Research-Stack/docs/compression_signal_shaping_synthesis.md
2026-05-08 14:50:03 -05:00

317 lines
5.7 KiB
Markdown

# Compression Signal-Shaping Synthesis
Date: 2026-05-08
Runner:
```text
4-Infrastructure/shim/compression_signal_shaping_synthesis.py
```
Receipt:
```text
4-Infrastructure/shim/compression_signal_shaping_synthesis_receipt.json
```
Receipt hash:
```text
fc06057b20dc2281161e7380a63557171ab4d87ee7a82277fe2e8d74b1446f68
```
## Primary Read
Across the local compression, compressed-sensing, signal-root,
semantic-topology, and docmd payload notes, the new pattern is not another
universal compressor.
The new pattern is:
```text
signal-shaped route compiler
```
Shape the route space before coding, then pay exact residual, witness, decoder,
and container bytes after coding.
## Approach Classes
```text
PAQ-style context mixing
shapes probability context
decision-diagram route search
shapes candidate route space
T16 candidate pipeline
shapes weak event detection
Phi response-family selection
shapes response curve
nonlinear compressed sensing
shapes regular nonlinear measurement maps
generative compressed sensing
shapes latent proposal manifold
invertible generative inverse priors
shapes invertible / flow charts
holographic fractional recursive fold
shapes boundary descriptor and bounded memory
signal invariant roots
shapes signal morphology feature space
semantic topology regimes
shapes fold / prune / tear decisions
LLM control-plane compression
shapes prompt / logogram / metaprobe representation
docmd static payload strategy
shapes runtime payload
```
## What New Pops Up
### N1 Signal-Shaped Route Compiler
Combine signal invariant roots with decision-diagram route search:
```text
chunk
-> feature vector
-> route family
-> codec trial
-> exact residual
```
Candidate equation:
```text
route =
argmin_r LB(r | phi_signal(chunk), topology_regime, history_state)
```
First test:
```text
wiki8 chunk sweep with:
entropy
XML tag density
DCT energy
transient edges
autocorrelation
cosine reuse
```
Promotion gate:
```text
chosen route beats bz2 / zstd baseline after feature and witness bytes
```
### N2 Runtime Staticization As Compression Prepass
Use the docmd lesson:
```text
do not ship branches you can rebuild
```
Candidate shape:
```text
tiddlers / articles
-> static route pages
-> external search index
-> manifest
```
First test:
```text
build a small TiddlyWiki / article slice as both live and static outputs
compare initial gzip payload and search-index cost
```
### N3 Witness-Budgeted Latent Route
Generative and invertible models should only propose routes:
```text
latent z proposes transform
exact residual repairs
uncertainty decides hold
```
Cost:
```text
C =
bytes(z)
+ bytes(model_id)
+ bytes(residual)
+ bytes(witness)
+ bytes(decoder)
```
Promotion gate:
```text
C < incumbent
and decoded hash equals source hash
```
### N4 Fractional History Route Scheduler
Use bounded memory for nonstationary corpus regions:
```text
h_t =
sum_{tau<t, window W}
K_alpha(t - tau) * residual_tau
route_t =
R(chunk_t, h_t)
```
Promotion gate:
```text
history bytes counted
and total compressed size improves
```
### N5 Topology Regime Guard
Use semantic topology before tokenbook merges:
```text
beautiful -> fold
ugly -> prune
horrible -> isolate / hold
```
Candidate classifier:
```text
regime =
classify(invariant_overlap, torsion, round_trip_loss, contradiction)
```
Promotion gate:
```text
fewer bad merges without losing byte wins
```
### N6 Physical Signal Probe Feedback
Borrow the CAD-force-probe habit for compression routes:
```text
route hypothesis
-> measurable perturbation
-> negative control
-> receipt
```
Promotion gate:
```text
positive route beats baseline
and matched negative control fails or underperforms
```
## Unifying Equations
Signal feature vector:
```text
phi_signal(c) =
[
H(c),
tag_density(c),
DCT_energy(c),
transient(c),
autocorr(c),
cosine_reuse(c)
]
```
Route selection:
```text
r* =
argmin_r LB(r | phi_signal(c), semantic_regime(c), history_state)
```
Exact cost:
```text
C_total =
bytes(payload)
+ bytes(sidecar)
+ bytes(residual)
+ bytes(decoder)
+ bytes(witness)
+ bytes(container)
```
Promotion:
```text
promote iff
H(decode(r*)) == H(source)
and C_total < incumbent
and failure_rules == none
```
Negative control:
```text
valid_gain iff
C(candidate) < C(baseline)
and C(candidate) < C(matched_bad_route)
```
## Immediate Experiment Ladder
```text
E1 wiki8_signal_feature_baseline
extract per-chunk signal features and compare feature clusters to codec outcomes
E2 route_classifier_without_new_codec
choose among existing routes only: raw, bz2, zstd, xml_token+bz2, tokenbook+bz2
E3 topology_guard_tokenbook
apply semantic / topology guards before tokenbook merge
E4 docmd_static_wiki_slice
export a small tiddler / article slice to static pages plus external index
E5 bounded_history_scheduler
route stream chunks with finite fractional residual memory
```
## Failure Rules
```text
feature score treated as byte gain -> invalid
sidecar / witness / residual / decoder bytes omitted -> invalid receipt
latent or generative prior used as hidden source payload -> invalid
semantic merge without round-trip / contradiction check -> hold
history kernel unbounded or uncounted -> fail closed
docmd-style staticization reported as Hutter compression -> overclaim
negative controls omitted from new route claim -> weak claim
```
## Claim Boundary
This synthesis proposes testable route-shaping experiments. It is not a Hutter
Prize result, not proof of a new compressor, and not a guarantee that signal
features will improve wiki8.