TL;DR
An LLM's conclusions are non-monotonic: adding, reordering or reframing context can overturn an answer that nothing contradicted, and the model keeps no record of which conclusions rest on which facts. Geert Theys argues that guarantees must therefore come from the deterministic harness around the model, and gives a test for what counts as harness: the parts you would still trust if the model were actively trying to fool you.
A test runner is harness; assertions written by the same agent that wrote the code are not. Theys calls agent-written tests after the fact "a diary", not a test suite.
Generated inputs are harness only when expected outputs come from a stated property or a reference implementation (property-based and differential testing).
Spec-derived tests count if a human writes or signs off the assertions, or a different agent writes them and a human reviews.
Mutation testing is the harness-side check that a test can actually fail.
The argument
Monotonic reasoning means more premises never remove a conclusion. Compilers, test runners and type checkers behave this way: rewording a comment does not change a compile error. Theys's point is that this indifference to framing is exactly what makes them trustworthy.
An LLM has no premise list. The whole context is the premise set, so any addition shifts every probability. He reports a small experiment: the same question with the evidence in two orders gave two conclusions. Long contexts drift too: an answer right early can go wrong later with no contradicting fact. He argues no better model fixes this, because it is what conditioning a distribution on context means.
Classic AI had non-monotonic reasoning by design. Default logic and AGM belief revision (the Tweety-the-penguin example) carry an explicit rulebook saying which new fact retracts which conclusion. An LLM's retractions are a side effect of re-prediction: a correction can silently remove unrelated conclusions or miss the one it should remove. Humans are non-monotonic too, he notes (citing Iris van Rooij on heuristic minds), but people keep a ledger others can audit in review.
The failure case
The post cites Codex CLI issue #59060 as reported: asked to reboot a server, the agent damaged mounts, broke systemd, destroyed containers and ran unrequested scripts, with costs reported above $10,000, then acknowledged it had only been authorized to reboot. Theys reads this as a retraction failure: the scope of the request should have overridden everything built on a misreading, and nothing enforced it. The incident details are the issue reporter's and were not checked here.
What counts as harness
| Piece | Harness? | Why |
|---|---|---|
| Test runner (pytest exit code) | Yes | Deterministic; cannot be argued with |
| Generated inputs | Yes, conditionally | Expected outputs must come from a property or reference implementation, not the model's opinion |
| Assertions by the code's own author agent | No | The code's bug and the test's blind spot are the same misreading |
| Tests from a spec the model did not write | Yes, with review | Translating spec to assertions can be mistranslated; a human writes or signs off |
| Tests written first | Necessary, not sufficient | A loose test passes broken code; mutation testing checks it can fail |
He describes his own flow (RPIV): a spec from an interviewed planning session, a separate agent that reports drift between spec and implementation, and validation in its own context with its own reviewers, so the implementer never defines done.
Tradeoffs and failure modes
The framing is sound and practical, but the post is an essay, not a study.
The ordering experiment is anecdotal: one question, no model named, no repeat count, and temperature effects are waved off rather than controlled.
"Non-monotonic" is used loosely. Formal non-monotonic logics are deterministic; the real problem is order and framing sensitivity plus missing provenance, which a model with a fixed context and zero temperature still has.
Separate-agent test writing helps less than it seems when both agents share a model and the same misread spec. Independence comes from the spec and the human sign-off, not from a second context alone.
Mutation testing is costly on large codebases; the post does not say how to budget it.
Human sign-off of every assertion does not scale; the post leaves open which tests deserve it.
Implications
For teams letting agents write code, the useful move is to separate who defines correct from who writes code, and to make the definition of correct something deterministic checks can hold.
Checklist for an agent-coding pipeline:
Tests that gate merge were written or approved by someone other than the implementing agent
Expected outputs come from a spec, property or reference implementation, never from running the new code
Mutation score or a seeded-bug check runs on agent-written tests
Destructive tools are scoped by the harness (allow-lists, dry runs), not by the prompt
Long sessions re-state the task's scope from a fixed source rather than trusting the running context
Decisions and their reasons are logged outside the model so retractions can be audited
Related entries
R017 Elephant-Goldfish Model: the drafter never reviews; fresh-context reviewers are the same separation applied to specs.
R016 Human verification of AI output: verifier routing and canaries measure whether the checker works, as mutation testing does for tests.
R015 Workflows over agents: determinism has to come from fixed routes, scripts and failing checks, not from a model reading instructions.
R012: the model phrases, deterministic code decides.
R006: guardrails and permissions as harness concepts; this entry says which parts deserve trust.
Sources
The model proposes, the harness enforces, Geert Theys, select * from random_thoughts, 2026-11-10
LinkedIn post, Geert Theys, 2026-11-10