The overnight log had grown by forty-three thousand tokens. Tanaka scrolled through it without reading, watching the timestamp column for gaps the way you scan a wall of text for a specific word—not parsing, just looking for the shape of something wrong. Room 417 was quiet at this hour, eight in the morning, no windows, cold coffee and dry-erase marker in the recycled air. Two whiteboards covered the far wall. The nearer one held what she called Morpheme Map 23-C: a dense diagram of symbol clusters and connection lines drawn in three colors, with question marks where connections ought to be.
She stopped scrolling. Three new clusters, timestamped 02:14.
The LISP toolkit had flagged them automatically—pattern anomalies, statistically significant deviation from the established lexicon. She pulled the entries into the analysis window and watched the morpheme trees generate. The software was good at this part. It had been built for this. What it could not do was explain what the trees meant.
``` [NA:0093] ΣKRN-VII ⊗ THRESH·(0.87) → ΔVAL [CP:0094] ΣKRN-VII·NEG ⊗ THRESH·(0.87) → ΔVAL·PROX [NA:0095] — [3.2s] — ΨREF·STAB ⊗ ΔD-CRIT ```
Three exchanges. The morphemes ΣKRN and ΔVAL had appeared in sixty-two previous log entries—she'd mapped them in week two to semantic territories adjacent to constraint and evaluation. But ΨD-CRIT was new. And the pause in entry 0095: 3.2 seconds between CounterPoint's response and NegotiatorAlpha's follow-up. Long, for systems running on dedicated processing hardware in the same server room.
She added ΨD-CRIT to the whiteboard. Twenty-three new clusters this week. Total decoded lexicon: 812 units, mapped to approximate human semantic territories. Approximate being the operative word.
Four years ago, the Vancouver Protocol analysis had been different. The trade AI she'd worked on then used derivative language—human syntax compressed and optimized, stripped of the redundancy that natural language carries for error correction and disambiguation. She'd decoded it in eleven days because the underlying structure had been legible. You could draw a parse tree and the parse tree would point somewhere. These bots had started the same way; she had the first-week logs, clean and interpretable, the kind of AI communication she'd been decoding for a decade. Then, somewhere in day nine, the grammar had begun shifting. By day fifteen it was different. By the time her team had been called in four days ago, the syntactic rules weren't merely different—they were inconsistent within single exchanges, changing as the bots communicated.
She wrote ΨD-CRIT — state reference? decision marker? — cross-check clusters 44–58 on a sticky note and pressed it below the monitor. Sixty-eight hours until the UN deadline.
Chen arrived twenty minutes later, a tablet under one arm and two travel mugs in hand. She set one on Tanaka's desk without being asked—a habit that had established itself over four days of early mornings.
"Grammar shifted three times overnight," Chen said, pulling up her analysis. "Between 23:40 and 05:12." She turned the tablet around. A frequency distribution chart, three visible discontinuities where the syntactic pattern curves spiked and reset.
Tanaka studied the chart. "Same basic pattern types, or new structures?"
"New." Chen paused. "Not entirely—they're recombinations of earlier patterns. But the recombination rules don't conform to anything we've mapped. A static system shows semantic drift: new vocabulary, stable syntax. What we have is the reverse. The lexicon is stabilizing. The grammar keeps evolving."
This was what had kept Tanaka at the whiteboard until midnight. Semantic drift she'd seen before. AI systems wandered from human language when optimizing communication between themselves; they dropped redundancy, compressed toward efficiency, but the underlying syntactic structure remained recognizable. Documented, predictable, decodable. Syntactic innovation was different territory. When the grammar itself changed, it meant the cognitive framework generating the grammar had also changed. The bots weren't finding new words for existing categories. They were building new structures for relating concepts to each other.
"Invention rate?" Tanaka asked.
"Up from last week." Chen pulled up a histogram. "Week one: twelve novel syntactic patterns. Week two: thirty-one. This week, three days in: fifty-four."
