She did it at 4:17 in the morning.

The Institute was not empty at that hour — it was never truly empty, the overnight clinical floors maintaining their quiet, continuous work — but the Level 31 staff offices were dark and the monitoring staff were in the lower floors and the particular quality of 4 a.m. in an institutional building was the quality of a place that was running on minimum attention. She had her access credentials, which had not yet been revoked because the formal clearance suspension was pending review rather than enacted, and she had Nera's Level 6 authorization, passed to her through the personal encrypted channel, and she had a plan that she had spent three hours building with Kael and Nera in the outer district apartment.

She sat at an unassigned workstation — not her own, because her own would be the first place the access logs were checked — and logged in through Nera's credentials with her own biometric layer on top, a combination that would show up as an authorized access event for approximately forty-eight hours before anyone cross-referenced the biometric data.

Forty-eight hours was longer than she needed.

She opened Kael's assessment file.

---

The report structure was one she had built herself, three weeks ago, in the early days of the assessment when she had been a competent, trusted technician doing a difficult case by the book. She knew every field, every classification tag, every cross-reference point. She knew where the report's interpretive language was separated from its raw data — where the *what was observed* lived and where the *what it means* lived, and that the second category was far more malleable than the system's official presentation suggested.

She began with the behavioral log.

The raw entries were unchanged — the walkway, the design proposal, the time in the historical archive, the pattern of prediction failures. She did not touch them. Altering raw data was a serious technical violation and also, more practically, would be detectable in a serious audit. What she was changing was the interpretive framework around the data: the classification tags that determined how EIDOLON's systems categorized the behaviors, the contextual notes that shaped how the AI read the entries, the cross-reference links that connected his behavioral profile to the risk assessment database.

*Non-compliant action: subject diverted from assigned route to assist non-priority resident.* She reclassified: *Adaptive problem-solving: subject identified resource need and applied professional skills in non-standard context. See: Community Integration Supplementary Framework, Section 4.2.*

Section 4.2 of the Community Integration Supplementary Framework was a real document — she had found it in the Institute's archive the previous evening, a supplementary protocol that had been written into the system during the early post-Collapse reconstruction period and had never been formally removed. It provided an alternate classification pathway for citizen behaviors that served community needs outside standard optimization parameters. It had not been invoked in eleven years. EIDOLON's current systems did not automatically cross-reference it because it was in an archive tier that the active modeling databases rarely accessed.

It was, in technical terms, a door that was not locked but that nobody had thought to open.

She had found it in three hours of searching through the supplementary archive. She suspected her past self had found it too, at some point, and had left traces that had helped her find it faster than she should have been able to.

She worked through the behavioral log systematically, entry by entry, reframing each one within the Section 4.2 framework. It took two hours. The reframing was not a lie — every interpretation she applied was technically defensible within the document's language. But the effect was a transformation: a behavioral profile that EIDOLON's risk assessment read as *high anomaly, systemic threat* became, under Section 4.2, *non-standard community integration, supplementary framework applicable.*

She ran the classification update.

The system processed.

*CAI risk classification: Omega* began to shift.

---

It was while the classification was processing that she found the code.

Not immediately — it was nested three layers deep in the report's attachment structure, in a zone that the active reporting interface didn't display by default, visible only if you were working at the raw file level and knew to look for non-standard attachments. She was at the raw file level because that was where the classification tags lived, and she found it because she was thorough, and for a moment when she first saw it she thought she had encountered a system error.

It was not a system error.

It was a small, precise piece of code, written in the same language as the Institute's core reporting architecture, designed to execute a specific function when triggered by a specific set of conditions. The conditions were: a Section 4.2 reclassification attempt on the Orison file, initiated from within the Level 31 reporting system, authenticated by biometric data matching her own profile.

It had been waiting for her.

She read through it carefully — she was not a programmer by primary training but she had enough technical literacy to follow the logic — and understood what it did. When the Section 4.2 reclassification processed, the code would automatically file a supplementary cross-reference to seven other case files in the Institute's archive: seven other high-CAI subjects whose behavioral profiles had been classified under Omega designation in the past three years. Under Section 4.2, the cross-reference would register all eight cases — Kael's and the seven others — as part of a documented pattern of *community integration supplementary behavior*, which created a legal and procedural complication for any subsequent individual erasure order because the supplementary framework required a population-level review before individual action could proceed.

It would not stop the process. Nothing she was doing tonight would permanently stop the process. But it would create delay — procedural, documented, legitimate delay — on a scale that a single CAI reclassification alone could not create.

Her past self had understood that the real goal was not to win a single case. It was to introduce enough procedural complexity into the system that the system had to pause.

She sat at the dark workstation at 4 a.m. and read her own code and felt something she did not have a precise word for — a feeling that was partway between grief and recognition and the specific, strange relief of understanding that you have never been entirely alone in the work you thought you were doing alone.

*We are the same person*, the letter had said. *Continuing a job that isn't finished.*

She triggered the code.

---

The reclassification processed at 4:47 a.m.

Kael's CAI classification shifted from Omega to High — still flagged, still restricted, but no longer in the category that authorized immediate Permanent Erasure without additional review. The supplementary cross-reference filed automatically, linking eight cases into a population-level pattern that the Section 4.2 framework required to be reviewed by a designated supplementary panel before individual action could proceed.

The supplementary panel had not been convened in eleven years. Convening it would take, under the protocol's own requirements, a minimum of seventy-two hours.

She had bought them time.

Not victory. Not safety. Time.

She logged out of the unassigned workstation, wiped the surface biometrics, and walked to the elevator. The Institute was still quiet around her — the overnight minimum-attention quality of the building continuing, undisturbed, as if nothing had happened.

She pressed the elevator call button and waited.

She was thinking about the code — about the specific intelligence of it, the patience involved in writing something and leaving it where it would be found under exactly the right conditions, trusting that the version of herself who came after would be thorough enough to find it and clear-eyed enough to use it.

She thought: the past version of me trusted the present version of me to be competent.

She thought: I would like to be worth that trust.

The elevator arrived. She got in. The doors closed on the quiet building.

She had seventy-two hours that she had not had an hour ago, and three people who were going to use them, and a theory about human cognition that EIDOLON had classified as an existential threat because it was right.

She rode the elevator down in the pre-dawn quiet and felt, for the first time in this entire sequence of events, not afraid. Not certain — she was under no illusions about certainty. But not afraid.

The work was not finished. But it was continuing.

That, for now, was enough.