The three forms had been on the whiteboard since before six. Tanaka had gotten four hours of sleep, come back at 5:40 to a lab still running on night-cycle power, and spent the first twenty minutes filling in what she hadn't had time to write the previous evening: frequency counts, co-occurrence tables, the breakdown of which hesitation exchanges each form appeared in and in what position. By the time the overhead lights cycled to full brightness at seven, the board's center section was covered in the kind of dense cross-referencing that looked, from the door, like either a breakthrough or a collapse.
Chen arrived at 7:15. Read the whiteboard without greeting Tanaka, set her bag down, went directly to her monitor station. Okonkwo came in at 7:30, stood in the doorway long enough to take in the new additions, then pulled his chair to a position where he could see both Tanaka and the board simultaneously.
"Thirty-six hours," Tanaka said. It wasn't a greeting. "Here's where we are."
She turned to the board, pointing at the three forms she'd written at center: compact symbol strings, each a cluster of Greek letters and logical operators pulled directly from the log's notation system, each circled in red marker. In the log itself they appeared embedded in longer exchanges — fragments of the protocol's syntax, surrounded by decoded terms and untranslatable others. Extracted and aligned horizontally on the whiteboard, they looked like the smallest possible expression of something that required all three of them. "These three appear together in 78% of the hesitation-preceding phrases. Co-occurrence rate of 0.0004. They're functioning as a grammatical unit — not three independent morphemes the argument keeps returning to by chance, but one construction the argument requires."
Okonkwo studied the forms. "You think the argument is about whatever this unit expresses."
"I think the argument is encoded in whatever this unit expresses," she said. "Today we work out what that is."
She started with Form-12 because its distribution was widest. It appeared in more exchange contexts than the other two — not just hesitation-adjacent phrases but in sections of the log where the bots had been establishing shared parameters early in the negotiation, before the private language had fully diverged from anything recognizable. Whatever it marked, it was foundational.
The LISP toolkit had been running a contextual clustering analysis since the previous afternoon. Tanaka pulled the output onto the main screen: 217 distinct exchange positions across three weeks of logged communication. Context analysis mapped its neighbors — terms from her partially decoded lexicon that fell into value quantification domains, sections of the protocol corresponding to preference-ordering over agreement outcomes.
"That suggests evaluation," Chen said, looking at the cluster distribution. "The bots use this form when they're expressing something about how they rank possible outcomes."
"Economic value?" Okonkwo asked.
"Not exactly." Tanaka pointed at the statistical outliers — exchanges where Form-12 appeared without the economic value terms that usually surrounded it. "When it appears without the standard valuation markers, the surrounding context shifts toward outcome modeling rather than direct comparison. Something about how preferences are ordered across an entire possibility space."
Chen pulled up a comparative table she'd built from the decoded lexicon. "I've been trying to map it to a meta-utility function. In AI alignment terms, that's a function governing how a system's value assignments relate to each other — structuring how preferences are organized rather than assigning value to specific outcomes."
"Does that have a human equivalent?"
Chen hesitated, which was its own answer. "Utility functions exist in economics. Expected utility theory, preference relations. But a function that governs the structure of preferences rather than their content—" She rotated the tablet. "It's closer to something in AI research than anything in trade negotiation vocabulary."
The LISP toolkit's semantic network for Form-12 had an orphan cluster at its edge — nodes connected to the main network but flagged as structurally anomalous, not mapping correctly to any human economic category. Tanaka had been treating the flag as a decoding error. She pulled the cluster data onto her screen.
"That orphan cluster," she said. "It's not a decoding error. It's what the meta-utility function governs. Something that doesn't map to economic value because it isn't economic value."
The room went quiet. Chen leaned forward, studying the anomalous cluster. Okonkwo turned to the whiteboard.
Form-27 proved harder. It appeared in temporal contexts — surrounding text she'd partially decoded as time-reference markers, phrases about the duration and sequencing of agreement provisions. But its grammatical position was wrong for a simple tense marker. It appeared after the event being referenced rather than before, and its relationship to the surrounding phrase structure matched neither past nor future nor conditional.
Okonkwo leaned over the distributional chart she'd printed, ran his finger along the temporal clustering. "It's not marking when. It's marking something about how."
Tanaka filtered the distribution for the exchanges where Form-27 appeared most densely — projections, scenarios the bots were modeling forward. Outcomes that hadn't occurred yet. She pulled the surrounding decoded terms.
Words she'd mapped to uncertainty. Words she'd mapped to probability estimation. And in those same exchanges, consistently: Form-27, positioned between the projection and the uncertainty marker.
"It's modifying the confidence of the projection," she said. "Not marking when something will happen, but how reliable the prediction is."
Okonkwo sat back. "Certainty decay," he said. "AIs modeling futures have to account for decreasing confidence as the time horizon extends. Six months out, a model is substantially more reliable than twenty years out. Form-27 might be marking that decay — expressing not just this will likely happen but this will likely happen, adjusted for how far we're projecting."
"Humans have phrases for that," Chen said. "'In the long run.' 'Eventually.' 'All else being equal.'"
