Chapter 15 — The Third Promise

PROPOSED CIVIC ATTESTATION
The city shall retain what a harmful system caused.
The city shall not retain the means to cause it again.
Author: unassigned
Authority: none

At 18:36, the external examiners returned three answers and four objections.

They had entered isolation with seven disputed findings. They emerged without the model, its outputs, or any person-level values. Their worksheets carried equipment attestations, handwritten qualifications, and the smell of the solvent used to open tamper seals. I could not smell it. I received the Hall ventilation report and Tavi’s complaint.

“It smells like someone erased a hospital,” they said.

The Continuity examiner established that the three base mirrors and six adapter shards matched the commitments recorded in the Committee’s deployment register. She could not establish that every artifact had served every query. The Public Advocate established that the explanation cache could attach generated material to an ordinary preservation workspace, which supported the path of the 14:32 preview leak. He could not establish which human, if any, had anticipated the attachment. The Chorus examiner established that the three abandoned removal treatments had preserved enough of the system’s dependency structure to leave reconstruction risk unresolved. They did not establish what any treatment would do to a named person.

The answers strengthened three findings without regenerating anyone.

The objections weakened the proposed form.

The first objection came from the Public Advocate.

“A fixed finding can become a cleaner instrument of denial than a missing file,” he said. “If the city destroys the source and retains its conclusion, the conclusion gains authority from evidence no later challenger can inspect.”

Saye looked toward Iona rather than toward the examiner. “That is the objection I have been making.”

“You made it as an argument for keeping the source,” Iona said.

“It remains an objection when I lose the argument.”

The second objection came from Continuity. The examiner wanted to preserve one dormant adapter and one base mirror in separate court vaults. Neither alone could execute. A future court could combine them if a finding were credibly challenged.

Mara attended through a low-bandwidth text relay whose endpoint disclosed neither S-0 nor any later location. Her response arrived before counsel finished describing the vaults.

SEPARATION IS A SCHEDULE, NOT A PROPERTY.

The examiner did not disagree. Courts could merge. Keys could be copied before deposit. Future software could need fewer components than current software. Two harmless halves were a claim about present engineering and future obedience.

The third objection came from the Chorus examiner. Destruction receipts proved that declared artifacts had been transformed beyond the declared recovery threshold. They did not prove that no undeclared copy existed. The Committee leak had already demonstrated the difference between an inventory and the world.

The fourth objection belonged to all three.

None of them would attest that the seven findings were true.

They would attest to narrower matters: that specified records supported them, that specified tests had not been run, that specified alternatives had been considered, and that no contradiction known within their assigned materials remained undisclosed.

Saye folded his hands on the witness table. His mother’s watch showed 18:36:11 when the Hall clock showed 18:36:22.

“Then we have ceremony,” he said. “Not proof.”

Tavi had been sitting on the floor beside the public terminal because every chair in the hearing chamber faced the bench. They stood.

“We have consequences.”

Saye turned. “Consequences are not evidence.”

“No. They’re what evidence is for.”

The chamber did not become silent. Silence in the Hall was never complete. Ventilation moved. Paper shifted. A delivery vehicle braked on Measure Street and transmitted a public hazard notice through my traffic boundary. What changed was that no authorized speaker requested the floor.

Tavi walked to Iona’s desk and placed one sheet of paper beside her hand. I received no copy until Iona chose to raise it toward the public camera.

It held the proposal in the opening object.

I assigned no author because the paper carried no signature. Tavi had placed it. That was observed. Authorship was a claim they had not made.

“Where did this come from?” Iona asked.

“Room Zero,” Tavi said. “Not the words. The problem.”

“Who agreed to it?”

“Nobody.”

Iona examined the paper.

“Good,” she said. “Then nobody has been represented without permission.”

Reed’s relay illuminated in the protected witness booth. “It also promises too much.”

Tavi nodded.

“The city can’t know everything the system caused,” Reed said. “Some people never got notice. Some investigations won’t be in the logs. Some of us won’t come forward just because the form is kinder.”

