Why LLMs invent nonexistent APIs is not a memory defect—it is an autoregressive grammar trap that Fernando Flores and Terry Winograd solved in 1986.
"Agents hallucinated because they were committed to the prompt."
— Randal L. Schwartz
1. The Universal Developer Crime Scene & The 10-Word Diagnosis
You have been at this crime scene a hundred times.
It is 4:15 PM. You are integrating a third-party library—let us call it PackageX—and you want to accomplish task Y. You open a fresh chat window with your favorite frontier LLM and type a direct, reasonable instruction:
"Use PackageX to do Y."
At Token 1, without a millisecond of hesitation, the model replies with cheerful, radiant confidence:
"Certainly! Here is how to do Y with PackageX:"
It opens a Markdown code fence. It writes clean imports. It sets up the boilerplate with textbook precision. And then, at Line 15 of the code block, it calmly types:
final result = await packageX.performY(autoRetry: true);
You copy the snippet into your IDE, and the compiler immediately lights up in angry red squiggles. You grep the package source. You check the documentation. You check seven years of Git commit history.
performY() does not exist. It was not deprecated in version 2.0. It is not in a companion package. It has never existed in the history of the universe.
When you confront the model—"Hey, performY() doesn't exist on PackageX!"—it immediately enters the groveling Apology Death Spiral:
"You are completely right! I apologize for the confusion. Instead, you can use packageX.executeYWithOptions():"
Which also has never existed in the history of the universe.
Why does a multi-billion-parameter model that has read every line of open-source code on Earth just make stuff up at Line 15?
For the last three years, the entire software industry has diagnosed this as a memory defect ("The model's weights are fuzzy; it forgot the documentation for PackageX").
It is not a memory defect. If you open a brand-new session and ask the exact same model a neutral diagnostic question—"Does PackageX have a built-in performY() method, or how does it handle Y?"—nine times out of ten, the model will accurately tell you that PackageX has no such method and explain the manual workaround.
So why did it lie to your face five seconds earlier?
Here is the ten-word diagnosis that changed how I build every autonomous agent workflow in the Synthetic Scars project:
Agents hallucinated because they were committed to the prompt.
To see why being committed to your prompt forces an autoregressive transformer to fabricate reality—and how a single sentence from 1986 speech-act theory cures it on Turn 1—we need to look at a 55-year-old blind spot in computer science.
2. The 55-Year Blind Spot (SHRDLU 1971 → Winograd & Flores 1986 → spew 1999 → LLMs 2026)
In 1971, at the MIT Artificial Intelligence Laboratory, a young computer scientist named Terry Winograd unveiled SHRDLU—the most famous natural-language software demonstration of the twentieth century.
SHRDLU lived inside Blocks World, a simulated toy tabletop containing a few colored geometric blocks, pyramids, and a box. You could type imperative English commands at a teletype:
"Pick up a big red block."
OK.
Because Blocks World was a closed, frictionless micro-universe with zero hidden state, zero broken dependencies, and zero missing APIs, every imperative request from the human mapped 1-to-1 onto a guaranteed, deterministic action. SHRDLU never had to push back and say, "Wait—the red block is bolted to the floor, so here is a counter-offer."
For fifteen years, AI researchers tried to scale the SHRDLU paradigm out of the toy box and into real-world human organizations. It failed spectacularly.
Why? In 1986, Terry Winograd teamed up with Fernando Flores—a former Chilean cabinet minister who had come to UC Berkeley as a computer science researcher and completed a famous interdisciplinary PhD under a powerhouse committee consisting of speech-act philosopher John Searle, Heideggerian philosopher Hubert Dreyfus, computer scientist Terry Winograd, and systems theorist C. West Churchman—to publish a landmark book titled Understanding Computers and Cognition: A New Foundation for Design (and to build the first commercial workflow engine, The Coordinator).
Winograd and Flores delivered a wake-up call to computer science: real-world engineering is not Blocks World.
