizes a payment is not whoever holds the account, and neither of them keeps the record. A phone and a spreadsheet in the same hands is one person holding all three. Nothing about the arrangement is dishonest. The control is simply absent, and its absence is not visible in any artifact the association produces.

The payment is final. Once the recipient is enrolled, an authorized Zelle payment cannot be canceled, and the network's own words cover the rest: "Unlike a credit card, Zelle does not offer a protection program for any authorized payments made with Zelle." Regulation E (12 CFR Part 1005) reaches an unauthorized transfer and an incorrect one. Whether a payment the resident authorized and sent twice is "incorrect" is a question the rule does not answer, and the network's answer is that it cannot be reversed. The correction is a second payment in the other direction, arranged by hand, with nothing binding it to the first. Two unlinked movements where there should be one reversal, and nothing in any record says they belong together.

The record does not outlive the person keeping it. Membership, ownership changes, balances, who owed what four years ago: a spreadsheet on a volunteer's laptop and an app tied to a personal phone number. Boards turn over every year or two, and the file goes with the person or it does not.

And then the largest one: the reports. This ledger renders each of them on request from the same posted entries. The statements: the balance sheet; the statement of revenues and expenses, operating and reserve shown apart, read against the adopted budget; the cash flow statement; and the fund summary across operating, reserve, special assessment, and capital improvement funds. Under them, the trial balance and the general ledger. Beside them, the operating reports: budget versus actual, budget variance and its trend, A/R aging, A/P aging, and reserve funding. A full GAAP presentation also includes a statement of changes in fund balances, which reconciles each fund's beginning and ending balance with the results of operations and interfund transfers. That statement is not among them; the close evidence pack rolls each fund forward and counts the period's transfers, and does not split the change into operations and transfers. Add the resale certificate, produced on demand against a statutory clock. The budget the board adopts is the one input among them that comes from outside the ledger, and the variance reports run against it.

Every one of the rest is a derivation from the same primary facts, and a ledger performs each derivation at the cost of a query, as many times as anyone asks.

Without one, the monthly treasurer's report is rebuilt by hand from the bank statement, and the year-end statement is rebuilt from the same statements and the spreadsheet by the accountant engaged to prepare it. Most of the list above is never produced. Each rebuild is an independent opportunity to get it wrong, which is the same distributed error as before, with a person doing the deriving instead of a model. And each rebuild produces nothing that was not already determined the moment the money moved.

That is parasitic load. The work is real, it is skilled, it recurs without end, and it returns no information the system did not already have. The free rail did not remove the cost. It moved the cost onto a volunteer and off the invoice.

The Surface Is Finite

Determinism is not a virtue that can be applied anywhere by wanting it enough. It is available where the surface is enumerable, and not otherwise.

An association's money can do a bounded number of things. In this system that enumeration is the 48 transaction types the posting engine accepts, routed through 10 flows and evaluated by 18 guards. Assessments. Receipts. Late fees. NSF fees. Vendor payments. Reserve transfers. Loan draws, payments, and accruals. Trust movements. Reversals. Period close. The list grows, slowly, by decision, and every addition is a line in a registry. The registry refuses in one direction. A guard the manifest names and the code does not supply stops every posting it would have covered. A type the registry does not list is not refused; it runs the universal chain, and the invariant guard passes it with a note that it did not recognize the type. Seven types ride that default today. Nothing posts unevaluated.

A finite named surface can be decided once per pattern, written into a guard, and applied by the same code on every occurrence. The marginal cost of the eight thousandth application is one pass through that chain and two rows in the decision log. Neither the model nor the spreadsheet keeps the decision, so each derives it again.

Where the surface is not enumerable — drafting, summarizing, conversing, reading an unfamiliar document for the first time — no such registry exists and a model is the right instrument. That boundary is the whole rule. Enumerate the surface. Where it is finite, decide once and keep the decision. Where it is not, a model earns its cost.

Above the ledger, a model has work, and more of it than a query interface or a summarizer. It can run the tasks: read the invoice, propose the account, schedule the payment, draft the delinquency notice, write the narrative section of the board packet. Everything it proposes goes through the same choke point as a person's entry and leaves the same decision record. That is why a probabilistic step is safe there and not underneath. Above the ledger its output is a proposal that the guards evaluate and the record names. Underneath, its output would be the record.

