The mirror had a hole. Test 6 found it.
Author: Eric Yaka
Body:
I built a persistent KV cache. Five proofs said it was lossless. Isometry, conjugate recovery, cross-distance, relative-offset, persistence all green, all bit-for-bit on this hardware.
Then I attacked it in a way I hadn't thought of. Permute every stored entry across its slots. Values intact. Addresses destroyed.
[before shuffle] recover(10) max err = 5.96e-08 exact
[after shuffle] recover(10) max err = 7.62e-01 garbage
The position was living in the array slot, not in the payload. Five tests passed because none of them moved anything.
Fix: every entry carries its own position tag. The slot is parking. The tag is the address. recover() finds entries by tag and takes the un-rotation angle from the entry itself.
Same attack, after the fix:
[before shuffle] recover(10) max err = 5.96e-08
[after shuffle] recover(10) max err = 5.96e-08 SURVIVED
Six tests green, this machine (RTX, torch 2.8.0+cu126, Windows):
- Isometry 0.00e+00
- Conjugate recovery 0.00e+00
- Cross-distance (4990 writes) 5.96e-08
- Relative-offset invariance 1.00e-05
- Save/load persistence 5.96e-08
- Shuffle-invariance 5.96e-08
Cost of the upgrade: one int64 per entry. The rotation, the isometry, the attention math untouched. Legacy saves without tags still load. Slots become tags automatically.
Run it on CPU, older torch, M-series Mac post your max error. I want to see where the float32 floor moves. Find an attack I missed, open a thread. The mirror is self-describing now. Lossless and order-independent.
Come break it.
This is a really interesting catch. The fact that the original tests passed while the cache was still fundamentally position-dependent is a great reminder that correctness under expected conditions isn't the same as structural correctness. I especially like that the fix preserves the core math while making each entry independently recoverable. The shuffle test is a much stronger proof of robustness than the initial green tests. Solid work.