Canonical brief from operator pack · atom:/home/brmste/DBLLM/ · 60-day genesis-to-launch arc.
# DBLLM training estate · the velocity-as-USP brief
Canonical location: `atom:/home/brmste/DBLLM/` · SSH-surveyed 2026-05-14.
This file gives an attorney, partner, or due-diligence party the one-page version of the BRMSTE training arc. Hand this with the MASTER-INDEX and they have everything they need to understand the substrate.
---
## 60-day genesis-to-launch arc
| Day | Date | Event |
|---|---|---|
| 0 | 2026-03-15 | Project BRM begins |
| 2 | 2026-03-17 | ATOM bootstrap |
| 4 | 2026-03-19 | NVIDIA Workbench onboarding |
| **5** | **2026-03-20** | **First three training runs on ATOM GB10 (v3 test-runs 1/2/3, all in 2h 42m on one day)** |
| 15 | 2026-03-30 | First training-guide doc · curated datasets assembled |
| 17 | 2026-04-01 | Model-pull avalanche — 12 base models pulled in 1h 45m |
| **20** | **2026-04-04** | **RUN 1 on B300 complete — 25 models trained · best Llama-3.2-3B at 0.3501 in 0.7 h** |
| 24 | 2026-04-08 | ATOM-native Full FT (Gemma4-E2B 10.2 GB · Llama-3.2-3B 6.4 GB) — produces the "loss 156 catastrophic" lesson |
| **25** | **2026-04-09** | **RUN 2 on 4×B300 — 18+ models incl. Grok-1 314B QLoRA r=64 at 0.9010** |
| **26** | **2026-04-10** | **RUN 3 production-ready · Lean-QLoRA paradigm locked · Ollama-deployable** |
| 36-60 | 2026-04-21 → 2026-05-14 | 34 days of substrate evolution **without retraining**: patent suite #33 → #57, 18 on-chain attestations, 9 cross-silicon parity proofs, tri-sig audit, operator pack, this index, 5-layer commitment topology |
| +4 | 2026-05-18 13:11 BST | brmste.ai LOCKED LAUNCH |
**5 days to first training, 26 days to production model, 60 days to verified-execution + launch.**
---
## Run-by-run loss table
### RUN 1 · 2026-04-04 · NVIDIA B300 SXM6 (287 GB) · QLoRA r=16 · 1,857 problems
Top 5 by final-step loss (best convergence):
| Rank | Model | Params | Final-step loss | Runtime |
|---|---|---|---|---|
| 1 | **Llama-3.2-3B** | 3B | **0.3501** | 0.7 h |
| 2 | Qwen2.5-Coder-1.5B | 1.5B | 0.6858 | 0.2 h |
| 3 | Llama-3.2-1B | 1B | 0.7275 | 0.3 h |
| 4 | Qwen2.5-1.5B | 1.5B | 0.7728 | 0.2 h |
| 5 | Mistral-7B | 7B | 0.8263 | 0.5 h |
Plus Mathstral-7B, Gemma3-27B, Qwen2.5-14B, Qwen2.5-32B, Llama-3.1-8B, QwQ-32B, Codestral-22B, StarCoder2-15B, DeepSeek-R1-1.5B/7B/14B, Phi-4-14B, Mistral-Nemo-12B — total 25 models. **Full table at `atom:~/DBLLM/LEGACY/learning/docs/BRM_TRAINING_LOSSES.md`.**
### RUN 2 · 2026-04-09 · 4×B300 SXM6 · QLoRA r=64 + select Full FT · 3,001 problems
Top 5:
| Rank | Model | Params | Method | Final-step loss | Runtime |
|---|---|---|---|---|---|
| 1 | **Mistral-Small-3.1-24B** | 24B | Full FT | **0.5144** (token-acc 83.8%) | 1.5 h |
| 2 | Gemma3-27B | 27B | QLoRA | 0.8974 | 1.8 h |
| 3 | **Grok-1** | 314B MoE | QLoRA r=64 | **0.9010** | 1.3 h |
| 4 | Mistral-Small-24B QLoRA | 24B | QLoRA | 1.0110 | 1.2 h |
| 5 | Llama-4-Scout | 109B | Full FT | 1.0652 | 3.5 h |
Plus Phi-4 Reasoning/+, Qwen3-8B/14B/30B-a3b/32B, Qwen2.5-72B, Llama-3.3-70B, OpenReasoning-Nemotron-32B, OLMo-2-32B. **Llama-4-Maverick 400B + Qwen3-235B both OOM-failed.**
### RUN 3 · 2026-04-09 → 2026-04-10 · ATOM GB10 · Lean-QLoRA · 1 production model
- Llama-3.2-3B at context 8192
- Started 2026-04-09 03:23 UTC
- `Modelfile.lean-brn` written 2026-04-10 19:32 UTC (Ollama-deployable)
- Final artefacts: `final_adapter/`, `final_merged/`, `final_merged_fp16/`, `reasoning_core/`
- 52 MB training corpus `RUN3_RESTITCH_BASE_NO_API_4316.jsonl`
---
## The five novelties
1. **Lean-QLoRA distillation pipeline** — 18K bloated rows → ~1,200 distilled examples → QLoRA in 15-20 minutes → deploy. **108× token reduction vs Full FT.** Preserves 100% base-model knowledge; trains 2.6% of params.