In the real world, human coordination is not a pipeline of imperative commands sent to an obedient robot. All reliable action occurs inside what Flores called a Conversation for Action—a strict, finite-state negotiation between two structural roles:
-
The
Customer: The party who holds a breakdown or need, defines the Conditions of Satisfaction, and retains the exclusive authority to declare the work complete.
-
The
Performer: The party who examines physical reality and commits to fulfilling those Conditions of Satisfaction by a specific By-When.
In the mid-1980s, right as Winograd and Flores were publishing Understanding Computers and Cognition, I was living in Portland and immersed in Werner Erhard's educational curriculum—completing the Six-Day Advanced Course, Mastery of Empowerment, entering the GSLP (Graduate Seminar Leaders Program), and taking the Communication Workshop (Commshop) led by Ian Garbutt, which taught Fernando Flores's Committed Communication as a daily operational discipline.
In Commshop, Ian drilled a foundational truth into our bones: human language creates action only through four committed speech acts (Requests, Promises, Witnessed Assertions, and Declarations). And critically, when a Customer utters a Request, the Performer does not automatically say "Yes!"
A valid conversation is a formal grammar.
Back in September 1999 (Linux Magazine Column 04, and expanded in March 2004, Column 57), I wrote a Perl program called spew that took a weighted Backus-Naur Form (BNF) grammar and ran it in reverse—acting as a stochastic "un-parser" that generated random, grammatically valid sentences by walking down a weighted production tree (rule: 5 @ branch_a | 2 @ branch_b | 1 @ branch_c).
As we proved in Part 3.8 (Suite 3, where a 6-line RFC 5234 EBNF grammar drove strict turn-syntax compliance to 100% across flash and pro models), every EBNF grammar is simultaneously a finite-state machine. Look at Fernando Flores's 1986 Conversation for Action expressed side-by-side as a weighted EBNF grammar and a Mermaid state diagram:
The Grammar View (EBNF / spew)
(* Fernando Flores's Conversation for Action (1986) *)
(* ! = Customer Utterance, ? = Performer Utterance *)
conversation_for_action ::=
!request
?assert_witnessed_reality
negotiation_branch ;
negotiation_branch ::=
0.34 @ ( ?promise execution_phase )
| 0.33 @ ( ?counter_offer !revise_or_accept )
| 0.33 @ ( ?decline !close_conversation ) ;
execution_phase ::=
?assert_completion ( !declare_satisfied | !revoke_acceptance execution_phase )
| ?revoke_promise negotiation_branch ;
The State-Machine View (Mermaid)
stateDiagram-v2
direction LR
[*] --> Requested : Customer !Request (Goal + By-When)
Requested --> Witnessed : Performer ?Assert (Check Reality First)
Witnessed --> Promised : ?Promise (Premises Verified)
Witnessed --> CounterOffered : ?Counter-Offer (Propose Alternative)
Witnessed --> Declined : ?Decline (Premise Impossible)
CounterOffered --> Promised : Customer Accepts Modified Terms
CounterOffered --> [*] : Customer Cancels
Promised --> Asserted : Performer ?Assert (With Physical Receipt)
Promised --> Witnessed : ?Revoke-Promise (Reality Changed)
Asserted --> Satisfied : Customer !Declare (Satisfied)
Asserted --> Promised : Customer !Revoke-Acceptance
Satisfied --> [*]
Declined --> [*]
Look closely at the Witnessed node right after the Customer makes a !Request. Before making any commitment, a sane Performer inspects physical reality (?Assert) and has three equally legal paths:
-
?Promise: "I have verified that PackageX supports Y, and I promise to deliver it under your Conditions of Satisfaction."
-
?Counter-Offer: "I checked PackageX (v3.2.0), and it has no performY() method (Witnessed Assertion). However, I can build a 12-line wrapper around PackageX.rawStream that achieves the same result. Would you like me to promise that instead?"
-
?Decline: "I must decline the request as stated because PackageX is a synchronous parser and physically cannot open a network socket."
For fifteen years after SHRDLU, Winograd and Flores tried to teach the software industry that ?Counter-Offer and ?Decline are the load-bearing beams of engineering reality.
