Chapter 104: The Provenance of a Phrase
[Noe Valley apartment Friday, February 6, 2016 9:15 AM]
Coffee from the Blue Bottle on 24th, carried home in a cup he'd forgotten to wash yesterday. The apartment still smelled like packing tape. Six boxes against the wall, unopened since August. Monica had stopped mentioning them in October.
His phone buzzed at 9:17. Priya. He opened it expecting the usual a flagged paper, a curation note, the mug complaint in subject line form.
It was a link.
Hooli careers page. A posting for "Senior Research Engineer, Large Scale Language Model Curation." He'd seen a dozen of these across companies since November the field was forming around Gardner's published work and the copycats were hiring. He almost closed it.
Then he read the second paragraph.
"Candidate will demonstrate proficiency in domain adversative token partitioning across heterogeneous source registers, with emphasis on medical adjacent and legal adjacent corpora isolation boundaries."
He read it twice. Not because the language was unfamiliar. Because it was too familiar.
Domain adversative token partitioning. Heterogeneous source registers. Medical adjacent and legal adjacent corpora isolation boundaries.
That was Priya's language. Not the field's hers. From an internal curation methodology document dated November 12, specific phrasing she'd developed to distinguish their isolation protocol from the standard approaches. Three compound terms in a row that appeared in exactly one document in the world, and that document had never left the building.
He called her.
"You've seen it," she said. Not a question.
"When did it go up?"
"Fourteen hours ago. I found it at midnight. I have not slept, and I would like you to know that I am not being dramatic about that. I am being factual."
"Can you send me the internal doc with those exact phrases?"
"I am looking at both side by side right now. The posting uses twenty two of my compound terms. Nine. In the same order. The third paragraph of our methodology section is reproduced almost verbatim with the serial numbers filed off." She paused. "Someone has my document, Ethan. Not a version of it. That specific version, revision 4.2, which seven people have read."
He set the coffee down. "Who can reach revision 4.2?"
"Sarah's team. Marcus, for infrastructure review. The integration channel copy that went to Pied Piper that was 4.1, but the diff between 4.1 and 4.2 is twelve lines, and whoever wrote the posting has those twelve lines."
[Gardner Analytics 11:30 AM]
Sarah was already in the glass office when he arrived. Red marker uncapped, untouched. She'd printed the Hooli posting and the internal doc on facing pages of the same sheet.
"Marcus," she said. "The entry baseline."
The log protocol built the first week of August, formalized after the Gilfoyle truce at Sightglass. Jared's named human discipline. Every query timestamped, every credential tied to a person.
Marcus pulled the integration channel logs in twenty minutes. The channel to Pied Piper was the shared architecture bridge for the decoder handover. It had logged entry windows every Tuesday and Thursday between 2 PM and 6 PM since November. That was Richard's team. Routine, scheduled, logged under Jared's protocol.
Then Marcus cross referenced the access windows against the curation document's revision history.
Revision 4.2 was pulled through the integration channel on December 9, a Tuesday, at 3:47 PM. The channel's log showed the request originating from a Pied Piper credential. The download included the full curation methodology directory not just the architecture documents the channel was built for.
"He pulled everything in the directory," Marcus said. Flat. "Not just the decoder files. The curation doc was in a parent folder. The channel didn't restrict by file type."
"Could this have been automated?" Ethan asked.
"The pattern is manual. Single pull, single session, human paced. Someone sat down and downloaded it."
Sarah asked the question he was avoiding. "Who at Pied Piper has credentials on that channel?"
Three people. Richard. Gilfoyle. Jared.
The room went quiet. Ethan felt the stillness arrive in his forearms first that specific buzz below the sleeve. It meant a threat had landed and his body was processing it before his mind finished.
Sarah looked at him. Not asking how he knew. Never that. Asking what he was going to do with what they had.
"I want this analyzed by someone with no relationship to Pied Piper," she said. "Procedurally clean. If this goes anywhere board, legal, the partnership we can't have run it ourselves."
She was right. She was always right about procedure.
That put the forensics on Priya's desk.
[Same office 4:00 PM]
Priya took the logs at noon and worked through the afternoon. By four she had a preliminary timeline: the December 9 access, a second pull on January 14, both inside Tuesday Thursday windows, both from Pied Piper credentials.
"It's ongoing," she said. "Two pulls, six weeks apart. The methodology went first. The second pull accessed the decoder detail the withheld layer. Whoever this is took two trips."
"Can you confirm it's the same credential both times?"
"Same credential. Same user ID." She removed her glasses the tell, the one that meant her model was shifting. "I need to ask you something and I need a sourced answer."
"Go ahead."
"The claim that this originated from Pied Piper's credentials. How do you know the channel wasn't compromised externally? How do you know these credentials weren't cloned?"
He didn't have a sourced answer. He had Marcus's protocol named humans, rotated keys, audit logs and it said the credentials were clean. But the protocol was Marcus's best practice, not proof.
"Marcus's baseline shows no credential anomalies," he said.
"That's a partial answer."
"It is."
She put her glasses back on. "I'll take it as preliminary. I will not take it as conclusive, and I want that noted."
[Same office Saturday, February 7, 11:50 PM]
Daniel Oyelaran had taken the incident response on Friday evening and run it through the weekend. He'd rotated every integration credential. He'd tightened the channel to architecture only, cutting the curation directory's path off entirely. He'd written a twelve page forensics brief documenting the exfiltration route, the timeline, the credential trail, and the specific documents confirmed in Hooli's possession.
It was excellent work. Ethan read it twice Saturday night and found no flaw in it.
The brief closed with a recommendation: suspend the shared architecture channel pending review. Neutral language. Procedurally correct. The kind of recommendation a competent VP of Operations writes when the evidence points where this evidence pointed.
Ethan sat with the printed brief, annotated in his own hand. His phone was face down on the desk. Richard's name was on the last incoming call, from Wednesday. Two days before the posting went up, before any of this, when the partnership was still the bet he'd made on a nine that didn't betray.
He did not pick up the phone. He did not call Richard tonight.
The reasoning was clean: do not confront until the forensics are complete. That was discipline. That was correct process.
And underneath the correct process, in a place he was not yet examining, sat the alternative. That a nine had turned out to be a bad person was a story about someone else's character. The alternative he was not choosing that the instrument had never measured what he needed it to was a story about his own.
He chose the first one. He would keep choosing it for a while yet.