This system has one such instrument on the bill screen, and it is not a language model. When a bill is scanned, a random-forest classifier trained per association on its own corrected bills, with a keyword fallback, proposes an expense account with a confidence and a reason, and the form pre-selects it when the confidence is 0.6 or better. The proposal is a response to a request. The account on the bill line is whatever a person with bill-editing permission submits, and when the bill posts, the entry passes the same choke point as every other entry and the decision record names that person and the guards that evaluated it, not the proposal. The guards check that the entry balances, that the period is open, and that funds do not cross. They do not check which operating expense account was chosen. A wrong account inside a fund posts, and the record says who chose it. A wrong fund is refused.

The Invariant

This argument does not depend on how good models get.

Accuracy is not the binding constraint. A classifier at four nines still distributes its errors, one cause each, instead of concentrating them in one. Reproducibility is not the binding constraint either; it has been solved at the kernel level and the solution returns the output without the reason. The constraint is the fab's: the record has to walk back to the rule and the step that produced it, and survive someone asking, years later, by what method — and that requirement predates the technology by about five centuries and is indifferent to it. CommunityPay set that position down in February, before this essay, in Enterprise Correctness in the Age of Infinite Generation: correctness is a property of the output, governance is a property of the process, and a perfect model still leaves no trail of why. What is added here is the arithmetic, the evidence rule, and the review.

A system usually trades cost against quality. Spend more, get better. Spend less, accept worse. Engineers spend careers on that curve, and most architectural argument is about where to sit on it. Parasitic load names one axis: power drawn, no work returned. The arithmetic above supplies the other: the same draw returns a different answer some of the time. That is a design off the curve.

That is the position here. Deriving a determined decision again costs compute that scales with every transaction, and returns an answer that is right most of the time. Deriving it once, recording the decision under a hash, and applying it again costs a pass through the same code, returns the same answer every time, and carries the rule that produced it. Cheaper on one axis. Exact on the other.

And accounting is the rare domain where exact is the specification rather than the ambition. A wrong account on a balanced entry posts, reconciles, and closes, and the difference between the two books is whether every entry names the rule that placed it.

So the substrate is deterministic. Not because models are bad, and not as a position a better model would change. On the two axes that decide this — what it costs to run, and whether it is right — the deterministic path takes both, and in accounting taking both is the entire job.

The model sits above the ledger. It reads the record. It does not author it.

Sources

• Thinking Machines Lab (H. He et al.), "Defeating nondeterminism in LLM inference," September 10, 2025. https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
• L. J. Skitka, K. L. Mosier, and M. Burdick, "Does automation bias decision-making?" International Journal of Human-Computer Studies 51, no. 5 (1999): 991–1006; L. J. Skitka, K. L. Mosier, and M. Burdick, "Accountability and automation bias," International Journal of Human-Computer Studies 52, no. 4 (2000): 701–717.
• L. Rozenblit and F. Keil, "The misunderstood limits of folk science: an illusion of explanatory depth," Cognitive Science 26, no. 5 (2002): 521–562.
• N. H. Mackworth, "The breakdown of vigilance during prolonged visual search," Quarterly Journal of Experimental Psychology 1, no. 1 (1948): 6–21.
• R. Reber and N. Schwarz, "Effects of perceptual fluency on judgments of truth," Consciousness and Cognition 8, no. 3 (1999): 338–342; A. L. Alter and D. M. Oppenheimer, "Uniting the tribes of fluency to form a metacognitive nation," Personality and Social Psychology Review 13, no. 3 (2009): 219–235.
• M. Bar-Hillel, "On the subjective probability of compound events," Organizational Behavior and Human Performance 9, no. 3 (1973): 396–406; A. Tversky and D. Kahneman, "Judgment under uncertainty: heuristics and biases," Science 185, no. 4157 (1974): 1124–1131.
• Fed. R. Evid. 902(13), 902(11), 803(6); 12 CFR Part 1005 (Regulation E), § 1005.11(a)(1).
• Anthropic, Commercial Terms of Service, section D.3, effective June 17, 2025. https://www.anthropic.com/legal/commercial-terms
• Early Warning Services, LLC — Zelle ownership and bank-to-bank settlement; "Security," https://www.zelle.com/security (cannot cancel once enrolled; no protection program for authorized payments).