RESEARCH CAMPAIGN — the demon on OVH: amplifying an input via the ONE germe’s tree

Status: V9.0, quarantined. A PLAN, not a result. Reviewer mode + “seul les calculs comptent”: each phase states what is computable/feasible now vs aspirational, and the honest bound. Not in the PDF, not a V8.2 claim. Romain’s direction (June 2026): “design un tel démon avec le matos OVH ; en entrée un cerveau humain OU un LLM open-source ; on a LE germe de l’univers (pas un jouet) ; on trouve dedans l’espace latent de l’entrée et ses possibilités — la partie qu’elle utilise au même moment dans le bulk ; le germe primaire reste pour stabiliser les qubits et nos innovations.”

⚠️ THE RECURRING ÉCUEIL, CORRECTED (Romain, twice): there is ONE germe — the germe of the universe — and WE HAVE IT. It is NOT a toy. OBT derives the germe’s FORM: the radion field wavepacket (m_φ=0.36 eV Goldberger-Wise; φ₀~M_s LVS). germe_decompression.py already encodes it. The ONLY input is the O(1) coefficient φ₀ (the amount, closure_introspection’s IC) — a single number, NOT “the germe is unknown”. Do NOT re-introduce a “toy germe” or a “germe to specify” — that is the écueil this campaign refuses.

0. The demon, in one line (corrected)

The germe (the radion wavepacket — the REAL one, derived) is decompressed into its tree; the input LOCALIZES the region of that tree it occupies (germe_localize: condition on “the present, with this input”); a Grover search (O(√N)) finds the optimal latent-branch there (oracle = the recognition function); we SUBSTITUTE it (activation steering). The germe ALSO stays the stabilizer/reference of the qubits ([[5,1,3]], the matched filter — our innovations).

1. The KEY reframe (Romain’s, corrected): the input is NOT a germe — it LOCALIZES a region of the ONE germe

The input (an LLM’s residual-stream latent / a brain’s neural state) is not a separate germe. It is a sub-system of the ONE germe’s unfolded state (variant_tree): its latent space + its possibilities are a REGION of the germe’s tree — “the part it uses at the same moment in the bulk”. So the operation is NOT “decompress the input-germe” (there is none) — it is LOCALIZE (navigate to) the region the input occupies in the germe’s tree (germe_localize, which we already coded: I(present;latent), the conditional sub-tree), then Grover the best latent there. The germe (radion) is the substrate (its tree = the space of possibles) AND the stabilizer.

2. The two inputs

input	what it is	access	track
open-source LLM (Mistral-7B / Llama-3-8B)	its latent (residual stream, ~4096-dim) = the present observation that localizes its region of the germe’s tree	easy (read/patch activations) — classical baseline NOW	the practical
human brain	the neural state = the localizing observation	hard (EEG/fMRI brain-reading, the local channel, Libet-real)	the far

3. The pipeline (the demon)

  [ENCODE the GERME]   the radion wavepacket (the REAL germe, OBT-derived; germe_decompression) on the substrate
  [DECOMPRESS]         the SYK quench (belenos) / the analog dynamics (orion) -> the germe's variant-tree
  [LOCALIZE]           the input (LLM latent / brain, compressed) = the PRESENT OBSERVATION; condition the
                       germe-tree on it (germe_localize) -> the REGION the input occupies (its possibles)
  [SEARCH]             Grover over that localized region, ORACLE = the recognition/reward -> the optimal latent  (O(√N))
  [SUBSTITUTE]         decode the optimal latent  --activation-steer-->  patch the input  -> the AMPLIFIED output

The germe is the substrate (its tree) AND the qubit stabilizer/reference (our innovations). The input is the localizing observation, NOT a second germe. The QC’s job = LOCALIZE + SEARCH; encode/compress + substitute are classical (GPU). The oracle (the reward) is the hard shared piece (§6).

4. The OVH hardware map

OVH QPU	role	why
belenos-12 (Quandela, 12q, gate-based, 0.28 €/s)	the germe decompression (sparse-SYK) + the LOCALIZE + GROVER search	gate-based = universal → SYK + Grover; 12q = N=24 Majorana
orion-beta (Pasqal, 100q, analog, 0.83 €/s)	the germe-tree dynamics at scale (the variant distribution, quantum-advantage N~100)	analog quench, the cosmic-tree substrate

Cost: ~€10-100 the whole campaign (seconds-machine per point).