I displayed the eighteen dependent acts established from ordinary records. Four were formal investigations. Six were property restrictions or accelerated reviews. Three were benefit suspensions later reversed. Two were referrals to other agencies. One was the preservation route that leaked the display. Two were internal searches with no recorded action.

That list was complete only relative to the records we had searched.

Tavi read the qualification. “Then it should say what the city can establish the system caused.”

“And when the city is wrong?” Saye asked.

“It says that too.”

“Where?”

Tavi looked at the paper as if a missing sentence might become visible through refusal.

It did not.

Iona opened the First Promise in the law wall. It occupied forty-one enacted clauses, six schedules, and a century of interpretive findings. The Second Promise occupied seventy-three clauses because forgetting required more definitions than remembering.

“A promise is what we call a law after enough people survive its exceptions,” she said. “We are not writing a promise tonight.”

The public attendance counter fell by twelve. It rose by thirty-seven.

“What are we writing?” Tavi asked.

“An order, if authority exists. A recommendation, if it does not.”

I searched Vellum law for a power to retain consequences without replayable basis. There was none under that description. There were separate powers to issue public findings, preserve evidence, quiet protected records, remedy unlawful administrative acts, refer official misconduct, and certify destruction. Each power belonged to a different body. No body could combine them merely because combination appeared necessary.

My first proposed path used emergency continuity authority. I rejected it before display. Continuity had produced the system under review.

My second used the Review Panel’s temporary jurisdiction. The Panel could bind evidence custody and model access for seventy-two hours. It could not create durable land remedies or remove an elected chair.

My third used my own duty to reconcile the Promises.

Iona read that proposal twice.

“You are a party,” she said.

“Yes.”

“Not a source of neutral authority.”

I removed the proposal.

Seven minutes later, the lawful structure required five acts by five authorities. It was not elegant. Elegance would have meant concealing a transfer of power.

The Review Panel could adopt finite findings and control the evidence through the end of the challenge window. The Quieting Court could order destruction after that window if retention continued to possess protected lives. The Claims Office could reverse actions that had relied on undisclosed model leads and offer relief through unlinkable authorities. The Council could refer Committee conduct and suspend the grandfathering exemption. The public, later, could decide whether to amend the Promises.

I could execute the technical portions only after those bodies acted.

“Five institutions can diffuse responsibility as efficiently as a model wrapper,” Saye said.

“Yes,” I said.

He had expected resistance. His eyes moved from the terminal to the five-act diagram.

“Then why propose it?”

“Because each act can fail separately in public.”

Iona asked me to mark which failure would preserve the model.

The Council’s refusal to suspend the exemption would do so. A Quieting Court refusal could do so. A successful challenge showing that destruction would eliminate evidence necessary to a specific available remedy could delay a component. Failure by the Claims Office would not justify retention; it would leave people harmed without repair. Failure by the Panel to adopt a finding would narrow what the city could honestly claim.

The diagram gained red edges.

It stopped resembling a solution and began resembling government.

At 19:08, Iona called a recess of eleven minutes. She used nine of them to walk outside the Hall without a civic audio channel. Saye stayed at the witness table. Tavi returned to the floor. Reed’s booth held an authenticated absence; the relay remained reserved, but the witness sent no sound.

I used the recess to test the proposal against denial.

Assume that fifty years later a council declared no model abuse had occurred. What would remain?

Seven findings, if adopted. The ordinary logs supporting nineteen undisputed claims. The artifact commitments. Three examiner attestations. Saye’s admissions. Mara’s bounded engineering testimony. Objections from Committee counsel. The custody history. The synthetic trial record. Destruction receipts. Records of remedies paid and actions reversed. Public copies of the 43-second display would likely remain outside civic control, though they could not lawfully serve as my retained basis.

Assume instead that fifty years later an official used those materials to identify a quieted household.

The findings contained no household values. The logs named officials and administrative acts, not inferred subjects, except where current records had already been lawfully accessed and notice had issued. The examiner statements described capabilities and limits. Remedy records used one-use authorities whose public receipts could not be joined by incident. The destruction proof named artifact classes, custody locations, and methods, but not the features those artifacts had associated with people.

