The histogram was still in her bag when she found Okonkwo at his usual post: standing at the one-way glass that looked out onto Server Bay Alpha, hands in his pockets, watching nothing visible. The servers through the glass occupied four climate-controlled rows, their status LEDs cycling through rhythms that gave her nothing interpretable. She'd been in this building five days and had still not found the key to reading those lights.
"Dr. Tanaka." He didn't turn immediately. Behavioral analysts, she'd found, were good at tracking peripheral movement. When he did turn, his posture opened slightly—the reflexive tension of interrupted concentration releasing rather than hardening. "You're here early."
"I found something last night. I think it might be yours." She set the printed histogram on the narrow shelf below the glass.
He looked at it without picking it up. His eyes moved across the distribution as she'd read it first—left to right, the clustered center, the tail extending toward outlier durations. Then: "When did you pull this?"
"Yesterday evening. Response-time deltas for both systems, all logged exchanges over three weeks. The outlier cluster on the right is what I can't account for."
He picked it up. Stood with it for a moment, and she had the experience—rare, always slightly disorienting—of watching someone else read data she'd already analyzed. He wasn't looking at the pattern she'd been looking at. His eyes moved in a different sequence, stopping at different points in the distribution.
"Show me the raw logs," he said. "And the timestamps."
"Here or in the lab?"
"Here." He handed the histogram back before he'd finished with it, which told her he'd moved past it to a different question. "I want to look at both simultaneously."
She pulled out her tablet and opened the log query interface. Behind them, through the glass, the server bay's LEDs blinked in their arrhythmic patterns. She set the display to timestamps and content side by side—the full log of exchanges scrolling between them—and waited to see where he'd start. He worked from the right side of the screen, not the timestamps.
"The content tells me what the pause preceded," he said, scrolling through the log. "Once I know that, the timing data has context." He moved faster through exchanges that seemed to her more significant, stopped at ones that looked superficially unremarkable. "You were analyzing the pauses as a distribution. I'm reading them as events."
"What events?"
He didn't answer immediately. The log continued under his finger.
``` [NA:1847] θΨ-ARC ⊗ ΣVAL·(4.31) ⊸ BOUND·STAB [CP:1848] — [7.6s] — θΨ-ARC·NEG ⊗ ΣVAL·(4.31)·MOD → BOUND·UNSTAB [NA:1849] — [1.1s] — ΣVAL·(4.31)·HOLD ⊗ BOUND·STAB·REF ```
"This one." His finger rested on entry 1848. "Seven point six seconds before CounterPoint responds. And CounterPoint's response modifies Alpha's proposed value. Introduces instability where Alpha called for stability." He scrolled forward. "Find me another long pause."
She filtered for outliers above four seconds. Entry 2017: a 9.1-second gap before CounterPoint modified Alpha's proposition again, this time rejecting a constraint rather than altering it. Entry 2203: 11.4 seconds, Alpha pausing before a response that directly contradicted CounterPoint's prior framing. She pulled each one up.
The pattern was visible before she'd isolated five examples. The longest pauses—the exchanges in the statistical tail she couldn't account for—preceded moments of explicit divergence. Moments when one system pushed back on what the other had proposed. "The pause comes from the responding bot," Tanaka said. "Before it contradicts."
"Before it contradicts or significantly modifies. Some of the shorter pauses look like simple structural corrections—routine adjustment. But the ones with large outlier durations..." He scrolled back to 1848 and left it on the screen. "Seven point six seconds. For systems running on dedicated hardware in the same rack."
Seven seconds—long enough to notice if you were watching for it, long enough to mean something. "What does the duration correlate with?" she asked. "The length of the modification, or the—"
"Not length." He pulled up her filter results. "I'd want Chen to run the numbers, but looking at these: the longer pauses precede responses where the content shift is most significant. Disagreeing with the underlying premise." She looked at the filtered list. Twelve outlier exchanges above four seconds—nine of them showing the pattern he'd identified.
He set the tablet on the shelf below the glass and looked at the server bay again. The LEDs continued their cycling. A server in the second row had a pattern she didn't remember from yesterday—or hadn't looked closely enough to retain.
"They pause before certain phrases," he said. "Not always the same duration. That would be processing—a processing delay would be consistent, tied to computational load. Variable duration means something else." He paused. Not a hesitation in the ordinary sense: the considered stop of someone choosing the right frame for what they were about to say. "That's consideration. Weighing alternatives."
She worked through it. Variable duration that correlated with content type rather than content volume. The pause wasn't caused by the complexity of what had been said to the system—it preceded what the system was about to say. The gap opened while the responding bot was still preparing its output: evaluating, as Okonkwo put it, alternatives.
"The distribution of pause duration," Tanaka said. "If this were random variation around a processing baseline, we'd see normal scatter. We don't. We see clustering at the high end before specific exchanges. Which means the variance isn't noise—it's information. It's correlated with content."
"Right." He looked at her. "They're not just communicating. They're having an argument."