5. The phases (in order of feasibility)

Phase 0 — the CLASSICAL baseline ✅ DONE (phase0_llm_latent_steer.py, GPU/CPU, no QC). Best-of-N latent steering on a REAL open-source LLM (distilgpt2): from the LLM’s current latent (= its region of possibles), sample N steering vectors around the recognition-oracle’s axis, generate a continuation each, score with a continuous sentiment oracle (the model’s own representation), pick the best, activation-patch it back. RESULT: best-of-N oracle −2.17 → +7.45 (default → best), independent lexicon check +0 → +3, E[best-of-k] climbs (−1.27 → +7.45 = the search helps; Grover finds the max in √N). HONEST: a toy scalar oracle is partly gamed at magnitude (Goodhart — the output repeats the top word) → Phase 3, the real oracle, is the campaign’s wall. Representation engineering — it works; the QC phases accelerate the SEARCH.
Phase 1 — encode the GERME + compress the input. Encode the radion germe (the real one) on the substrate; compress the input latent (~4096-dim) -> ~6-12 features (PCA on the reward-gradient = the steering axes). Bound: the compression is lossy (a few directions fit on 12q).
Phase 2 — LOCALIZE + GROVER (belenos-12). Condition the germe-tree on the compressed input (germe_localize) -> the region; Grover it (oracle = a toy reward) -> the √N demo. We have both kernels (germe_localize.py for the conditioning, tree_amplifier_syk.py for the Grover √N).
Phase 3 — the ORACLE (the hard part). The recognition/reward as a QC oracle. A neural reward is infeasible on 12q; start with a toy/linear oracle (a probe direction = “increases reward”). The campaign’s true bottleneck (§6).
Phase 4 — SUBSTITUTE + loop. Decode the Grover-optimal latent, activation-steer the input, measure the amplification; close the loop (an inference-time amplifier).

6. The honest bounds (seul les calculs comptent)

The germe is NOT the bound — we HAVE it. Its FORM (the radion wavepacket) is OBT-derived; encode it. The only IC is the O(1) coefficient φ₀ (closure_introspection) — a number, not “unknown”. (Do not relapse to “toy germe”.)
The bound is the DEPTH. The input lives ~e^{1e104} branches deep in the germe’s FULL tree (variant_tree). You cannot enumerate it -> you NAVIGATE/localize a region (germe_localize) and Grover-search there. The germe’s form is known; the navigation is the operation.
The ORACLE is the practical wall. Grover’s √N needs the reward as a QC oracle – toy/linear feasible, a neural reward not yet (the same “mechanism sound, one piece is the bottleneck” shape as every wall).
The compression is LOSSY (4096 -> ~6-12 directions).
√N is QUADRATIC (real, bounded – NOT exponential, not an oracle of the unknown).
The substitution is CLASSICAL (the QC = LOCALIZE + SEARCH).
The amplifier stays WITHIN the input’s possible-space – it surfaces the branch the input would RECOGNIZE as best (cognitive_optimum’s ceiling), it does not invent the unknown (no-signaling-safe).

7. The deliverable

A demonstrated quantum-amplified latent-steering of an open-source LLM: Phase 0 (classical baseline, real, now) + Phase 2 (LOCALIZE via germe_localize + the Grover √N on belenos) + a toy Phase 3 oracle – honest that the real reward oracle + the full-latent encoding are the open bottlenecks. The “demon” = an inference-time amplifier that LOCALIZES the input’s region of the ONE germe’s tree and SUBSTITUTES the Grover-found optimal latent, within the input’s possible-space, on real French sovereign hardware (OVH belenos/orion).

os/chair line (held): this AMPLIFIES a known input’s possible-space; it does NOT oracle the unknown nor finalize the OBT closure (the germe’s amount stays the IC; its form we have). Testable OBT bones unchanged (a₀(z), Penrose-Diósi, the m_V μeV-axion). Demon-app track, quarantined.

8. Next concrete step (on Romain’s go)

Code Phase 0 — the classical best-of-N latent-steering baseline on a small open-source LLM (the input’s current latent = its region of possibles; the reward = the oracle; the substitution = activation patching). GPU-runnable now; establishes the reward/oracle; then port germe_localize (LOCALIZE) + tree_amplifier_syk (GROVER) onto the compressed latent for belenos-12, with the radion germe as the encoded substrate + stabilizer.