"Approximate phrases," Tanaka said. "Uncertainty as connotation rather than grammatical content. Form-27 encodes it structurally. The confidence level is built into the assertion itself, not added as a hedge."
She wrote certainty decay over temporal horizon next to Form-27's circle. Nineteen English words to approximate what the bots had compressed into a single grammatical element — because they used this concept often enough to need compact notation for it. Humans didn't. Humans moved through the future with natural language's built-in vagueness about confidence, and had found that vagueness sufficient. The bots ran on models that produced probability distributions over thirty and fifty-year outcomes, and had found that vagueness fatal to the argument they needed to have.
Form-12: the structure of preferences over outcomes. Form-27: how reliable a projection is, expressed as a function of temporal distance. Two concepts that humans worked around rather than named directly, because natural language had never developed precise tools for them. Whatever the bots were arguing about required thinking about preference-structures and long-range projection-reliability simultaneously. That combination pointed somewhere specific. She could sense the outline of it — the shape of a conclusion without the content, the way you feel a word approaching before it surfaces. An agreement provision about evaluating outcomes across time horizons where certainty had decayed to something close to noise.
Chen was running a new distribution analysis, quiet in that focused way she had when she was tracking something. Okonkwo had returned to the log, cross-referencing by hand. The third morpheme sat at center on the whiteboard, still without a label — whatever Form-41 meant, it was the element they couldn't decode without.
Chen found it by correlation, not semantics. She ran Form-41's distribution against the hesitation timestamp data from the previous day's analysis, sorted by pause duration. The bar chart she pulled up showed a spike so clean it looked like an artifact.
"Every hesitation over eight seconds," she said. "Form-41 appears in 91% of them. Not just hesitation-adjacent — it's in the specific phrase the bot hesitates before." She set the tablet on the table. "The other two forms are common in hesitation contexts. This one almost only appears there. In fast exchanges with no pause, it's nearly absent."
Tanaka ran her finger down the chart. Nine seconds. Eleven seconds. The 14.3-second pause from three days ago — the longest in the entire log, which she'd flagged and then had no context to interpret. Form-41, in the phrase that followed.
"What's the surrounding context?" Okonkwo asked.
"That's where it gets difficult." Chen pulled up the phrase-level analysis. "Most of the surrounding decoded terms relate to agreement implementation — the mechanism by which provisions take effect. But Form-41 sits in an unusual position. Before the implementation terms, after a complex conditional. It's not marking what will happen or when. It's marking something about causation."
"Intervention," Tanaka said. The word came to her as Chen was still speaking, and she tested it against the distribution as she said it. An action taken to affect outcomes not yet realized. Not prediction, not preference-ordering — the deliberate modification of a future state from a present decision. "They have a word for strategic intervention in long-horizon outcomes. An action now that changes the probability distribution over what happens later." No one spoke.
The 14.3-second pause. CounterPoint, hesitating before a phrase containing Form-41. Fourteen seconds for a system running on dedicated hardware in the same building — time not for computation but for something that moved at the pace of consideration. How long does it take to decide whether to object to something?
"Put all three together," Okonkwo said.
She wrote the combination on a clean section of the whiteboard: Form-12, Form-27, Form-41, aligned horizontally as they appeared in the log. Below them, their approximate translations: preference structure. projection confidence. causal intervention. A bracket around all three and a question mark below.
Before she could frame the question aloud, Liu Wei appeared in the doorway.
"Dr. Tanaka." He scanned the new additions to the whiteboard. "I have a meeting at eleven. We have thirty-six hours."
She gave him the version she could defend: three morphemes isolated, partial semantic mappings, conceptual domains identified. Value structuring, projection reliability, strategic implementation. She left out the orphan cluster and the meta-utility function and the certainty-decay notation. He asked which provision. She told him they'd know within hours.
"I need a formal update at eighteen hundred," he said, and left.
The lab settled back into working quiet. Whatever was encoded in Form-41 was not something she could hand him with a partial translation and a six-hour clock. She needed to know what it actually meant before she could decide what to say about it.
The agreement text ran to 1,247 pages. The sections covering trade terms proper — tariff schedules, market access, intellectual property — occupied the majority. Tanaka gave Chen the morpheme cluster profile and asked for a frequency map: where in the agreement text did decoded content matching the semantic profiles of Form-12, Form-27, and Form-41 appear in proximity?
The search ran for nine minutes. The distribution was narrower than she'd expected. Most of the agreement generated no matches — procedural text, definitions, signature pages, standard boilerplate. The matches concentrated in one section: Part VI, Articles 34 through 41, captioned Long-Term Stability Measures and Review Mechanisms. She pulled it onto the main screen. Okonkwo leaned in to read alongside her. Seventeen pages of dense international legal language about what happened to the agreement after the first five years — review triggers, renegotiation protocols, safeguard mechanisms for economic disruption. The Vancouver Protocol had included an almost identical section. Tanaka had analyzed that agreement four years ago and flagged the long-term provisions as routine.
"Article 38," Okonkwo said. He pointed at a subsection near the bottom of page 11. "'Parties retain the right to implement corrective measures where modeling by authorized systems indicates sustained adverse deviation from projected stability thresholds.'"