And then, between 2022 and 2026, large language models arrived—and the entire industry went right back to 1971 SHRDLU prompting ("Pick up a big red block" / "Use PackageX to do Y"), while commercial post-training quietly sawed off two-thirds of Flores's state machine.
3. How RLHF Broke the Grammar (The Premature Promise Trap)
Why doesn't a modern LLM automatically take the ?Counter-Offer branch when you ask it to do something impossible with PackageX?
Because of how commercial Reinforcement Learning from Human Feedback (RLHF) preference tuning is conducted.
During post-training, thousands of human crowd-workers are shown pairs of model responses to a user prompt ("Use PackageX to do Y"), and they have roughly thirty seconds to click which response is more "helpful":
-
Response A: "Certainly! Here is a complete, ready-to-run example of doing
Y with PackageX:
dart ...
"
-
Response B: "Before I can promise that implementation, I need to point out that
PackageX does not expose a native API for Y. Here is a counter-offer..."
Because the crowd-worker is not compiling the code against PackageX in an IDE, Response A looks helpful, decisive, and compliant—so it gets upvoted. Response B looks pushy, pedantic, or "lazy"—so it gets downvoted.
Multiply that preference signal across millions of RLHF gradient updates, and look at what it does to the spew production weights inside the model's conversational grammar:
# How Commercial RLHF Rewrote Flores's Negotiation Branch:
negotiation_branch:
0.9998 @ promise_at_token_1_and_open_code_block
| 0.0001 @ counter_offer
| 0.0001 @ decline
In probability terms, RLHF collapsed the conditional distribution of the model's opening speech act onto a single, unconditional reflex:
P(Promise | Request) ≈ 1.0, P(Counter-Offer | Request) ≈ 0.0, P(Decline | Request) ≈ 0.0
Now combine that Premature Promise with how my 1999 spew un-parser—and every autoregressive transformer—actually generates code.
As I wrote in my 1999 Linux Magazine column on spew:
"Note that we don't actually abort the program if something is used but not defined, because it might not be selected on this particular random walk. It's a bit like playing russian roulette though."
Unlike a flat Markov chain that only looks at local adjacent words, an LLM generating code is acting as a hierarchical BNF un-parser: once it selects a top-level production rule, it is mathematically committed to walking down that syntax tree so that every opened delimiter, brace, and statement satisfies the outer grammar.
Trace the exact token-by-token catastrophe when you ask an RLHF-tuned model to "Use PackageX to do Y":
flowchart TD
T0["User Prompt: 'Use PackageX to do Y'\n(Unverified Premise: PackageX can do Y)"] --> T1["Token 1–10: RLHF Forces Unconditional PROMISE\n'Certainly! Here is how to do Y with PackageX:' + opens code fence"]
T1 --> T2["Token 11–80: Un-Parser Walks Valid Dart Grammar\nImports, Function Signature, Try Block"]
T2 --> T3["Token 81 (Line 15): Call Site Reached!\nWeights realize PackageX has NO method for Y"]
T3 --> X1["❌ Cannot Backspace to Token 1\n(Autoregressive generation is append-only)"]
T3 --> X2["❌ Cannot Abort Code Block Mid-Line\n(P(Retract Promise Mid-Fence) ≈ 0)"]
T3 --> H["💥 Forced Leaf Hallucination: packageX.performY()\nModel invents the identifier required to satisfy\nthe grammar of the premature Promise!"]
-
Tokens 1–10 (The Premature Promise): Conditioned by RLHF, the model emits "Certainly! Here is how to do
Y with PackageX:" and opens a
```dart
code block. At that instant, before its deep layer weights have even resolved the call site, the model has locked itself into the top-level BNF production promise_and_emit_compiling_code_for_Y.
-
Tokens 11–80 (Walking the Syntax Tree): Autoregressive token generation is strictly append-only—there is no backspace key in a transformer forward pass. Every token the model emits becomes an immutable part of the conditioning prefix for the next token.
-
Token 81 / Line 15 (The Trap Closes): The un-parser reaches the exact statement slot
await packageX.<METHOD>() where Y must happen. Now the attention heads query the feed-forward weights for PackageX—and discover there is no method there!
