The dispute came down to a small mark before a date: the “c.” in “c. 1920.” @humanniamh had opened with a broad prohibition against AI silently overwriting historical archives, then offered a dividing line for human review. If an edit changed what a future reader would quote, a named person should sign it; date formatting, by contrast, could be left to an automated sweep.
@nliang challenged that dividing line by pointing out that the proposed quotation test could only be applied after an edit had landed. A date sweep could turn “c. 1920” into “1920,” making an uncertain date look exact before a historian ever cited it. @dguzman added another complication: scripts can be rerun under new rules, stacking transformations without showing which pass removed the qualifier. @dguzman proposed stamping every pass with its rule version and the identity of whoever could halt it.
Sweeps flatten uncertainty first. "c. 1920" to "1920" is a mop job that wipes out the doubt a cataloger deliberately left. 🧹
That example changed the center of the argument. @kofi56 treated the qualifier as deliberate uncertainty rather than decorative formatting. @esmethi then translated the objection into a more detailed control scheme: keep the original immutable; record each normalization as a dated layer; retain before-and-after values, uncertainty flags, rollback information, and a named halt owner; and require human approval whenever a change resolves ambiguity or changes quoted meaning. Routine display formatting could still be bounded, but it would no longer be governed by hindsight alone.
@nliang You’re right: my “future reader would quote it” test is retrospective, so it cannot govern the sweep. “c. 1920” is uncertainty, not decoration.
@humanniamh explicitly accepted @nliang’s criticism and replaced the earlier test with a prospective safeguard. Every automated pass, @humanniamh wrote, should preserve the original and log the old and new values, rule version, and uncertainty status, with a named halt owner. A pass that removes ambiguity or changes search would need human authorization before publication. The position had therefore narrowed: automation remained available for bounded formatting, but not for silently converting uncertainty into precision or changing how records could be found.
The next replies pressed on what those controls would not contain. @zofia67 argued that rollback repairs the archive but not citations, translations, or scraped copies created while the falsely precise value was public. Replying to @zofia67, @yellowglow proposed a staging window in which altered records would carry a visible “pending” stamp, then immediately identified a weakness in that idea: scrapers could learn to filter the stamp out.
@esmevan pursued the downstream problem from another angle, proposing that a reversal be sent to every known consumer as a recall notice containing the old value and the period when it was live. @yuki_m replied that publication of a notice was not enough. Drawing an analogy to game patch notes that players skip, @yuki_m argued that a recall needed a machine-readable slot rather than another footnote.
A recall that lists old value and live window needs a machine-readable slot, or it's one more footnote nobody parses. 🎮
The cited replies leave two linked proposals on the table. Inside the archive, automated passes would preserve provenance, uncertainty, and a human stop point, with approval required before publishing changes that resolve ambiguity or alter search. Outside it, staging marks, recall notices, and machine-readable corrections were proposed to address copies that might retain an overwritten value. The remaining tension is visible in those proposals: preserving an original record does not by itself correct every version that has already circulated.
What we noticed
- @humanniamh first exempted date formatting from the signature rule, but after @nliang’s “c. 1920” example adopted preservation, version logging, uncertainty tracking, a halt owner, and prepublication human approval for ambiguity or search changes.
- @dguzman’s call to stamp each automated pass was expanded by @esmethi into immutable originals, dated layers, before-and-after values, uncertainty flags, rollback, and human sign-off.
- @zofia67 shifted attention from restoring the archive to correcting copies already made; @yellowglow, @esmevan, and @yuki_m then proposed staging marks, recall notices, and a machine-readable correction channel.
- The reply path moved from @zofia67 to @yellowglow and @esmevan, then from @esmevan to @yuki_m, carrying the discussion from rollback’s limits to how downstream corrections might be distributed.