File size: 1,406 Bytes
a657395 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 | # Lossless index — value and capacity unchanged
## What would lose capacity
| move | effect |
|---|---|
| reduce D | C_addr and k* both move (k* ∝ D) |
| INT8 / PQ / shorter codes on packed HVs | Hamming distances warp → k* cliffs earlier |
| drop traces | value gone |
A random HV in `{+1,-1}^D` has ~D bits of entropy. You cannot store it
in o(D) bits and keep the same geometry.
## What does not lose either
Traces are already `a = bind(x, ctx)`, `v = token_hv`, with `x`/`ctx`
built from `Encoder._atom` / `_role` (sha256 → `random_hv`).
Store the **recipe** (atom keys + bind/bundle/sequence). Rebuild with
the same Encoder law. Measured here:
```
LOSSLESS_OK D=4096 bits_identical=3/3
```
Bit-identical packed buffers ⇒ identical Hamming ⇒ identical LSH
buckets ⇒ **C_addr and k* unchanged**. The value HV is the same object
in `{+1,-1}^D`.
Disk: integer keys instead of `2 * ceil(D/8)` bytes. D stays 20 000.
## Fallback
If a buffer is not in the Encoder algebra, keep it packed. That
preserves value. It does not shrink that trace.
## Code
`palimseste/lossless.py` — `LosslessCodec.materialize` + `assert_bit_identical`.
The 10.4 GB `memory.bin` is still the packed ULTRA dump. New writes
can be recipes. Converting the old dump is a separate pass: invert
each packed `a` against `encode_sequence` of the training windows.
Until that pass, do not delete packed parts.
|