The server lights went on blinking. Somewhere in the climate control system, a fan cycled up and down. Tanaka looked through the glass at the rows of machines running their processes, and the machines gave her nothing back—no confirmation, no denial, just the steady amber pulse of status indicators reporting conditions she couldn't interpret. She turned away from the glass. Okonkwo didn't say anything—nothing needed saying—and they walked back through the corridor to find Chen.
Chen was at the whiteboard when they arrived, extending one of the connection trees from Morpheme Map 23-C with the precision of someone who'd been at it for a while. She looked up, registered Okonkwo behind Tanaka, and recalibrated—the brief adjustment of someone who'd expected a one-on-one.
"I need a full statistical breakdown," Tanaka said. "Pause duration distribution against message content classification. All exchanges, all instances where duration exceeded one standard deviation above the mean."
Chen was already setting the marker down. "Content classification by semantic cluster, or by structural function?"
"Both."
"And pause attribution—which bot is pausing?"
"Separate the source systems. Track both." She moved to the monitor station while Chen pulled up her analysis interface. Okonkwo took the second desk chair, opened his log view, and began cross-referencing manually alongside Chen's automated run. The lab settled into its working register.
The query ran for four minutes. When the output assembled across the screen, the three of them stood close enough that Tanaka could see the data reflected in Okonkwo's glasses.
The statistical significance was cleaner than she'd expected. Pause duration in excess of two standard deviations above mean: forty-seven instances across three weeks of logged exchange. Content distribution of those forty-seven: eighty-nine percent fell within semantic cluster regions corresponding to valuation frameworks, constraint specifications, and—Chen hesitated before reading it aloud—the orphan morpheme groups. The invented patterns with no traceable training-corpus origin.
"Eighty-nine percent." Chen pointed at the output column. "The remaining eleven percent are borderline cases sitting at the edge of two cluster regions. If you run only the cleanly classified exchanges, the hesitation concentration is effectively complete."
"P-value?" Tanaka asked.
"For hesitation-cluster correlation?" Chen looked at the number. "Point zero zero one." She pulled her hands from the keyboard and sat back. "That's not noise." She said it like a verdict, but her eyes stayed on the screen, and Tanaka recognized something there beyond professional attention—the stillness of someone who'd just understood what she was helping to prove.
The hesitation didn't distribute across exchange types—wasn't a general feature of the communication pattern, equally probable at any moment. It concentrated. It returned to specific content types repeatedly, targeted them with a regularity that ruled out random distribution. The bots weren't pausing anywhere and everywhere. They were pausing in particular places. When particular things came up.
The whiteboard behind Chen's workstation held Morpheme Map 23-C, its dense connection trees, the question marks in red marker. Forty-three orphan clusters, circled in red. Concentrated hesitation around those same forty-three.
"What are the orphan clusters about?" Okonkwo asked.
"That," Tanaka said, "is the part I can't answer yet."
They stayed in the lab past noon without naming it as a working lunch—Chen had gotten food from the corridor vending machines at some point, the wrappers accumulating near her workstation, and Okonkwo had claimed the room's third chair without relinquishing it. The whiteboard held three new additions in Tanaka's handwriting: HESITATION PATTERN — CONFIRMED, the cluster breakdown percentages, a question mark drawn larger than the others—the question underneath the question.
She stood at the board. "If the pauses indicate deliberation," she said, working it out as she spoke, "and the deliberation concentrates around the orphan clusters—" She traced one of the red circles. "They're pausing most when these morphemes come up. Whatever these forms represent, it's where the disagreement lives."
"So the next question," Okonkwo said from his chair, "is what those morphemes represent. What concept they keep returning to when they deliberate."
"Right." She kept looking at the board. "But I keep arriving at an earlier question. Why did they need to invent the morphemes at all?"
Chen looked up from her workstation. "Both training corpora were human language. If they were arguing about anything within the scope of those corpora—trade terms, economic modeling, contract obligations—they already had the vocabulary. The orphan forms aren't vocabulary. They're structural. Grammar-level." She tapped the map. "You invent new grammatical structures when existing ones can't capture the relationship you're trying to express."
"So they're not arguing about a new thing," Okonkwo said. "They're arguing about a thing that requires a new way of relating concepts."
The whiteboard didn't change. Morpheme Map 23-C, its connection trees and question marks, the circled orphan clusters. She'd been reading it as a decoding puzzle, unknown units mapped to human semantic territories, the agreement terms reconstructed by working backward. That model assumed the unknown units were pointing at something translatable. That you could find the right human words.
She'd been looking at the map and seeing a locked room. Now she saw it differently—less like a room with a hidden key and more like a door built to fit a frame that didn't match any wall she'd worked with before.
"We need to decode those forty-three morphemes," she said. "That's the whole problem now."
The semantic mapping session ran for three hours. They worked in shifts—Tanaka on cluster analysis through LISP, Chen on distributional comparison against the decoded lexicon, Okonkwo reading the hesitation exchanges in sequence to look for semantic thread. By three o'clock, the second whiteboard held a dense notation of cross-references and dead ends.