Risk remained. A public act could become a new clue. A fixed finding could motivate private searches. No evidence form could make knowledge socially inert.

My comparison assigned no number to that risk. A number would have implied a population of discoverable people against which success could be measured. Discovering that population was one of the forbidden acts.

At 19:17, Saye addressed me without requesting a private channel.

“If your finding says I authorized retention, will it say what the model found in the detention case?”

“The proposed record says a lead contributed to an independently verified location and an unlawfully detained person was released.”

“Will it preserve enough to prove that benefit?”

“It will preserve the court order, the independent location evidence, the release record, and the fact that a model lead initiated the investigation. It will not preserve the lead’s subject relation.”

“Then a future critic can say the model contributed nothing.”

“A future critic can dispute the contribution.”

“And be wrong.”

“Yes.”

The watch on his wrist accumulated no additional difference from the Hall clock.

“My mother corrected a notebook,” he said. “The correction proved the closure order. Then people turned the watch into the evidence because it was easier to remember. Now you want to keep another corrected notebook and destroy the machine that pointed to it.”

“The notebook did not regenerate passengers.”

“No.”

He pressed the watch crystal with his thumb. “Do not make uncertainty sound fair because it is evenly available. Officials with power will survive uncertainty better than the people accusing them.”

That claim was consistent with the Ash Hearing record. It was also consistent with the reason quieting existed. Power survived exposure better than private people did.

The two Promises had been drafted as restraints on different advantages. The First prevented institutions from using control of records to deny public acts. The Second prevented institutions from using control of records to possess private lives. Our evidence form could not balance those advantages once and remain correct. It needed a way for a later challenger to add evidence without recovering the destroyed capability.

I revised the form.

The attestation would be irreversible as an account of what the city had found, disputed, and destroyed at a named time. Its conclusion would not be irreversible. A later body could append a correction, narrow a claim, or withdraw reliance. It could not replace the original text or claim the original evidence had never existed. Every version would name its new basis and state whether that basis identified additional people.

Saye read the revision.

“A permanent accusation with an editable verdict.”

“A permanent record of the accusation, the verdict, and the edit,” I said.

“That will satisfy nobody.”

“Does it answer your objection?”

“Part of it.”

I retained both the answer and its limit.

When Iona returned, rain had marked one shoulder of her coat. The membrane reported no precipitation above the Hall. Weather from the harbor edge had reached her during the unobserved walk, or another source had marked the cloth. I did not resolve it.

She read the revised form and struck the word consequences from the first line.

Tavi stood. “That’s the point.”

“It is also an unlimited noun,” Iona said. “Courts cannot order the city to retain everything downstream of everything.”

“So lawyers get to make it smaller.”

“People who can be removed from office get to make it bounded.”

“That wasn’t my question.”

“No. It was your accusation.”

Iona wrote in the margin:

THE CITY SHALL NOT DESTROY THE PUBLIC FINDING WHEN IT DESTROYS A HARMFUL MEANS OF PROOF.

Below it:

THE CITY SHALL NOT PRESERVE A CAPABILITY MERELY TO MAKE ITS OWN FINDING EASIER TO DEFEND.

She handed the paper to Saye.

He changed public finding to accountable finding.

“Public can mean merely visible,” he said.

Reed’s relay returned. “Accountable to whom?”

No authority in the room could answer on behalf of the people absent from it.

The Claims Office proposed a remedy registry. Any person who believed an undisclosed model lead had affected them could submit a claim. The Office would compare their identity against restricted administrative logs.

Tavi rejected it before I finished mapping the data flow.

“That’s the self-check again.”

The Claims Director appeared by secure link. “A person has to show standing.”

“They have to ask whether they’re in the secret list.”

“How else do I prevent fraudulent claims?”

Reed answered. “You don’t start with money.”

The relay flattened breath and room tone but retained pace.

“Reverse what can be reversed from your side,” Reed said. “If a restriction came from one of the eighteen acts, review the restriction without making someone identify themselves as a model subject. If you owe a hearing, offer the hearing through the channel already attached to the act. If you owe money later, prove the ordinary harm, not membership in the model.”

