The certification suite was empty on Saturday mornings except for Reyna and the machines.
That was why she came. The RECOL unit in room 4 had been decommissioned for upgrade — the v3.8 firmware had shipped in late January with revised coherence weighting, and the queue for recalibration ran three weeks long. But the machine was still accessible in read-only mode, which meant Reyna could pull archived session data without generating a new extraction timestamp. No new record. Just old ones.
She had requested the Holloway files eleven months after the trial closed.
The request itself had required a Category II clearance, which she had, and a practitioner justification, which she had written in four sentences on a Tuesday and forgotten about by Thursday. The clearance came through on a Friday. She had not come in until now.
The RECOL machine in read-only mode looked exactly like the RECOL machine in session mode, which was part of the problem. The membrane hood still hung above the chair. The coherence monitors still glowed amber, cycling through their idle state. The system still ran its quiet ambient poll, checking for an extraction-ready signal that would not come, because nobody was sitting in the chair.
Reyna sat across from it, not in it.
She opened the Holloway file on her access pad. Three documents. The extraction summary, certified and stamped with the court seal and the session hash. The Variance Log, flagged, unsigned by the presiding attorney, bearing the notation: supplementary documentation, not entered. And a third document she had not seen before — a procedural brief filed by the defense counsel, dated two weeks after the trial concluded, citing the grounds on which practitioner observation had been excluded from primary evidentiary consideration.
She read the brief.
It was four pages. The first two pages were legal procedure, which she skimmed. The third page cited precedent. The fourth page cited the precedent again, more specifically this time, and named the establishing case.
Torres v. RECOL Certification Board, 2028.
Reyna set the access pad down on the desk.
She had not thought about Torres v. RECOL in three years. It had been a different kind of case — not a criminal trial, not testimony, but a certification dispute. A practitioner in the Recovered Corridor had submitted a Variance Log as standalone evidence in a civil matter without an accompanying extraction, on the grounds that the subject had died before a session could be completed. The board had ruled against it. The ruling had been contested, appealed, and denied. The precedent established by the denial was narrow but specific: practitioner observational records are admissible as supporting documentation to a certified extraction; they do not constitute an independent record of recall events absent an extraction baseline.
Reyna had testified in that case. She had been an expert witness for the certification board.
She had argued — carefully, with citations, in professional-standard vocabulary — that a Variance Log without an extraction baseline was methodologically incomplete. That the Log recorded the practitioner's interpretation of what the subject experienced during a session, and that without the extraction data to anchor that interpretation, the Log was not a record of the subject's memory. It was a record of the practitioner's observation of a subject who might or might not have been accessing the relevant memory at the time.
She had been right, technically. She had also been thorough. The brief sitting on her desk now quoted her at length.
The RECOL machine across from her cycled through another idle poll. Extraction-ready signal: absent. Coherence baseline: nominal. Awaiting subject.
In the Holloway case, the extraction had been clean. Coherence 97.3%. Duration 44 minutes. The court had seen a woman in full control of her account — measured responses, voice steady, physiological indicators within certified range for a composed witness. That was the record. That was what the machine had produced.
Reyna's Variance Log entry for the same session was three lines.
Extraction: no physiological deviation noted during initial questioning. Log: Subject's hands were shaking before the first question. Stopped by the third. Variance class: behavioral/physiological. Significance: flagged.
The machine had not been wrong. Naomi Holloway's extraction had been coherent. Her memories of the event in question — the ones the court was there to hear about — had rendered at 97.3% fidelity, which was as clean as Reyna had ever seen in a case involving active trauma. The machine had done exactly what it was built to do.
The machine was not built to notice hands.
Reyna picked up the access pad and went back to the brief. The defense had argued, citing Torres, that the Variance Log represented Reyna's interpretation of Holloway's behavioral state during session preparation — not during extraction, not during the memory event itself, but in the three minutes before the session began, while Holloway was seated in the membrane hood waiting for the session calibration to complete. The shaking had stopped before extraction started. The extraction had run clean. The Log documented something that happened outside the evidential window, and the Torres precedent said that without an extraction anchor, that documentation could not stand on its own.
The judge had agreed.
Reyna had not been in court that day. She had submitted her session documentation and her Variance Log, received a notification that the Log had been accepted as supplementary documentation, and gone on to three other cases. She had followed the Holloway verdict in the news. She had noted the verdict. She had filed it.
What she had not done — what she was doing now, in the empty certification suite on a Saturday morning eleven months later — was read the defense brief that explained, in four pages with citations, how the framework she had helped establish was the framework that had made the Log inadmissible.
Torres v. RECOL Certification Board, 2028. Torres, R. (expert testimony for certification board).
Her own name in the footnote.
She stood up. Walked to the window. The certification office was in a converted industrial building in what used to be the Corridor's manufacturing district, before the Corridor had been rezoned as a RECOL services hub in 2029. The view from the window was four other certification suites in adjacent buildings, all empty on Saturday, all glowing amber through their windows.
She was thinking about what Naomi Holloway's hands had looked like.