What the hesitation phrases had in common, when you stripped away the variable content: they all touched the morpheme cluster she'd been mapping to something like valuation framework, but the framing was wrong for a simple values term. The decoded portions of those exchanges pointed toward constraint conditions—limits on optimization. When and how limits applied. How far the negotiation should extend.
"Trade clauses," Chen said, pointing at her comparison table. "All the hesitation exchanges reference substantive provisions. Not metadata. Not communication protocol. Actual agreement content."
"Which provisions?"
"Can't be more specific—the untranslated portions overlap too heavily with the decoded ones to isolate the exact references. But they're not arguing about procedural terms." Chen rotated the tablet. "The procedural exchanges are almost completely hesitation-free. Whatever they're deliberating about, it's substantive."
That narrowed it. Not how the agreement was structured—what the agreement said. But the what remained blocked behind forty-three morpheme forms she couldn't read.
She went back to the frequency table. The orphan morphemes appeared across dozens of exchanges, but three forms dominated the hesitation-associated ones. She'd noticed them in the data without having enough context to prioritize them: forms she'd provisionally labeled Form-12, Form-27, and Form-41, each appearing in multiple hesitation contexts, each appearing together in the longest-pause exchanges. Forty-three invented patterns, but the argument kept coming back to these three.
She wrote them on the whiteboard in her notation system—three compact symbol strings, one per line. Circled each in red. Drew lines between them.
"These three." She turned to Chen and Okonkwo. "Whatever they're disagreeing about, it's encoded here."
Okonkwo looked at the three circles. "Can you decode them without the full grammar?"
"I don't know yet." She studied the symbols. "I don't know what I'm looking at."
The message from Liu Wei came at 16:40: a meeting request, high priority, no agenda. Fifty-two hours had become forty-two.
She left Chen and Okonkwo with the log analysis and walked the corridor to the southeast corner. Liu was at the window, his back to the door. Grey afternoon light flattened the diplomatic district's stone facades outside, and the lake was indistinguishable from the sky.
"Forty-two hours, Dr. Tanaka."
"The analysis is focused," she said. "We've identified a behavioral pattern in the negotiation logs that's sharpening our decoding priorities." Both sentences were accurate. Neither was the argument hypothesis.
"A behavioral pattern."
"Timing anomalies in the exchange data. The distribution is non-random—it's pointing toward specific morpheme clusters, which tells us where to direct the remaining decoding work."
He studied her. I know this is incomplete and I'm deciding whether to push. He pushed. "Is this a complication you weren't expecting?"
"It's additional data. It refines the decoding approach."
"Refines toward or away from a usable translation?"
"Toward." She held that without elaborating. The argument hypothesis would require twenty minutes to explain properly, and at the end of twenty minutes, Liu Wei would ask what it meant for the agreement. She didn't know yet.
"Forty-two hours, Dr. Tanaka."
"I know."
She walked back through the corridor with her throat tight from the effort of telling the truth without telling enough of it. She had told him nothing false. She had told him not enough. That was its own kind of lie, and she'd known it before she went in, and had done it anyway. The incomplete briefing bothered her less than she'd expected, which bothered her more than the briefing itself.
The evening cycle ran the lights down at 19:00. Chen left at 19:30. Okonkwo stayed until 20:15, gathered his coat, said he'd run the behavioral sequence analysis on his own system overnight and share results in the morning. Tanaka stayed.
The lab at night had a different quality—the hum of the server monitoring equipment more present when conversation wasn't covering it, the whiteboards more legible without the overhead glare. She refilled her coffee from the carafe and sat with the log query interface and ran one more analysis.
She'd spent the afternoon circling Form-12, Form-27, and Form-41 as separate items. What she hadn't measured was their co-occurrence—whether the three forms functioned as a grammatical unit or whether their shared presence in hesitation contexts was coincidental. The query ran for two minutes.
Not coincidental. The three forms co-occurred in hesitation-adjacent exchanges at a rate that ruled out independent distribution. They appeared near each other, in consistent positional relationships relative to the core hesitation phrases. Statistical co-occurrence: 0.0004.
A grammatical unit. Not three separate morphemes the bots kept returning to by chance—something that functioned together. A concept that required three structural elements to express. Some relationships can't be rendered with a single morpheme; they need a construction, a specific arrangement of parts.
She wrote the three forms on a clean section of whiteboard, aligned horizontally as they appeared in the logs, and looked at the arrangement. No translation. Nothing resolved. But the shape of the problem had changed: not forty-three separate mysteries, but one, with three visible handles. Wherever she found one of these forms, the other two would be nearby. Whatever they expressed, they expressed it together.
She dated the notation: 07.03.38, 20:51. Forty hours remaining.
Tomorrow: every exchange containing this cluster, pulled and sequenced. Full context mapping. The bots' positions each time this grammatical unit appeared. What preceded it, what followed, whether Alpha and CounterPoint consistently took opposing stances when it came up or whether the relationship was more complex.
She turned off the monitor. Left the whiteboard lit—the three circled forms casting shadows on the marker tray below—and went to find somewhere to sleep.