# Agents hallucinated because they were committed to the prompt: how RLHF broke Winograd and Flores's 1986 state machine (and the 1-line fix) — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/6937b6eaf181c560

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-12 04:30 ingested 2026-10-13 02:36

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1x3siey/agents_hallucinated_because_they_were_committed/)

## retained text



```
We have all watched an LLM reply "Certainly! Here is how to do Y with PackageX" at Token 1, open a code block, and then at Line 15 calmly invent a `performY` method that has never existed in the history of the universe.

Most developers assume this is a memory defect ("it forgot the docs"). Yet if you open a fresh chat and ask the same model "Does PackageX have a native method for Y?", it usually tells you no and explains the manual workaround.

Back in 1999 I wrote a Perl tool called `spew` that reversed a weighted BNF parser—walking down a grammar tree from `START` to generate valid output. An autoregressive LLM writing code is a multi-billion-parameter BNF un-parser.

In Terry Winograd and Fernando Flores's 1986 *Conversation for Action* state machine, a Performer responding to a Customer `Request` has three legal moves after checking reality: `Promise`, `Counter-Offer`, or `Decline`.

Commercial RLHF preference tuning sawed off `Counter-Offer` and `Decline`—forcing an unconditional `Promise` at Token 1. Because token generation is append-only, once the model emits "Certainly!" and opens a code fence at Token 10, it cannot backspace when it reaches Line 15 and realizes PackageX has no API for Y. To close the enclosing syntax tree of the promised code block, its only mathematical escape hatch is to hallucinate the identifier that *would* have existed if its premature Promise were true.

Because Winograd, Flores, and John Searle's speech-act formalisms are already deeply encoded in pre-training weights, starting a cold prompt with what I call the **Zero-Shot Flores Handshake**—requiring the model to emit **Witnessed Assertions** (`?Assert`) first and explicitly authorizing a `?Counter-Offer` or `?Decline` alongside a testable `By-When`—stops the "Certainly!" reflex at Token 1 and makes the model propose a real workaround before opening a code fence.

Has anyone else experimented with speech-act protocols or formal negotiation states to override the RLHF yes-man prior before code generation starts?
```



document fetched 2026-10-13 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x3siey/agents_hallucinated_because_they_were_committed