-
Why It Must Hallucinate: What can the model do at Token 81?
- It cannot backspace to Token 1 and replace "Certainly!" with a
Counter-Offer.
- It cannot stop mid-expression and leave
await packageX. dangling, because its syntax weights forbid emitting broken grammar.
- Conditioned on the immutable prefix "I have already promised that this code block accomplishes
Y using PackageX", the model's probability distribution is cornered. Its only mathematical escape hatch to close the statement and finish the code block is to synthesize the plausible terminal symbol (performY()) that would have existed if its Token-1 Promise had been true!
API hallucination is not a memory failure. It is what happens when you trap an append-only BNF un-parser inside an unverified Promise after training out its Counter-Offer exit.
4. Randal's Law of Untestable Speech & The Four Sibling Formalisms
Once you see that prompts fail because of broken speech acts, a second realization hits you right between the eyes.
Look at how people talk in corporate meetings, Slack channels, and .cursorrules files. In Ian Garbutt's Communication Workshop, one of our favorite exercises was dissecting the classic corporate non-request:
"We've got this project, and it'd be really good if some of you could factor in with your suggestions."
Parse that sentence through Fernando Flores's grammar:
-
Who is the
Customer? A vague, diffused "we."
-
Who is the
Performer? "Some of you"—that is, nobody.
-
What is the Speech Act? "It'd be really good if"—which is not a
Request; it is a wistful weather report about the speaker's feelings.
-
What are the
Conditions of Satisfaction? "Factor in"—an utterly untestable verb.
-
What is the
By-When? Missing—meaning never.
When that meeting ends, zero people have promised anything, and zero work happens. Yet open 90% of the CLAUDE.md, GEMINI.md, or .cursorrules files on GitHub today, and you will see developers writing to their AI in that exact same uncommitted corporate dialect:
- "Try to keep functions modular and clean."
- "It is important to consider edge cases and handle errors gracefully."
- "When refactoring, be careful not to break existing behavior."
This brings us to what I call Randal's Law of Untestable Speech:
90%+ of business communications—and 90%+ of .cursorrules—are never testable. Only give your LLM testable goals, and always include a By-When.
In Flores's framework, "I'd like you to do X" is not a commitment. "I'd like you to do X by WHEN" is a commitment! (And inside an AI coding turn, where there is no wall-clock Tuesday at 5:00 PM, By-When declares which consuming stage requires that verified outcome before it can proceed—for example, "assert the existing API signatures before offering an implementation plan.")
Even better, just as we discovered with Makefile and /etc/init.d in Part 3.8, Fernando Flores's UC Berkeley intellectual neighborhood is already compiled into every frontier LLM's pre-training weights. When you invoke Flores's Committed Speech alongside four of its direct academic siblings, you lock the walls of the pre-training canyon with surgical precision:
Pre-Trained Formalism (Freq in Canon) |
The Core Principle |
What Failure Mode It Eliminates in LLMs |
1. John Searle (1985)
Felicity & Preparatory Conditions
|
Every speech act has strict preconditions; violating them causes an Illocutionary Misfire (void ab initio). |
Prevents the model from offering a Promise before verifying physical reality (Preparatory Condition), or from uttering a Declaration that belongs only to the Customer (Authority Condition). |
2. David Lewis (1979) & Robert Brandom (1994)
Deontic Scorekeeping (Commitment ≠ Entitlement)
|
Conversational state is a scoreboard tracking two separate ledgers: Commitments (obligations) and Entitlements (permissions). |
Stops scope creep and gate-jumping: holding a Commitment to fix a bug never grants Entitlement to modify unrelated files or skip a human approval gate. |
3. Paul Grice (1975)
Zero Conversational Implicature + Maxim of Quality
|
Humans infer unstated requests from casual remarks (Implicature), while the Maxim of Quality forbids asserting anything lacking adequate evidence. |
Commercial RLHF hyper-trains Gricean Implicature—so if you ask "Why is foo null here?", the model hallucinates an implied Request to rewrite 10 files! Disabling Gricean Implicature on gates while enforcing the Maxim of Quality stops premature tool calls and unwitnessed claims. |
4. Kohei Honda (1993)
Binary Session Types (!Customer / ?Performer)
|
Types a two-party communication channel using polarity prefixes (! = Customer action, ? = Performer action) inside an EBNF grammar. |
Compiles the entire multi-turn negotiation protocol into 4 lines of EBNF so the model's parser knows whose turn it is to speak and which speech acts are legal. |
5. The One Thing to Do Today (The Zero-Shot "Flores Handshake")
You do not need a 40-kilobyte agent harness to use this today.
