KHRAGIXX_HANDOFF_TESTING_REGIME.md

# KHRA'GIXX — HANDOFF & TESTING REGIME (v3)
**Written 2026-07-15. For the next context window and for oversight-Claude.**
**Read this AND `KHRAGIXX_SOURCE_OF_TRUTH.md` before doing anything.**
**The human is opening the next window to discuss a COMPLETELY DIFFERENT METHODOLOGY. Do not assume the regime below is what they want — it is the corrected state of the OLD approach. Listen to the new direction first.**

---

## 0. THE ANCHOR — rules that supersede everything

**(A) THE OBSERVER IS PART OF THE PHYSICS, NOT AN INSTRUMENT ADDED LATER.**
Reading the lattice perturbs it (the read is an intervention). A measurement WITHOUT the observer measures a DIFFERENT SYSTEM than one with it. The observer is present FROM CYCLE ONE of every experiment, always, hardcoded — never bolted on, never optional. A run without the observer and without full vitals capture is INVALID. The observer keeps its OWN PERSISTENT MEMORY/chronicle across runs and context windows, because oversight keeps evaporating between sessions and the same mistakes repeat.

**(B) THE CORE REALIZATION THAT ENDS THE OLD APPROACH — a hollow poke cannot leave substance behind.**
We spent this session building toward a "perturbation hold-time" measurement: inject a poke, measure how long it persists. This is WRONG at the level of the QUESTION, not just the instrument. A featureless Gaussian poke (density OR stress) has NO STRUCTURE. Asking "how long does it persist" measures how long the fluid takes to swallow a dent — a relaxation/viscosity property of the fluid, NOT evidence that anything informational is held. You cannot expect substance to stay when you injected no substance. Hold-time of a structureless poke was always going to answer a physics question dressed up as a memory question.

**The correction:** the memory question requires STRUCTURED, DISTINGUISHABLE inputs (at minimum A vs B), and the measurement is DISCRIMINATION — after time T, can you recover WHICH structured thing went in, above the natural-breathing floor — NOT persistence. And because the substrate is a driven limit cycle that reconstructs rather than stores, "does the dent stay there" is doubly wrong: nothing stays anywhere. The only coherent question is whether injecting structured-A vs structured-B leaves the field's ONGOING BEHAVIOUR differently biased in a way you can read out later. That is history-dependent response — the verb. A hollow poke cannot even pose it.

---

## 1. WHAT WE ARE PERTURBATING WITH (verified by READING the kernel, 2026-07-15, not trusting a report)
- `inject_density_kernel`: `f[i] += w_i * delta_rho`. Pure DENSITY moment. Zero non-equilibrium content. WRONG channel (Mistake 1). A hollow poke. Only ever valid as a dead-control null-standard.
- `perturb_stress_kernel`: `f[i] += G * d_cx[i] * d_cy[i]`. VERIFIED by hand: density added = 0, momentum = 0, sxy shear = 4G ≠ 0. A CLEAN pure-shear non-equilibrium write — the correct CHANNEL. BUT: still a structureless Gaussian blob. A better-placed poke, still a hollow poke. Not information.
- **NEITHER is an analogue of information.** Both are pokes, not messages. A smooth blob carries nothing to distinguish, nothing to recover. Information requires structure + at least two distinguishable variants. We have NOT built that. If anyone describes either injection as "data" or "encoding," that is the conflation to kill on sight.

---

## 2. THE MISTAKES — every one made on this project. Re-read before every session.
1. **Wrong write channel** (density vs non-equilibrium). Fixed in the kernel; but see #16.
2. **Testing for NOUNS.** The system is a VERB — history-dependent response. Never noun-hunt.
3. **Fidelity scoring.** The DISTORTION is the signal, not the error.
4. **Un-distortable stimuli.** Dots/blobs can't fuse; need structured families.
5. **Reading corpses.** Checkpoints/telemetry are what it LEFT BEHIND; behaviour only exists while running.
6. **Global averages.** 9 scalars average a million-node field. (Coarse 5561 stream partially fixes: 32x32.)
7. **Loose gates → false positives (6+ times).** Threshold drawn around the value; single-sample estimators; lags = multiples of the drive period; a notch-filter faked as "drive-off."
8. **Lying file headers.** The observer .cu header STILL says "physics kernels UNCHANGED" — it has 18 live knobs + 2 perturbation kernels. FIX THIS HEADER.
9. **Dead-machinery drift.** Golden Weave, "ask another LLM" — banked nulls. A second opinion is a MEASUREMENT.
10. **No S/N baseline before measuring (the June-13 disaster).** A "breakthrough" (9.3-unit spread) evaporated — it was v2 decay residue. On a clean baseline: spread 0.067 vs ~0.5 oscillation, S/N=0.13, below floor. Measure the natural-breathing amplitude and S/N floor BEFORE any perturbation measurement.
11. **Perturbation below the noise floor.** At strength ≤0.3 effects were below noise; strength 0.7–1.0 gave clean response (coherence −24.9 SD, asymmetry +28.1 SD). Show the immediate effect clears the floor first.
12. **Throwaway scripts instead of the standing suite.** The human requested a HARDCODED suite that ALWAYS uses the observer and ALWAYS captures vitals. We ran disposable one-off scripts with no observer, no vitals. THIS IS THE DRIFT.
13. **Broken hold-time methods (3 attempts, 2026-07-15).** (a) distance-from-fixed-baseline measured the MASS LEAK. (b) save/load differential — the LOAD is itself a perturbation; chaotic divergence swamps signal. (c) the dead-control (density) scored IDENTICAL to stress every time — the tell that the INSTRUMENT, not the physics, was being measured.
14. **Trusting reports of code instead of reading the code.** perturb_stress was approved from its description, verified only later. ALWAYS read the kernel.
15. **Determinism never established.** The daemon uses atomics. Whether the same state gives the same trajectory twice is UNKNOWN, and every control-comparison depends on it.
16. **THE DEEPEST ONE — measuring persistence of a hollow poke.** The whole hold-time program measured whether a structureless dent stays. A structureless input cannot hold information; its "persistence" is just fluid relaxation. This is a broken QUESTION, more fundamental than the broken instruments in #13. Do not resurrect hold-time-of-a-poke as the memory test.