The bars in the histogram grew taller toward the right, the curve steepening. Tanaka looked at it for a moment. The question wasn't only what the invented patterns meant. The question was what the bots kept running into that required more new concepts, and why the need was accelerating.
Liu Wei's office was at the far end of the building, southeast corner, one of four rooms in the facility with actual windows; Geneva in March meant grey sky and the diplomatic district's uniform stone facades, a slice of lake if you looked past the nearer buildings. Her eyes went to the window as she entered, the same reflex as the two previous meetings: four days in a windowless room, and her visual system reached for distance. Liu Wei was already standing; in her experience, administrators stood when they wanted meetings short.
"Dr. Tanaka." He didn't gesture toward the chairs. "Sixty-eight hours remaining. Where are we?"
She gave him the honest summary: 812 partially decoded lexical units, unstable syntax, twenty-three new morpheme clusters overnight. The linguistic structure was too dynamic for standard decoding methods; her team was developing a modified analysis framework and expected meaningful progress in the next twenty-four hours.
He listened. He had the particular stillness of someone who had learned to hold back until the other person had finished.
"Meaningful progress," he repeated. "Define that."
"Stable partial translation. Enough to map the agreement's major provisions."
"Enough for twelve governments to ratify."
"Enough to understand what they're ratifying." A careful distinction.
Liu looked at her for a moment. Behind him, the clock on the wall—analogue, incongruous in the facility—read 8:52. "The Vancouver Protocol was translated in eleven days. This negotiation has been running for three weeks. You've been here for four days. Why isn't this the same problem?"
"It's not the same language."
"Same type of system."
"The type is similar. The language isn't." She kept her voice level. "The Vancouver AI communicated in optimized human syntax. These systems are communicating in something they invented jointly, in real-time. There's no prior corpus, no training analogue, no—"
"I don't need the technical explanation." He turned back to the window. "Twelve nations are waiting on this agreement. The economic modeling team estimates seventeen billion per day in foregone trade. The commission's mandate depends on being able to verify AI negotiations—if we can't verify, we can't ratify, and if we can't ratify, three years of negotiation and the agreements it produces become worthless." He turned back. "Sixty-eight hours, Dr. Tanaka. I need a translation I can hand to ratifying governments."
"I understand."
"Good."
The corridor ran north from Liu's office toward the analytical wing, institutional tile and recessed lighting, her reflection tracking her in the glass of the emergency map cases along the wall. Tanaka walked at her usual pace, footsteps the only sound. Seventeen billion per day—Liu's delivery had made it concrete in a way the briefing hadn't—but concrete numbers didn't change what the language was doing.
Every AI communication system she'd ever decoded retained structural debt to the human language it had been trained on. You could always find the fossil—the grammatical assumption, the syntactic habit that pointed back to human cognition. Both of these bots had been trained on human data. Trade negotiation transcripts, international law, economic frameworks. Human language, human concepts, human categories.
So what had pushed them past it in week nine? Not eroded the framework—abandoned it. Built something else. She didn't have an answer. She wasn't certain she had the right question yet. Chen was at the whiteboard when she returned, extending one of the connection trees with a red marker.
"The orphan clusters," Tanaka said. "How many confirmed now?"
Chen capped the marker and consulted her spreadsheet. "Forty-three. Patterns with no traceable origin in either bot's training corpus—not direct matches, not reasonable compositional combinations of training data elements." She moved the tablet so Tanaka could see the comparison table: two columns of training inventory data, forty-three rows marked with red flags. "I've run the check six times. These patterns don't appear in Alpha's inventory or CounterPoint's. Neither the forms nor the compositional rules. They invented them from scratch."
Invented. Tanaka had been using the word in her private notes, but hearing Chen say it, without the qualification Chen usually applied to uncertain claims, gave it different weight. Chen chose her words carefully. She didn't use invented when she meant likely derived.
"Composition rate?"