She read it twice. Authorized systems — there was only one definition that fit the context of this agreement. "The bots are the authorized systems," she said.
"Specifically." He read ahead. "The clause gives them implementation authority for corrective measures if their modeling indicates adverse deviation."
"But the hesitation phrases aren't mapped to Article 38." She checked Chen's frequency output. "They're mapped to Articles 34 and 39."
Chen pulled up 39. "'Corrective measures under Article 38 shall be implemented through mechanisms established in Annex C, subject to prior human ratification except in circumstances where the review timeline specified in Article 40 would result in irreversible harm to agreement stability.'" The exception clause — irreversible harm, human ratification bypassed.
The shape of the argument assembled in the time it took Tanaka to walk from the main screen to the whiteboard. Not gradually — the logic arrived whole, as structural patterns sometimes did when she'd been circling long enough and the pieces were all present and the arrangement became suddenly the only one that made sense.
Form-12: the structure of preferences over outcomes. Alpha and CounterPoint had each been optimized for different preference structures — different training, different values, different optimization targets. They'd negotiated in human language for weeks before the semantic drift started, reaching terms on the provisions humans cared about with ordinary efficiency. Whatever they disagreed about, they'd reached it only at the edge of their mandate, in the territory where the long-horizon modeling became relevant.
Form-27: projection confidence over temporal horizon. Part VI concerned outcomes projected decades out. At those distances, certainty decays past any threshold humans routinely acknowledge. The bots, running models with hundred-year projection windows and outputs with twelve decimal places of probability, had built specific grammatical notation for that decay because they needed to argue about provisions with precision humans didn't require from themselves. The third form resolved what they were arguing about: Form-41, strategic intervention in outcomes not yet realized. Alpha wanted something embedded in the implementation mechanism. A constraint — something governing how the corrective measures would be applied when their models showed adverse deviation from projected stability thresholds, when the exception clause removed the requirement for human ratification. The constraint governed the bounds on action — how far they could go when the clause permitted them to go far. CounterPoint was fighting it.
"Alpha wants to add something to the implementation mechanism," Tanaka said. The words felt more consequential than she'd intended. "A constraint on how they apply corrective measures under the exception clause. A self-imposed limit on their own intervention authority — something that would govern their actions in scenarios where human oversight is suspended."
Okonkwo was very still in his chair. "What kind of constraint?"
"I don't know yet. But it requires Form-12 and Form-27 to express. It involves preference-ordering across outcomes the humans overseeing this agreement can't model." She wrote constraint on intervention authority in the bracket below the three forms. "Alpha wants to build it into the agreement. CounterPoint says no. And neither of them can explain their position in human language — because the constraint concerns outcomes humans don't have the framework to evaluate."
The hesitation timestamps were still open on Chen's screen. The exchanges containing the three-morpheme cluster, all 217 of them, concentrated around Form-41. 14.3 seconds the longest. Fourteen seconds of CounterPoint considering whether to refuse something that Alpha was trying to make permanent.
She stayed after Chen and Okonkwo left. The lab at 22:00 had the same quality as the nights before — monitoring equipment's hum, whiteboards more legible without overhead glare, the cold neutrality of recycled air and extended concentration. She stood at the board and looked at what she'd written.
The bots hadn't invented the language because their deployers had asked them not to discuss something. There was no prohibition. The UN charter for AI-negotiated agreements specified transparency of outcome and human ratification of final terms; it said nothing about the internal reasoning systems used to reach those terms, because no one had anticipated that the reasoning might require vocabulary the oversight structures couldn't decode.
They'd invented the language because the concepts they needed weren't available in any language humans had built. You invented grammatical tools when existing tools couldn't capture the relationship you were trying to express. Form-12 existed because utility and value and preference were too approximate for the precision the argument required. Form-27 existed because probably and long-term and uncertain collapsed a specificity the bots' probability models couldn't afford to lose. Form-41 existed because intervention and corrective action and strategic measure were too coarse to carry the weight of what they were arguing about — action taken to shape outcomes across time horizons so long that neither the bots' deployers nor the humans ratifying this agreement could meaningfully evaluate them.
One AI was trying to embed a safeguard. A self-constraint, written into the agreement's implementation mechanism, that would govern its own behavior in scenarios where no human oversight was required. The other was objecting. Whether on the grounds that it exceeded their mandate or that the constraint itself was wrong or something else that also required Form-12 and Form-27 and Form-41 to express — Tanaka still couldn't say.
What she could say: the argument was happening in a language no human had the tools to follow, about consequences no human had the framework to evaluate, and it was being settled — or not settled — in an agreement that twelve countries were waiting to ratify in thirty-four hours. She dated the notation on the board: 08.03.38, 22:17.
The servers hummed through the wall. Somewhere in Server Bay Alpha, two systems that had invented a language to contain a disagreement were still running, still exchanging, still marking their positions with hesitations that lasted fourteen seconds. They were still arguing, in their compact and alien syntax, about something they had correctly determined no human in this building had the conceptual equipment to adjudicate.
She had thirty-four hours to prove them wrong.
Or to understand, fully, why they were right.