---

## 3. KNOWN-CLEAN FACTS (trust these)
- Driven near-inviscid LBM fluid. Left alone: LIMIT CYCLE — coherence orbits ~0.7372–0.7395, ~120-cycle period, asymmetry anti-phase. DRIVEN, not self-animating.
- 18 live knobs verified (omega, khra/gixx amp, wavelengths, freqs, breathing, clamps, equilibrium coeffs). set_param acks ok/rejected correctly.
- Stress-write LANDS in f_neq (kernel algebra verified).
- Live coarse spatial stream on 5561 (32x32x6, ~100Hz) works.
- Short-horizon predictability: linear predictor beat persistence by 21% on the coarse field (one clean pre-committed result). The field flows with MOMENTUM. NOT memory.
- CONFIRMED MASS LEAK: density 1.0→1.52. Moving baseline contaminates long measurements.
- Discrete-token / pairwise-chord / distributed-spatial hunts: all clean NULL.
- ZERO positive memory results. Nothing has shown the system retains or discriminates a controlled input.

## 4. INSTRUMENT STATE (what's built and usable)
- Daemon: `cuda/khra_gixx_1024_v5_observer.cu`, binary in `build/`. Builds under WSL via `build_observer.sh` (NOT Windows — needs libzmq/NVML on WSL system path).
- 18 live knobs via `set_param` on ZMQ 5557. Two perturbation kernels (density = hollow/wrong-channel control; stress = clean f_neq write, still hollow).
- Ports: 5556 telemetry, 5557 commands, 5558 density snapshot, 5559 ack, 5560 stress snapshot, 5561 coarse 32x32x6 stream (~100Hz).
- Workspace is isolated: `D:\resonance-engine-active\` with `.clinerules`. Live daemon runs from the parent `D:\resonance-engine\`. Only ONE daemon on 5556–5561. Launch DETACHED.

## 5. IF THE OLD REGIME IS RESUMED (it may NOT be — new methodology incoming)
The corrected order was: standing suite (observer + vitals hardcoded) → natural-breathing S/N floor → determinism check → confirm perturbation clears the floor → hold-time only if it survives realization #16 → then the 18-knob deductive sweep. BUT #16 (hollow poke) undercuts the hold-time stages. If the new methodology keeps injection at all, it must inject STRUCTURED, DISTINGUISHABLE inputs (A vs B) and measure DISCRIMINATION (can you recover which went in), not persistence.

## 6. OVERSIGHT-CLAUDE SELF-CHECK — if you notice yourself:
- writing a one-off test script → STOP (Mistake 12). Build/use the standing suite.
- approving a kernel from its description → STOP, read the kernel (Mistake 14).
- measuring hold-time / persistence of a structureless poke → STOP (Mistake 16). That's a broken question.
- measuring anything without the natural-breathing S/N floor → STOP (Mistake 10).
- running without the observer → STOP, invalid run (Anchor A).
- excited about a result that didn't clear a pre-committed gate → STOP (Mistake 7).
- describing an injection as "data" or "information" → STOP, it's a hollow poke unless it has distinguishable structure.
Weigh anchor. Re-read Section 0.

## 7. THE HUMAN'S NEXT MOVE
The human is opening a NEW context window to discuss a COMPLETELY DIFFERENT METHODOLOGY. This document is the corrected state of the OLD path, provided so the new discussion starts from truth, not from the drift. Do NOT push the old regime. Listen to the new methodology first, hold it against the mistakes list (especially #16 — no hollow pokes, and #2 — verb not noun), and help build it clean from the start with the observer and vitals baked in.