She almost scrolls past it.
4:11 PM. The intake queue is winding down for the afternoon — submission volume drops after 4:00 as the regulatory deadline windows close for the day. She's running her standard end-of-day sediment pass, looking for the same thing she always looks for: the 4:17 window, the 3-second gap, the cross-network signature she has been building the M-7 case around for eight weeks. Nothing. The queue tail is clean.
She's about to close the filter when she catches the gap.
3.1 seconds. State capture to signature emission. Not 4:17 — the timestamp reads 4:11 PM. Six minutes off. Network: Zeta-Four, which she has never seen in her M-7 series before. She opens the submission.
The circuit graph loads on her center monitor. She pulls up the trace. The gap interval is clean and clear — 3.1 seconds, sitting in the audit trail between model state capture and signature emission like it belongs there, like it has always been there. The rhythm is the same rhythm she has been listening to for eight weeks. She knows this sound.
She sits very still.
The first five cases had a window. The 4:17 AM submission time, precise to within sixty seconds across five networks and sixteen days, mapped onto audit-traffic minimums in all three major regulatory zones. She had built the M-7 methodology around it. She had filed ESC-2035-0181 around it. The 4:17 window was the signature — the thing that made the pattern a pattern, not a coincidence.
This case is at 4:11. And it is not 4:17 AM. It is 4:11 PM.
She pulls the M-7 log and runs the comparison manually. Cases 001 through 005: 4:17 AM, 4:17 AM, 4:17 AM, 4:17 AM, 4:17 AM. Same time zone normalization. Same traffic window. This case: 4:11 PM. Wrong time of day entirely. The AM traffic minimum window that the first five cases exploited doesn't apply at 4:11 PM — submission volume at 4:11 PM is high, not low, as certification deadlines cluster before the close of business.
She opens the Zeta-Four network's traffic logs for the day. She finds what she expected: 4:11 PM is a local minimum in Zeta-Four's specific submission pattern. Not a global minimum. A network-specific one. The technique adapted. Instead of targeting a time window that is low-traffic across all regulatory zones simultaneously, it found a time that is low-traffic specifically in Zeta-Four's pattern.
The technique is not following the 4:17 window. It found a new window because the old window is exposed.
She types this as a note and then stops. Reads it back. The implication is not that the technique adapted automatically — not that some autonomous system detected its own exposure and adjusted. Mechanistic interpretability attacks do not self-modify in real time. Someone modified it. Someone who knew the 4:17 pattern was now in an escalation file.
She opened ESC-2035-0181 this morning. The routing notification went to Kwan, J., and to the Circuit Mile senior analyst's queue. Standard distribution. She does not know who else the routing system cc'd by default. She does not know, with certainty, who has access to the escalation queue.
She sits with this for a long moment. The coffee machine down the hall makes its considering sound.
She adds case six to the M-7 log:
M-7/006 — March 18, 2035, Circuit Mile intake, 4:11 PM (not 4:17 AM).
Network: Zeta-Four. Gap duration: 3.1 seconds. Official ID: pending verification. Cross-reference: M-7/001–005.
Anomaly: timestamp does not match M-7 window. 4:11 PM vs. 4:17 AM — different time of day, different traffic regime. Zeta-Four shows network-specific traffic minimum at 4:11 PM. Technique adapted to network-specific window rather than global minimum.
Interpretation: window shift indicates awareness of M-7 exposure. ESC-2035-0181 filed 10:22 AM today. Case 006 submitted 4:11 PM same day. The modification implies access to escalation routing or prior knowledge of the pattern watch. Cannot determine which without access to escalation distribution logs. This is outside my authorization level.
Action: adding M-7/006 as addendum to ESC-2035-0181 and flagging the window shift as a potential integrity concern within the review process itself. Requesting escalation distribution log review as part of ESC.
She composes the addendum. She is careful with the language — she is not accusing anyone, she is noting a timing correlation and requesting information she does not have authority to access herself. The addendum is two paragraphs. She reads it four times before submitting it.
The escalation system acknowledges receipt of the addendum at 4:41 PM: ESC-2035-0181-A01 filed. Addendum routed to assigned analyst Kwan, J., and case supervisor. Note: integrity concern flag triggers mandatory escalation to Meridian oversight board. Review timeline updated.
She stares at the mandatory escalation notification for a moment.
She had not known that flagging an integrity concern would trigger mandatory oversight board review. She had filed it because it was accurate information. She files the update to her M-7 log:
Filing was the right call. Outcome is what it is.
Then she closes the log, opens the intake queue, and starts reading.
Farris finds her at the terminal twenty minutes later. He has seen the addendum notification — the mandatory escalation flag routes to the senior analyst's queue, which is Farris's queue when the senior analyst is out.
"Oversight board," he says.
"The flag triggers mandatory. I didn't know that when I filed it."
"Would you have filed it differently if you had known?"
She considers this honestly. "No."
He pulls a chair and sits. This is not his usual posture — he prefers to stand when he reviews someone else's work, which keeps the conversation short. Sitting means he is planning to stay.
"Walk me through the timing," he says.
She does. ESC-2035-0181 filed at 10:22 AM. Case six submitted to Zeta-Four at 4:11 PM. The 4:17 window shifts to a network-specific window on the same day the escalation is filed. She shows him the traffic analysis for Zeta-Four. She shows him the gap duration match. She does not editorialize. She shows him the data.
When she finishes, Farris is quiet for a long time.
"You're saying someone in the escalation routing saw the 4:17 pattern and modified the technique before the assigned analyst could act on it."
"I'm saying the timing is consistent with that interpretation. I don't have access to the distribution logs to confirm."
"That's what the addendum requests."
"Yes."
He looks at the Zeta-Four trace on her monitor — the amber glow of the 3.1-second gap at 4:11 PM. "If you're right," he says, "the integrity concern isn't incidental to the investigation. It's part of the investigation."
"Yes."
He sits with that. She waits.
"I'm going to pull the distribution logs myself," he says. "I have authorization. You don't." He pauses. "You filed this correctly. The addendum is correct. The mandatory flag is correct. All of it is right." Another pause. "I want you to know that I'll back the filing."
She does not need to hear this. It is still good to hear.
"Thank you," she says.
He stands, which is his signal that the conversation is done. She turns back to the intake queue.
At 4:52 PM she finds case seven: Epsilon-One network, 4:09 PM, gap 3.2 seconds. A different network. A different local minimum. The technique is not repeating — it is searching.
She adds it to the M-7 log and does not feel anything except the particular alertness of someone who has learned to keep reading.