40 lines
4.0 KiB
Markdown
40 lines
4.0 KiB
Markdown
|
# KHRA'GIXX — QUICK GUIDE (pin this)
|
|||
|
|
**Full detail: `KHRAGIXX_SOURCE_OF_TRUTH.md`. This is the one-page version.**
|
|||
|
|
|
|||
|
|
## THE 9 MISTAKES — check every claim against these
|
|||
|
|
1. **Wrong write channel.** inject_density writes the PASSIVE density moment; memory (if any) lives in the stress moment the collision wipes every step. All injection memory results = VOID.
|
|||
|
|
2. **Noun-hunting.** We always asked "what's stored / does it come back." Wrong. The question is a VERB (see below). Still not fixed.
|
|||
|
|
3. **Fidelity scoring.** corr(readback, template) throws away the signal. The DISTORTION is the signal, not the error.
|
|||
|
|
4. **Un-distortable stimuli.** Dots/blobs can't fuse. Need structured families so reconstruction can show.
|
|||
|
|
5. **Reading corpses.** Checkpoints/telemetry are what it LEFT BEHIND. Behaviour only exists while running.
|
|||
|
|
6. **Global averages.** The 9 telemetry scalars average a million-node field to single numbers — structure destroyed before it leaves the GPU. Every telemetry null is a null about 9 averages.
|
|||
|
|
7. **Loose gates → false positives (caught 6+ times).** Threshold drawn around the value; single-sample estimators; lags that are multiples of the 49-cycle drive; a notch-filter faked as "drive-off." EVERY claim needs: pre-committed threshold, surrogate null, bootstrap distribution, threshold NOT within one noise-width of the value.
|
|||
|
|
8. **Lying file headers.** Daemon copies kept old headers. Trust the file map in the source-of-truth doc, not header comments.
|
|||
|
|
9. **Dead-machinery drift.** Golden Weave, "ask another LLM" = banked nulls. A second opinion must be a MEASUREMENT, never a model's impression of a summary.
|
|||
|
|
|
|||
|
|
## THE VERB WE'RE LOOKING FOR
|
|||
|
|
**RESPOND — history-dependent response.** Poke it the identical way at two points in its history: does it answer DIFFERENTLY, in a structured way that tells you which history happened? Chaotic-different ≠ memory. It's only memory if the response DISCRIMINATES the history (right-history vs wrong-history, above surrogate, pre-committed). **This has never been tested. It is the frontier.**
|
|||
|
|
|
|||
|
|
## WHAT'S CLEAN (trust these)
|
|||
|
|
- Substrate runs (canonical v5, frozen).
|
|||
|
|
- Free-run = one smooth global mode, one slow DOF (coherence↔asymmetry), no distributed spatial structure at checkpoint res.
|
|||
|
|
- **Mass leak: density 1.0→1.52, real bug, moving baseline — contaminates long experiments.**
|
|||
|
|
- Discrete-token / pairwise-chord hunts = clean NULL. The 49-cycle "beat" is just the injected drive echoing.
|
|||
|
|
- ZERO positive memory results exist. We have not shown it remembers anything.
|
|||
|
|
|
|||
|
|
## THE INSTRUMENT (new, 2026-07-14, never used for an experiment)
|
|||
|
|
- `khra_gixx_1024_v5_observer` — probe + fast coarse field stream on **port 5561**: 32×32×6 (rho,ux,uy,sxx,syy,sxy), ~100Hz, 24,592-byte KGCF frames. First channel showing WHERE, at cycle-scale, instead of 9 averages. Verified live.
|
|||
|
|
|
|||
|
|
## OPERATIONAL RULES
|
|||
|
|
- ONE daemon holds ports 5556–5561 — kill others first or bind fails.
|
|||
|
|
- Launch DETACHED (`setsid ... < /dev/null & disown`, poll a log) — executor times out at 30s and kills inline daemons.
|
|||
|
|
- NEVER touch canonical `khra_gixx_1024_v5.cu`. NEVER reset_equilibrium casually.
|
|||
|
|
- Fix the write channel (target stress moment, not density) BEFORE any poke experiment.
|
|||
|
|
- PowerShell MCP is broken; filesystem MCP works for all of D:\.
|
|||
|
|
|
|||
|
|
## NEXT STEP (not yet run)
|
|||
|
|
**Passive predict-test first.** Observer watches live 5561, predicts next frame, scored vs persistence + linear baselines, short windows (leak-safe), pre-committed threshold. Can it beat "next = current"? If no → clean null, language isn't there either. If yes → THEN poke (into stress moment) and test the verb. Do passive before active. Don't skip to poking.
|
|||
|
|
|
|||
|
|
## IF YOU WANT TO BURN IT DOWN
|
|||
|
|
KEEP: canonical v5, the 224-checkpoint corpus, the observer daemon, the two docs. You do NOT need to delete the VOID results — quarantining them (done, in the source-of-truth doc) frees you from them while keeping provenance. Only safe deletions: redundant `_mem*` / `probe_p2*` BINARIES (clutter). Leave all sources.
|