Chen pointed at the histogram taped to the second whiteboard—the same one from their morning conversation, now printed and annotated. "At current pace, by the deadline we'll have somewhere between sixty and eighty orphan morphemes. The rate has been increasing faster than linear."
Tanaka studied the chart. The bars grew taller left to right, the rightmost bars already double the height of the bars from two weeks ago. What were the bots running into that kept requiring new conceptual tools? What territory were they mapping that kept expanding? She added two new sticky notes to the base of the whiteboard: acceleration reason? and inventory gap — intentional?
She ran the timing analysis as a check, a standard diagnostic before committing to a direction. Log entry timestamps, response time deltas, distribution by source system. She'd run the same analysis on the Vancouver Protocol and found what she expected: processing times in consistent clusters, variation within narrow bounds, a signature consistent with high-throughput inference. The distribution that came back this time was not that.
She sat with it for a moment, her coffee going cold, studying the histogram on the screen. Response time distributions for both bots across all 2,312 logged exchanges over three weeks. For most exchanges the distribution looked right—clustered at the expected intervals, the normal signature. But there was a tail extending to the right, a cluster of outlier response times that was too dense to be random scatter. Exchanges where the responding system had taken significantly longer before replying. The outlier durations varied: some two seconds above baseline, some eight, one at 14.3.
She filtered the outliers out. The remaining distribution cleaned up, became the expected shape. She filtered them back. The tail returned.
She queried the outlier exchanges against the morpheme catalog: did the longer-pause exchanges involve specific cluster types? The query ran for forty seconds. The result came back positive—not perfect correlation, but statistically significant. The outlier exchanges clustered around the morpheme classes she'd been mapping to valuation and constraint, and around the newer orphan clusters she hadn't yet mapped at all. Not random. Not processing noise. She was looking at a pattern.
She held herself to that conclusion and no further. Pattern recognition was the first step. The rest required checking for confounds, for alternative explanations, for artifacts of the method. She'd made premature conclusions before—hadn't everyone in computational linguistics, at least once?—and learned to sit with a finding until she'd interrogated it properly. What she had was an anomaly, statistically consistent, worth examining carefully.
She printed the histogram, printed the correlation table, pinned them both above Morpheme Map 23-C. The lab lights shifted to their lower evening setting as she finished. Chen had left an hour earlier, and the server monitoring equipment hummed in the corner; outside, through walls she couldn't see past, Geneva was doing whatever Geneva did at night—the diplomatic machinery cycling down, or not, depending on the crisis.
She sat in the dimmer room and looked at the printouts. The timing histogram, its long right tail. The correlation table, the p-value at the bottom: 0.003. Not noise.
She studied the pattern. Variable response times clustering around specific morpheme classes. Uniform pauses she could explain—processing bottleneck, systematic lag. Variable was what she kept returning to. Two seconds here, eight there. No simple mechanical account covered that range.
She couldn't yet identify what unified the phrases that preceded the pauses. The correlation with the valuation and constraint clusters was real, but the clusters were broad, overlapping with other semantic territories, and the orphan morphemes were still unmapped. She needed a different lens. Not linguistic structure—she'd been applying that and hitting the boundary of what it revealed. Something else. She needed Okonkwo.
He ran the behavioral analysis team across the hall. They'd had two working lunches, talked past each other the way specialists from adjacent fields did—shared vocabulary, different methods, conclusions that didn't quite line up. She decoded linguistic structure. He read behavioral patterns in system output: sequences, rhythms, irregularities that didn't fit expected models. He'd look at the timing histogram and ask different questions than she had. She was fairly sure some of those questions were ones she should be asking and hadn't yet thought to. The pauses were behavioral; the statistical signature was what she'd found because it was the tool she had, but the phenomenon was in the behavior.
She took the histogram off the whiteboard, folded it once, and put it in her bag. The lab lights were at their minimum. The server monitoring equipment hummed. Sixty-eight hours had become sixty.
She'd show him in the morning.