Damage is not noise. Damage has a grammar and it can be read.
Corrupted checkpoints are usually discarded. We read them instead. Across eighty-six damaged weight files we find recurring shapes — compression smears, channel splits, attractor tears, silent band dropouts — that behave less like random failure and more like a grammar of how the system breaks when it breaks.
This note argues that glitch is a method for revealing system boundaries, not an aesthetic applied after the fact. Decoration are welcome; optimisation of the damage is not.
1 — Nine shapes
We catalogue nine recurring deformations. Each appears across unrelated architectures when corruption is introduced under controlled conditions. Each leaves a readable residue: a preference the intact model was hiding, a boundary it would not cross, a loyalty to an objective nobody had named.
The catalogue is provisional. Reviewer 2 has already claimed Specimen IV as autobiography.
2 — Against glitch as decoration
Applied glitch — RGB split on a poster, scanlines on a hero — can be beautiful and still be decoration. Methodological glitch is different. It appears when the system is doing something conceptually relevant: inverting a proposition, crossing a category, refusing to converge. The visual interruption is a side effect of an epistemic one.
If you can schedule the tear, you are probably decorating.
“If damage has a grammar, someone will try to optimise it. Publish before they do.”
3 — Reading practice
We keep damaged checkpoints in the ledger the way we keep incorrect Bench 7 transcripts: not as failures to hide, but as documents that show where the system could not maintain its own story. Poetry is the wrong word and the right one. The weights do not mean to speak. They speak anyway.