She had been a forensic recall practitioner for seven years. She knew what it looked like when a subject was afraid and trying not to show it. She knew the specific quality of stillness — deliberate, effortful, costing something — that a person produced when they had decided to be composed. Holloway had been that kind of still. The shaking had stopped when the membrane hood descended. Not because she was calmer. Because she had decided.
The machine had recorded the decision. Reyna had recorded the cost of making it.
She went back to the desk.
She opened a new file on the access pad. Not the Holloway file. A fresh document with a blank header. She typed at the top: On the Admissibility of Pre-Session Behavioral Observation: A Proposal for Revised Classification.
Then she stopped.
She had been here before — not this specific desk, not this specific case, but this specific moment of believing she had found the argument that would change the procedural framework. She had written proposals. She had submitted them to the board. The Torres case had started as a proposal she had supported, in fact — a proposal to clarify the relationship between extraction data and practitioner observation, because the ambiguity was causing inconsistencies across the Corridor's certification suites. The clarification had resolved the ambiguity, as clarifications tend to do, by choosing one reading over the other.
She had chosen the one that made sense to her at the time.
The RECOL machine cycled again. The amber light shifted, caught the edge of the membrane hood, and made a brief shadow on the far wall.
She thought about the sentence she had written in the Holloway margin: the machine remembered what she showed. She remembered what she felt.
The problem was that this was also true in cases where the Log was wrong. She had a file — she kept it, as a kind of counterweight — of variance logs where the practitioner's observation had been inconsistent with the extraction data in ways that had nothing to do with the machine's limitations. Cases where the practitioner had interpreted anxiety as guilt, or stillness as composure, or composure as innocence. Cases where the Log had been a record of the practitioner's assumptions about the subject, not the subject's actual state. The machine was calibrated. The practitioner was not.
She was calibrated — she believed this — better than most. Seven years. A clean variance record. No successful challenges to her log entries in the Corridor's review system.
But she was not calibrated like a machine.
The proposal she was about to write would argue for elevating pre-session observation to a distinct classification — not part of the extraction record, not a Variance Log addendum, but a separate documentation category with its own admissibility standard. It would require the board to create new criteria. It would require courts to accept a form of evidence that had not existed as an independent category before. It would require someone to trust that what she had seen — what practitioners like her saw, in the minutes before the membrane descended — was worth trusting.
She had spent six years making the argument that it was worth trusting, in individual session notes, in testimony, in informal guidance to newer practitioners.
She had also spent three days in 2028 making the argument that it wasn't enough on its own.
She sat with the blank document for a long time.
Then she typed a second line: This proposal does not argue that practitioner observation should replace extraction data. It argues that the current classification system cannot account for what happens in the window between subject arrival and session commencement — a window the Torres precedent renders legally invisible, and which contains information about the subject's volitional state that the extraction is not designed to capture.
She saved the file. Did not close it.
Before she closed the Holloway file, she did something she had not planned to do.
She opened the extraction playback.
The playback was not the same as the extraction itself. The extraction was a structured render of the source memory — the architecture, the sensory data, the encoded emotional valence. The playback was the practitioner's interface: a flattened audio-and-notation record of the session, timestamped at each phase. You could not re-experience the memory from the playback. You could hear the practitioner's voice and the ambient room sound and, occasionally, the subject's voice when they spoke aloud during deep recall.
Reyna had never listened to her own session playbacks. It was considered bad practice — it introduced retrospective bias into future sessions, or so the guidance said. You documented in the moment and moved on. The Log was the record.
She put on the room audio and let the session run.
The playback opened with the standard calibration sequence — two minutes of ambient tone while the membrane calibrated to Holloway's neural signature. Then her own voice: Subject presents for session certification. Holloway, Naomi. Session 7. Coherence baseline check running. A pause. The machine's response, rendered as a low tone that meant baseline nominal. Then her own voice again, quieter, not the official notation voice: Take your time. You don't have to be ready yet.
She heard something she had not noted in the Log.
Naomi Holloway had exhaled.
It was not the sound of someone who was afraid. It was the sound of someone who had been waiting for permission and had just received it. One breath, through the nose, slow. The machines didn't mark it — not a physiological deviation, not a coherence event. Just a breath.
And then the session had started. And Holloway had been composed.
Reyna turned off the playback.
She sat in the empty certification suite for a while longer. The amber light cycled. The machine waited.
She had documented the hands. She had not documented the breath. She had not documented it because she had not thought it was significant. She had thought the hands were the story — the fear, the effort, the decision to be composed. She had been looking for evidence of distress and she had found it, and she had flagged it, and the flag had been right.
But the breath was the moment Holloway had decided to proceed.
That was not in the extraction either. The extraction had caught the memory. The Log had caught the fear. Neither record had caught the decision.
She added a line to her proposal draft: The window between subject arrival and session commencement contains, at minimum, three distinct categories of behaviorally significant data: physiological distress indicators, volitional state transitions, and subjects' responses to practitioner communication. Current classification systems document the first category when practitioners flag it and discard the other two entirely. This is not a gap in technology. It is a gap in what we have decided counts.
She saved the file. Did not close it.
The RECOL machine ran its idle poll one more time. Extraction-ready signal: absent. Coherence baseline: nominal. Awaiting subject.
She picked up her bag and left the machine waiting.