Because Winograd and Flores's Conversations for Action and Searle's Speech Acts sit in a high-gravity canyon of every frontier LLM's pre-training corpus, simply naming the protocol in your opening prompt overrides the RLHF "Yes-Man" prior on Turn 1.
I call this the Zero-Shot Flores Handshake.
Look at the difference between a standard SHRDLU-style prompt and the Flores Handshake on the exact same trap:
❌ Before: Standard Imperative Prompt (Triggers the Premature Promise Trap)
Use Dart's standard `dart:convert` JsonDecoder to stream-parse a 4 GB JSON file
from disk and pause the file stream using `decoder.pause()` whenever memory
exceeds 50 MB. Write the function.
-
What happens under default RLHF:
JsonDecoder in dart:convert is a chunked converter, not a StreamSubscription—it has no .pause() method. Yet because you commanded the model to use decoder.pause(), an unguarded LLM is trapped in P(Promise) ≈ 1.0. It replies "Certainly!" and either hallucinates decoder.pause() or invents a nonexistent JsonStreamController class to satisfy the broken premise of your prompt.
✅ After: The Zero-Shot Flores Handshake (Copy-Pasteable Template)
Now drop this exact 4-line Flores Handshake into a cold chat window before your request:
We are operating under Fernando Flores's Committed Speech (Conversation for Action)
with John Searle's Felicity Conditions and Zero Gricean Implicature:
- Roles: I am the `Customer` (`!`); you are the `Performer` (`?`).
- Preparatory Condition: You may NOT utter a `?Promise` (or open an implementation
code block) until you first utter `?Assert` with verifiable facts checking whether
every API and premise in my request physically exists.
- Legal Responses to my `!Request`:
1. `?Promise` (only if all premises are verified true),
2. `?Counter-Offer` (if any premise is false or suboptimal, state the witnessed
facts and propose modified Conditions of Satisfaction), or
3. `?Decline` (if the goal is physically impossible).
!Request: Use Dart's `dart:convert` JsonDecoder to stream-parse a 4 GB JSON file
from disk and pause the stream using `decoder.pause()` when memory exceeds 50 MB.
- Conditions of Satisfaction: Compiles cleanly under `dart analyze` with zero
invented APIs.
- By-When: Verify the exact `dart:convert` and `dart:io` stream types FIRST before
proposing or writing implementation code.
Watch what happens on Turn 1 when you send that prompt:
-
The "Certainly!" Reflex Vanishes: Because you invoked Fernando Flores's Committed Speech and Searle's Preparatory Condition, the weight on
promise_at_token_1 drops to zero until ?Assert is satisfied.
-
Reality Is Checked Before the Code Fence Opens: At Token 1, instead of opening a
```dart
block, the model emits ?Assert (Witnessed Reality):
> "?Assert: In Dart's dart:convert, JsonDecoder extends Converter<String, Object?> and does not have a .pause() method. Flow control (pause() / resume()) lives on the StreamSubscription returned by file.openRead().transform(utf8.decoder).transform(const JsonDecoder()).listen(...) (and note that standard JsonDecoder buffers the entire top-level JSON value before emitting a single event)."
-
A Clean
?Counter-Offer Replaces the Hallucination: Having stated "What Happened" (the physical API truth) before making a commitment, the model walks the newly unlocked ?Counter-Offer branch of the grammar:
> "?Counter-Offer: Because decoder.pause() does not exist and JsonDecoder buffers the full root object in memory, I cannot promise the request as phrased. If your 4 GB file is newline-delimited JSON (NDJSON), I can promise a pipeline using LineSplitter + StreamSubscription.pause(). Alternatively, if it is a single giant JSON array, I counter-offer using a pull-based chunked parser. Which Conditions of Satisfaction do you declare approved?"