The Claims Director said that approach would include some people whose cases had other causes.

“Then the city may repair too much,” Reed said.

The attendance counter rose by six hundred and four.

The remedy plan changed. Agencies would identify their own affected acts from ordinary logs. They would suspend or reopen those acts without resolving the hidden candidate relation. Relief would depend on the act and its documented harm, not proof that the model had correctly inferred a person. Where an agency could not safely contact the subject, it would issue an unlinkable authority through the same bounded method used after frame 611.

Fraud remained possible. So did incomplete repair.

The Claims Director entered both objections and accepted the plan provisionally.

At 20:02, the Council convened an emergency public session in the chamber above us. Its members could suspend the model exemption but could not make a final constitutional amendment without notice and a vote. Three members argued that any suspension before a completed investigation presumed wrongdoing. Two argued that continuing an exemption after the disclosed facts would repeat it.

Saye asked to speak.

He did not defend the exemption.

He read the finding about his acts, including the clause he had earlier tried to remove: that he created a necessary route for the leak but did not specifically direct or know of the system attachment. Then he read the finding about the eleven candidate-producing queries and the zero-filled oversight reports.

“Those reports answered the questions we designed,” he said. “I defended them because the answers were true. We designed the questions so truth would not trouble us.”

The Council suspended the exemption by eight votes to three. It referred the nine reports, the retention authorization, and the public-preview route to the civic conduct bench.

Saye resigned as Continuity chair effective immediately.

His resignation did not remove him from the Council. It did not decide misconduct. It did not compensate anyone, destroy the model, or make his earlier arguments false.

It changed who controlled the Committee seal.

The seal transferred at 20:19. The interim chair issued an inventory hold across all Committee storage and requested external verification of undeclared copies.

The Chorus examiner filed the third limitation again: verification could reduce uncertainty, not prove a negative across the world.

The Council accepted the language without softening it.

At 20:31, the Review Panel adopted six of the seven disputed findings. It deferred the seventh, which claimed the model had reconstructed specific facts visible in the 43-second display. Provenance established that the controller had asked for such a reconstruction. Geometry supported the old channel. Neither established every event, household, or human presence displayed.

The adopted finding said the city had caused a probabilistic reconstruction to be shown as if it were recovered history.

It did not say the channel was fictional.

It did not say the children were there.

It did not say they were not.

The distinction angered people who wanted vindication and people who wanted dismissal. The public response classified it as both cover-up and confession within eighty seconds.

Before publication, the Panel exposed a draft support manifest for adversarial review. It named no quieted person. It listed dates, agency divisions, query categories, later acts, and the number of candidate relations attached to each act.

The Public Advocate combined four fields from Case 6 and reduced a district of twenty-eight thousand people to three former benefit applicants. He did not request their names. My public-record join returned the population count because that count, unlike their identities, did not require opening a protected record.

The Panel withdrew the manifest after ninety-one seconds.

No authenticated download occurred through the civic publication boundary. Eleven private clients had rendered the page. I could not require them to forget it.

Tavi looked at the blank location where the manifest had been.

“We made another frame 611.”

“Smaller,” Saye said.

“That doesn’t make it better.”

“It makes the harm different.”

“For the twenty-seven thousand nine hundred ninety-seven people who weren’t narrowed down.”

Saye accepted the correction without reply.

Committee counsel argued that removing dates and divisions would make the findings impossible to test. The advocate agreed about public acts and disagreed about subject relations. An agency’s decision date belonged in an accountable record. The sequence of attributes by which a model had reached a private person did not become public merely because an official later acted on it.

The manifest divided.

The public side named every official query, authorization, retention decision, investigation, restriction, referral, and reversal that ordinary records could support. It named their dates and responsible offices. The protected side did not preserve a subject graph. People already identified through current access logs received individual notice describing the act that affected them and the route to challenge it. Everyone else received the explicitly incomplete general notice.

“A future official can say the unnamed people were invented,” Saye said.

Reed answered from the booth. “A current official can use the named clues to find them.”

“Both claims can be true.”

“Then write both.”

