HARD PROBLEMS · 2026 · RECOVERY RECORD · SCIENTIFIC RESULT INCOMPLETE
The defensible result of this recovery was not “364 jobs recovered.” It was the decision to discard that progress figure as data. A surviving conversation reported 364 of 864 jobs complete, but the raw outputs, exact manifest and item-level hashes were absent. There was no trustworthy job 365 from which to resume.
Dijiang v0.4 was therefore set up as a new calculation lineage. A recovery audit, frozen specifications, hash locks and checkpoint tests were built. The target calculation and its scientific decision remained unfinished.
Old progress
conversation record only
Raw outputs
not recovered
v0.4
new lineage from zero
External validation
none
A count survived; the evidence behind it did not
The recovery audit began by listing what could not be found: primary outputs, the complete job manifest, the completed-cell list, the input registry and the hashes that would tie those objects together. A historical link in a conversation was not a durable archive.
That makes 364/864 a provenance statement about an earlier attempt, not a dataset available for analysis. Simple subtraction can produce a remaining count; it cannot identify which inputs ran or establish a safe continuation point. The audit makes that boundary explicit.
Recovery became a prospective design
For v0.4, the specification and job manifest were fixed before target outcomes could be inspected, then bound to content hashes. An incomplete job could resume only when its checkpoint still matched its job specification. A mismatch required a restart from the frozen input.
Internal automated checks and a deliberately interrupted checkpoint-resume test passed. Those tests support the mechanics of the runner and recovery path. They do not establish the target result, the underlying scientific claim or portability to another operator’s environment.
When provenance breaks, the honest recovery may begin by refusing to recover the progress number.
What remains, and what does not
The surviving artifacts describe the loss, the new protocol, the hash lock, the runner, checkpoint rules and internal tests. The missing numerical values and trends did not reappear. The old values were barred from the new analysis and from parameter selection.
No independent reproduction or external scientific review has closed the new lineage. Its final outcome is a fail-closed recovery decision, not a scientific finding. A result post would require completion under the frozen protocol and an independent check.
Project record
- Period: 2026
- Activity: loss audit, new execution specification, hash locking and checkpoint recovery design
- Internal evidence: recovery audit, status record, manifests, automated checks and resume test
- External validation: none
- Final outcome: the historical 364/864 report was excluded from analysis and v0.4 began as a new lineage; the scientific result is incomplete
- Not established: lost item-level values, any old numerical trend, the final target decision or independent reproduction