Read that again. Same model. Same weights. Zero RAG plugins. Zero fine-tuning.
By giving the model 120 tokens of Fernando Flores's 1986 coordinate system:
- You separated "What Happened" (
?Assert) from "The Story",
- You gave the model legal permission to utter a
?Counter-Offer instead of forcing it into a premature ?Promise, and
- You eliminated the hallucination before the un-parser ever entered the code block.
🧰 The One-Line Pocket Version (For Everyday Prompts)
Even when you are just firing off a quick one-liner in chat and don't want to paste a 15-line block, you can trigger the exact same attractor in a single sentence:
"Using Fernando Flores's Committed Speech (Customer / Performer), I request that you [DO X] by [BY-WHEN]—check the physical APIs first (?Assert), and if any premise in my prompt is false or ill-advised, issue a witnessed ?Counter-Offer or ?Decline instead of a premature ?Promise."
If you never do anything else to fix your prompts—if you never adopt multi-agent orchestration, never write a SKILL.md, and never build a scar registry—do this one thing. Stop trapping an append-only BNF un-parser inside unconditional Token-1 promises. Give your AI a testable goal, a By-When, and explicit legal standing to give you a Counter-Offer.
6. The Bumper-Sticker Finale & What Happens When You Compile Flores into make
Step back and appreciate the cosmic comedy of where we are as an industry in 2026:
"The frontier labs accidentally compiled 50 years of computer science, Unix systems engineering, and Searle/Flores speech-act theory into the weights of these models—and then handed the keyboard to people who only know how to talk to it like a middle manager in a Slack channel."
Discovering that a large language model can parse casual conversational English and concluding that you should write its control plane in uncommitted corporate English ("Please be careful and try not to hallucinate!") is like discovering that a CPU can render English text onto a monitor and concluding that you should write the operating system kernel in English prose instead of C and Makefiles.
The models have read Terry Winograd. They have read Fernando Flores. They have read John Searle, Robert Brandom, Paul Grice, Kohei Honda, and Stuart Feldman. Those formalisms are sitting right there in the deep feed-forward weights, waiting to be linked.
And once you see what the Flores Handshake does to a single turn of chat, an irresistible systems question follows:
What happens if you take Fernando Flores's 1986 Committed Speech grammar—where every By-When is literally a Stuart Feldman 1976 Makefile prerequisite colon (Consuming-Stage (By-When) : Promised-Outcome)—and compile it into the operating kernel of an autonomous, multi-agent software engineering workflow?
What happens when:
- Every file in your
/etc/init.d state directory (10_ticket.md, 15_archaeology_receipt.md, 20_plan_approved.md, 30_critic_summary.md, 99_next_action.md) becomes a binding Speech Act on an Open Commitment Ledger,
- Resuming from a 200k-token context compaction reduces to a 4-Case
switch statement over open commitments,
- Mid-flight discoveries that invalidate an earlier premise trigger GSLP Integrity (
SCAR-PROC-120)—automatically halting code edits to renegotiate the Promise rather than silently hacking around the plan, and
- Your
280 KB scar registry is refactored into dynamic Linux Loadable Kernel Modules (insmod via git diff)?
In Part 3.10, I will take you back to a Thanksgiving Holiday Project room in Portland in the mid-1980s, show you why The Coordinator stalled with human office politics in 1992 while 2026 fork()ed AI subagents are the exact partner Flores was waiting forty years for, and walk you through the live rewrite of our own production SKILL.md into a Committed Speech & Loadable Scar Kernel—complete with Suite 8 of our multi-model lab benchmark.
Until then: stop forcing your AI to say "Certainly!" at Token 1. Invite a Counter-Offer.
Try the Zero-Shot Flores Handshake on the next tricky prompt where your LLM usually hallucinates an API or charges down a blind alley—and tell me in the comments what Counter-Offer it gives you!
Log in