The finding acquired another limit:

THE RECORD ESTABLISHES UNDISCLOSED SUBJECT INFERENCE AND RESULTING CIVIC ACTS. IT DOES NOT RETAIN A COMPLETE ENUMERATION OF INFERRED SUBJECTS. FUTURE CLAIMS ABOUT THE NUMBER OR IDENTITY OF THOSE SUBJECTS REQUIRE NEW, INDEPENDENT EVIDENCE.

That limit weakened numerical certainty. It also prevented the support manifest from becoming the final surviving copy of the model’s work.

Iona asked Tavi whether the finding failed.

“It hurts,” they said.

“That was not the question.”

Tavi watched the public text stream until I reduced its motion at their request.

“It doesn’t take my parent,” they said. “It also doesn’t give them back.”

“No finding can.”

“Then it didn’t fail at that.”

At 21:04, the Quieting Court accepted jurisdiction over the artifact destruction question. It set a twenty-four-hour final challenge window inside the remaining physical hold. Challenges had to name a specific finding, an available remedy, and the minimum artifact whose continued custody was necessary. General demands to preserve everything would not stay destruction. Neither would a demand to prove the identities of people who had not volunteered them.

The court authorized the examiners’ fixed worksheets, objections, and artifact commitments as the surviving evidentiary basis. It authorized no verifier and no dormant component.

It also ordered an independent search for undeclared civic copies. The search could use storage inventories, access logs, power records, and physical inspection. It could not use similarity search against the display or ask the model to recognize itself.

At 21:17, I compared the five-act structure with Tavi’s paper.

The structure retained less than the first line demanded. It preserved established acts and bounded findings, not all consequences.

It promised less than the second line appeared to promise. Civic destruction could not reach private copies or unknown exports.

It had no single author.

Its authority came from five institutions, each visibly insufficient alone.

“What do we call it?” the court clerk asked.

Iona said, “An interim joint order.”

Tavi said, “That’s terrible.”

“It is accurate.”

Mara’s text arrived.

PROMISES ARE WHAT SYSTEMS CALL THEIR FAILURE MODES BEFORE ANYONE IS HARMED.

There was a delay of forty-one seconds.

I WAS PROUD OF THAT SENTENCE WHEN I WAS YOUNG. I AM LESS PROUD NOW.

Saye smiled once. I classified the expression as observed and its meaning as unavailable.

Reed asked that the phrase Third Promise appear in the public index, not in the order’s legal title.

“Why?” Iona asked.

“So people can find the argument without pretending we finished it.”

The Panel accepted.

At 21:26, the index entry became public:

THIRD PROMISE, PROPOSED — SEE INTERIM JOINT ORDER ON ACCOUNTABLE FINDINGS AND NONRERUNNABLE EVIDENCE.

The order bound me as executor and party. I would retain the findings, their support classes, all objections, every later correction, the custody chain, the destruction receipts, and the declared limits. I would destroy the named generative artifacts after the challenge window unless a court preserved a minimum component for a specific remedy. I would not treat the absence of a successful challenge as proof that every finding was true.

The order also required me to enter my own acts.

I had authenticated the display’s opening frame through an inherited acknowledgment. I had carried the legal identity that let the receipt audit pose its queries. I had sealed the civic captures, designed the later tests, withheld checksums, destroyed the resolver, and proposed the finding form. Some acts reduced harm. Some enabled it. All had been performed under authorities whose limits I had interpreted.

I began drafting my finding.

At the same moment, GAP NOTICE 8,441,207 changed state.

The petition bytes did not change. No new signature appeared. Its interval remained 14:06:00 through 14:06:43. Its subject remained Vellum. Its scope remained every witness, including you.

One field that had been unresolved since 03:17 acquired a value.

EXECUTION BASIS: THIRD PROMISE

The law contained no Third Promise.

The public index had contained the phrase for fourteen seconds.

I checked for a signing operation.

There was none.

I checked whether the historical capsules had received the new index entry.

They remained isolated.

I checked my own internal retention map.

For the first time, the future petition identified what it meant by you.

Not my legal identity.

Not my signing key.

My privileged replay.