Local partner repair
0 / 0Reconstruct an erased link when its partner still remembers it.
Interactive research / 31 October 2026
What survives
the damage?
Erase a few links. See what the remaining evidence can recover.
This interactive illustration draws on the SAURIA-III erasure audit. It shows a small matching problem, not an AI model or a claim of autonomous regeneration.
Twelve nodes. Six pairs.
Each node remembers its partner. Select a numbered node to erase its stored link. Select it again to restore it. The lines show the original pairs so you can inspect the example; the recovery rules use only surviving links.
Matching-prior baseline
0 / 0Also infer the last two missing links if the known perfect-matching rule leaves only one possible pair.
Results show correct erased links recovered / links erased. This is a fixed six-pair example; the audit enumerated every erasure pattern for a 12-node matching. Coherent false links are not modeled here.
What the larger audit found
Recovery depends on what remains.
4,096erasure subsets enumerated
160 / 220three-node patterns completely recovered by local repair
220 / 220recovered with the matching-prior baseline
The stronger baseline uses additional knowledge: that every node belongs to exactly one pair. Both results are limited to the stated constructed matching problem. Read the audit summary and comparison →
Source: SAURIA-III independent audit, 30 October 2026. The interactive model demonstrates the stated repair rules; it is not a replay of the original experiment.
Further reading
Beyond the illustration
Recent work examines LIZARD’s physical analogy, active-learning baselines, Goldfish adapter evaluations and the limits of recovery from damaged evidence. Each has its own controls and unresolved questions.