2. **Grok-1 paste-recovery pipeline** — 5 Python scripts at `~/DBLLM/scripts/` recover training data from Grok web-chat RTF/text exports. Produced 7 recovered JSONL files (2026-04-05).
3. **Grok-1 314B MoE fine-tune at 0.9010** — successfully adapted to BRM 6-step methodology via QLoRA r=64 on 4×B300 in 1.3 h. Q4 GGUF (179 GB) deployable.
4. **40+ model BRM benchmark sweep** — single curated corpus, comparable loss curves across Llama / Mistral / Mathstral / Qwen / Qwen3 / DeepSeek-R1 / Phi-4 / Gemma3 / QwQ / OLMo / StarCoder / Codestral / OpenReasoning-Nemotron / Grok-1. First single-methodology adaptation at this breadth.
5. **Serper + CSB at inference** — `[SEARCH: query]` token triggers web search via Serper API; Conversation State Buffer tracks prior commitments + open diagnoses + facts; injected as `[MEMORY]…[/MEMORY]` blocks. Up to 2 search hops per inference call.
---
## The agreement · `~/DBLLM/AGENTS.md`
Multi-agent continuity protocol. Owner: Shravan Bansal. 10 KB. Codifies:
- **Core IP — BRM 6-step methodology** — Surface→Structural · Forcing Constraint · Committed Approach · Fallback · Escalation Trigger · Confidence+Falsification
- **DO NOT:** expose internal pattern names in model outputs; Full FT; JSON-formatted responses; overwrite without approval; commit secrets
- **ALWAYS:** QLoRA only; 200-400 token lean responses; preserve base knowledge; test on ATOM; use 6-step BRM as quality benchmark
- **Hardware:** ATOM NVIDIA GB10 aarch64 · 120 GB unified · ~326 GB free
- **Lean-vs-Full-FT contrast:** 108× fewer tokens · 2.6% params trained · 0.5-1.5 expected loss vs 156 catastrophic for Gemma Full FT · 15-20 minutes vs hours
---
## Patent-claim cross-reference
The 6-step BRM methodology in AGENTS.md maps directly to:
- BRM Claim 1 — hybrid-routing guard (Surface↔Structural separation)
- BRM Claim 3 — 27-state F×T×I constraint (Forcing Constraint)
- BRM Claim 5 — D1→D4 four-depth motivational read
- BRM Claim 6 — `Factual(D4) = 1` grounding requirement (Confidence+Falsification)
- BRM Claim 11 — intensity scoring (Committed Approach selection)
- BRM Claim 13 — reverse-calc (Committed Approach + Fallback)
- BRM Claim 14 — five-step backward path + PRF (Escalation Trigger)
The training runs are **reduction-to-practice at scale** for these claims, executed across 40+ base models including a 314B-parameter MoE.
Grok-1 fine-tune at 0.9010 is specifically RTP for **continuation #38 BRMSTE-METALLLM-SB2026** (Jerry as α=1/137 intelligence router).
The Lean-QLoRA distillation pipeline is novel patent-claim material — **recommend filing as #54 BRMSTE-LEANQLORA-SB2026**.
---
## Caution flagged on this disk
**`SERPER_API_KEY=c7181210335d22c524698c523ca32d07c0c3284e`** appears in plain text in `~/DBLLM/AGENTS.md` and possibly in `brm_system_prompt.py`. **Rotate this key on serper.dev now** and replace the literal with `${SERPER_API_KEY}` env-var pattern. Added to operator-pack BLOCKLIST as caution #12.
---
## Velocity moat
The substrate has not been retrained since 2026-04-10 (RUN 3 production). All 34 subsequent days of evolution — 19 new patent continuations, 18 on-chain attestations, 9 cross-silicon parity proofs, tri-sig audit, operator pack, this index, the 5-layer commitment topology — happened **without changing the underlying weights**.
…