User Correction

The claimant working group proposed a new release system. Not correction.

Annotation. Users could keep reconstructions while attaching visible provenance layers: directly sourced, inferred, cross-donor, generated dialogue target, unknown lineage.

They could choose how much to see. Default would show uncertainty rather than hide it.

Mara called it "putting Mum's condition into software eighteen years late." Celia called it administratively expensive.

Ivo called it obvious. Nia called it enough to start.

The harder proposal was user correction. If a client knew a detail was wrong, should their correction become new source evidence?

Morrow had often said yes. That created feedback loops.

A person accepted a yellow cardigan, then later told the system yellow was correct, increasing confidence in the false detail. Ivo proposed separating autobiographical preference from factual source.

"You can say, this is how I want the scene represented, without pretending it's evidence."

Elias liked that. Mara did too.

"A memory can have an edit without calling the edit history."

The court approved a pilot. Nia became the first user.

She kept the yellow cardigan. Beside it, the provenance layer read: borrowed visual feature, Bell donor family, not supported by Rosa Vale records.

Nia looked at the scene.

"Still Mum," she said.

Then she corrected herself.

"Still my memory of Mum."

The difference was small. It mattered.

The new correction interface refused to use the word corrected. Mara had won that argument after threatening to withdraw from the claimant group.

"A person isn't a corrupted file," she said. "You annotate what the system did. You do not correct somebody's memory like spelling."

The final design displayed layers.

DIRECT SOURCE.

INFERRED FROM PERSONAL DATA.

CROSS-DONOR PRIOR.

GENERATED FOR CONTINUITY.

UNRESOLVED.

Users could reveal or hide layers. They could attach their own notes. They could mark a detail as emotionally meaningful without claiming factual confidence. Most importantly, the system never silently promoted an accepted generated detail into direct history.

Ivo hated the interface aesthetically. "It looks like tax software."

"Good," Mara said. "Maybe people will stop falling in love with it."

Nia volunteered to test it on her mother scene.

The pear-shaped sugar bowl glowed amber: CROSS-DONOR OBJECT, probable Venn source.

Helen-like hand posture: CROSS-DONOR INTERACTION PRIOR, source family cluster uncertain.

Farah's blouse: DIRECT PHOTO SOURCE.

Kitchen window: DIRECT PERSONAL PHOTO SOURCE.

Dialogue: mixed, mostly generated.

Nia watched the scene with labels visible.

It was less immersive. Less beautiful. More truthful about construction.

"I hate it," she said.

Ivo looked alarmed. "The interface?"

"No. I hate knowing exactly where the joins are."

Mara asked whether she wanted to stop.

Nia shook her head.

The reconstructed Farah reached toward a cup. The hand carried an amber outline.

Nia started crying.

Lena moved toward the stop control. Nia held up one hand.

"I'm okay. I just thought finding out would make the whole thing feel fake. It doesn't. That's harder."

Afterward she chose to keep the scene in her archive with provenance layers available by default.

"Still my memory of Mum," she said.

Elias almost corrected the word memory, then did not. It was her memory of experiencing a reconstruction of her mother. Human language did not need to become a technical manual every time.

The claimant group proposed another feature: a personal annotation field where users could write what they believed before and after provenance release.

That created arguments. Lawyers worried such notes could be mistaken for clinical evidence. Psychologists worried people would over-monitor themselves. Users wanted somewhere to preserve their own changing understanding.

Mara suggested plain text with a warning: THIS IS YOUR NOTE, NOT A SYSTEM FINDING.

Everyone agreed.

She looked disappointed. "I expected more resistance."

Celia said, "We're tired."

The difference between the old and new interface was small in code. Labels remained visible. Acceptance did not change source confidence. Users controlled whether they wanted more detail.

Small design choices had helped create the original problem.

Small design choices could not repair all of it, but they could stop repeating the same mistake.