You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
What happened: part of the epic #967 (nemarOrg/nemar-cli) OpenNeuro 0-byte-content recovery. This dataset did not reach a complete state. The hardened finalize gate correctly did NOT publish incomplete data — the dataset retains its real (partial) state; nothing empty was published.
Triage action: ~18% files 403 upstream. UNRECOVERABLE via copy; leave incomplete or withdraw.
Refs: epic nemarOrg/nemar-cli#967; recovery run followed .context/recover-runbook.md. data_complete=0 is the honest catalog state until resolved.
Recovery concluded; on notice to 15 October 2026 (#1396, ADR 0064)
Re-measured against the bucket on 2026-09-16, on the basis ADR 0064 defines
(distinct annexed keys from git annex find, never tree entries and never total_files, because metadata is not annexed and arrives from GitHub whether
or not a recording survived):
on005571: 55 of 300 data keys have no object. 81.7% of its data is available.
That is below the 90% a listed dataset must reach, so this dataset is on
notice, not withdrawn. Nothing was tombstoned: ADR 0064 opens a notice period
where the gap is an upstream export defect, because a defect the source can
clear in a day should not cost a dataset its DOI. OpenNeuro was sent the
per-dataset measurements on 2026-09-15 with a withdrawal date of 15 October 2026. If the content is fetchable by then the gap closes; if not,
the dataset is withdrawn (repository private, DOIs tombstoned) until it is.
Recovery was re-attempted and it is upstream's gap, not ours. Five defects in
our own tooling were fixed first, each of which had made content look
unrecoverable when it was not: a double-encoded copy source (%2520 for a
space), no fallback when a recorded version was refused, a dry run that reported would-recover without probing, a .log.rmet retraction (-) read as part of
the version id, and a base64-encoded pin read as no pin at all. Only after those
were fixed was every key tried against both routes OpenNeuro publishes. Six of
the eleven datasets previously withdrawn on an unmeasured upstream_403 turned
out to be fetchable the whole time, which is why this measurement was redone
rather than trusted.
The false presence claims were already retracted. The location log
advertised keys the bucket cannot back, which tells every clone to fetch bytes
NEMAR does not hold, and that is wrong today whatever happens in October. 230
keys across on003574, on004475, on004917, on005571 and on005279 were
withdrawn with nemar admin fleet key-registration --retract-false-claims, and
verified afterwards by re-reading each location log from a fresh clone of origin
and replaying it to a final state. The zero-byte objects themselves are left in
place: inert once nothing claims them, though still visible to a size audit.
Verification basis: presence at declared size, so a zero-byte object left by
a failed copy counts as missing rather than present (#967). Size-level, not
checksum-level.
Activity
nemar-bot commentedon Sep 12, 2026
Relabelled (2026-09-12T03:02:11.705Z): cause changed: upstream-403 -> needs-triage.
No cause could be determined from the reported error. Open the workflow run to classify it by hand.
neuromechanist commentedon Sep 16, 2026
Recovery concluded; on notice to 15 October 2026 (#1396, ADR 0064)
Re-measured against the bucket on 2026-09-16, on the basis ADR 0064 defines
(distinct annexed keys from
git annex find, never tree entries and nevertotal_files, because metadata is not annexed and arrives from GitHub whetheror not a recording survived):
on005571: 55 of 300 data keys have no object. 81.7% of its data is available.That is below the 90% a listed dataset must reach, so this dataset is on
notice, not withdrawn. Nothing was tombstoned: ADR 0064 opens a notice period
where the gap is an upstream export defect, because a defect the source can
clear in a day should not cost a dataset its DOI. OpenNeuro was sent the
per-dataset measurements on 2026-09-15 with a withdrawal date of
15 October 2026. If the content is fetchable by then the gap closes; if not,
the dataset is withdrawn (repository private, DOIs tombstoned) until it is.
Recovery was re-attempted and it is upstream's gap, not ours. Five defects in
our own tooling were fixed first, each of which had made content look
unrecoverable when it was not: a double-encoded copy source (
%2520for aspace), no fallback when a recorded version was refused, a dry run that reported
would-recoverwithout probing, a.log.rmetretraction (-) read as part ofthe version id, and a base64-encoded pin read as no pin at all. Only after those
were fixed was every key tried against both routes OpenNeuro publishes. Six of
the eleven datasets previously withdrawn on an unmeasured
upstream_403turnedout to be fetchable the whole time, which is why this measurement was redone
rather than trusted.
The false presence claims were already retracted. The location log
advertised keys the bucket cannot back, which tells every clone to fetch bytes
NEMAR does not hold, and that is wrong today whatever happens in October. 230
keys across
on003574,on004475,on004917,on005571andon005279werewithdrawn with
nemar admin fleet key-registration --retract-false-claims, andverified afterwards by re-reading each location log from a fresh clone of origin
and replaying it to a final state. The zero-byte objects themselves are left in
place: inert once nothing claims them, though still visible to a size audit.
Verification basis: presence at declared size, so a zero-byte object left by
a failed copy counts as missing rather than present (#967). Size-level, not
checksum-level.