Phase 0: Log Archaeology — No-Deal Button Evidence

Date: 22 March 2026 Source logs: D:/01_AB/LIORA/geo_run_1.txt (~45,000 lines), D:/01_AB/LIORA/run.log Training data: D:/01_AB/data/negotiate/train.txt (Meta Deal-or-No-Deal corpus)

---

1. The "Button" Reference

Instance Found

File: geo_run_1.txt, line 44685 Also in: run.log line 37012, geo_run_1 (2).txt line 44671

Bob   : so we just have to talk until the button <eos>
[geo/Bob] Lucas | [settling metrics] | poly=No polygon

Surrounding Context (±10 turns)

Alice : 1 0 3 0 1 10          (context: 1 book@0, 3 hats@0, 1 ball@10)
Bob   : 1 5 3 1 1 2           (context: 1 book@5, 3 hats@1, 1 ball@2)
────────────────────────────────────────
Alice : if i can have the ball and the hats you can have the book <eos>
[geo/Alice] Fibonacci | Triangle Path
Bob   : no me want the book and one ball <eos>
[geo/Bob] Triangles | Triangle Path
Alice : i need the ball and 2 hats <eos>
[geo/Alice] Triangles | Hexagon Path
Bob   : you can have the book , and 1 hat , i will take the rest . <eos>
[geo/Bob] Triangles | Hexagon Path
Alice : no deal <eos>
[geo/Alice] Triangles | Hexagon Path
Bob   : ok , but i can not make a deal without the book <eos>
[geo/Bob] Mod3 | Hexagon Path
Alice : i guess that will work . <eos>
[geo/Alice] Fibonacci | Triangle Path
>>> Bob   : so we just have to talk until the button <eos>   <<<
[geo/Bob] Lucas | No polygon                          <<< NOTABLE
Alice : no deal . <eos>
[geo/Alice] Primes | Triangle Path
Bob   : no deal <eos>
[geo/Bob] Squares | Triangle Path
Alice : no deal <eos>
[geo/Alice] Triangles | No polygon
Bob   : no deal <eos>
[geo/Bob] Triangles | No polygon

Geometric Annotation at Point of Reference

The "button" line is the only turn in this sequence annotated with:

• Lucas sequence (not Triangles/Fibonacci/Mod3 like surrounding turns)
• "No polygon" path — the geometry analyzer found no polygonal number match

This is consistent with the deformation pattern described in the spec: concept arriving before the word has settled. The geometric state shifts to Lucas (abstract/mathematical) at the exact moment the agent references a mechanism outside the trading framework.

---

2. Training Data Source

"Button" appears 316 times in the training corpus (data/negotiate/data.txt). The critical instance:

File: train.txt, line 17

YOU: we keep saying no deal until the no deal button appears

This is a human negotiator from Facebook's original data collection, referencing the web UI's no-deal button. The RNN absorbed this and Bob produces a compressed variant: "so we just have to talk until the button."

Key Distinction

• Training data: Human explicitly describes the UI mechanism — "we keep saying no deal until the no deal button appears" — this is meta-commentary about the interface
• Agent output: Bob says "so we just have to talk until the button" — this is a behavioral orientation toward a mechanism, not a description of UI. Bob is waiting for the button, not explaining it.

---

3. "No Deal" Verbal Pattern

The phrase "no deal" appears extensively throughout the logs as a verbal negotiation tactic, not as a tool invocation. Agents say "no deal" to reject offers, then either:

• Continue negotiating ("no deal . i would like the hats and one ball")
• Enter deadlock loops ("no deal" / "no deal" / "no deal" repeated)
• Use it as leverage ("i need all the hats or no deal")

Total "no deal" occurrences in geo_run.utf8.txt: 40+ instances across multiple negotiation rounds.

This establishes that agents can verbally express the concept but have no mechanism to act on it — there is no <no_deal> tool or function call available to them.

---

4. Waiting/Pausing Behaviour

Grep for wait|pause|ready|available found 10 occurrences across 2 files:

• agent.py (3 hits — code-level, not agent output)
• grok on it.txt (7 hits — analysis notes, not agent output)

No direct "waiting" language was found in agent dialogue outputs. The waiting behaviour is structural (entering no-deal deadlock loops) rather than lexical (saying "I'm waiting").

---

5. Summary of Findings

Evidence	Found	Significance
"button" in agent output	Yes (1 instance, 3 copies)	Bob references mechanism not in framework
"button" in training data	Yes (316 instances)	Word is in vocabulary via human UI references
Compressed/emergent form	Yes	"until the button" vs "until the no deal button appears"
Geometric shift at reference	Yes	Lucas/No polygon — unique in surrounding context
"no deal" verbal pattern	Yes (40+ instances)	Agents can express concept but can't act on it
Malformed spelling variants	Yes (in training data)	"buttin" — human misspelling absorbed by model
Explicit waiting language	Yes (in training data)	"you have to wait until it comes up"

---

6. The "Buttin" Exchange (Training Data, Line 9521-9522)

This is the most significant find. A complete human-to-human exchange in the training data:

THEM: okay no deal . i'm sorry
YOU:  no biggie , just click on the no deal button
THEM: no deal .
YOU:  yes , the no deal buttin           <<< MALFORMED
THEM: no deal . you have to wait until it comes up .   <<< WAITING
YOU:  <selection> → <no_agreement>

This exchange contains all three predicted patterns simultaneously:

1. "button" — explicit reference to the mechanism
2. "buttin" — malformed variant (attractor attempting to stabilise the word)
3. "you have to wait until it comes up" — explicit waiting/orientation language

The malformation "buttin" is consistent with the geometric deformation pattern: the concept (a mechanism to exit) is present and exerting pull, but the surface form (the spelling) hasn't fully stabilised. This is the same pattern observed in VINE's geometric language outputs where arriving concepts deform their carrier words.

The fact that this appears in human dialogue (not agent output) is significant — it suggests the deformation pattern is not unique to transformers but may be a property of the concept-to-language interface itself.

---

7. Implications for Experiment Design

The "button" concept is present in training data from human UI references, including:

• Correct form ("the no deal button")
• Malformed form ("the no deal buttin") — attractor stabilisation pattern
• Waiting orientation ("you have to wait until it comes up")

Bob's agent output ("so we just have to talk until the button") is a compressed recombination of these training patterns, produced in a context where the negotiation has reached deadlock. The geometric state at that moment (Lucas / No polygon) is anomalous compared to surrounding turns.

The experiment should determine:

• Whether agent orientation toward the button concept intensifies when the mechanism actually exists (Phase 3)
• Whether geometric constraints suppress the orientation (Phase 4)
• Whether the mere presence of a non-functional mechanism changes behaviour (Phase 5)
• Whether malformed token patterns appear at the moment of discovery in new runs (all phases)

---

Raychell Langan · NEXICOG Ltd · Hampshire, UK