Random became the first environment the model learned to inhabit.

Controller bench

Controlled model boundaries, not a single opaque agent

Real David prompt
       |
Gemma native-thinking checkpoint
       |
Gemma continuation seed
       |
David-GPT-2 prose continuation
       |
optional bounded checkpoint rounds
       |
Gemma / Sol final streamed answer

Architecture milestone / August 15, 2026

The public continuation API can sample eight explicitly named lineage points from raw GPT-2 through the exhausted Generation 001 run. Historical checkpoints load one at a time, emit separately attributed predictions, and leave the validation-selected G1 model resident. Gemma/Sol owns synthesis. Linux and direct API observation retain factual authority.

The lower loop is fine-tuning foresight built into the architecture: the model can highlight a completed exchange as educational, but exact transcript bytes enter an unapproved ledger and only operator authorization can launch training.

Dedicated model API

Continue text with the local checkpoint

Live dedicated continuation API
David model=/home/david/projects/david-gpt2/outputs/david-gpt2-random-notes/best
selected validation step=2000
knowledge=http://127.0.0.1:8892/api/knowledge/query
sitewide Sol / Gemma remains a separate service boundary

curl 'https://sol.system42.one/api/david-continuation/v1/continuations' \
  -H 'Content-Type: application/json' \
  -d '{
    "prompt": "An engineer from Orbital Control enters the room.",
    "prefix": "OLD SOL (V.O.)",
    "mode": "screenplay",
    "knowledge": {"enabled": true, "top_k": 3},
    "max_tokens": 200
  }'

Output will appear here.

David-GPT-2 + Gemma native-thinking test bench
David model=/home/david/projects/david-gpt2/outputs/david-gpt2-combined/best
Gemma model=gemma4-12b-q5-chat
Gemma dialogue mode checkpoints Ollama's message.thinking stream between calls.
Type /help for controls or /exit to quit.
[status] mode=david knowledge=off extended=False max_rounds=4 show_internal=True david_device=auto david_max_tokens=200 temperature=0.8 seed=37 gemma_checkpoint_tokens=192 gemma_primer_tokens=128 gemma_final_tokens=768

Sol prompt> [screenplay log] EXT: 

[David-GPT-2] loading checkpoint: /home/david/projects/david-gpt2/outputs/david-gpt2-combined/best
Loading weights: 100%|█████████████████████| 148/148 [00:00<00:00, 7292.98it/s]
[David-GPT-2] ready on cuda after 3.0s          
[David-GPT-2] CUDA memory allocated=475 MiB reserved=530 MiB
[David-GPT-2] generating David-role continuation
[David-GPT-2] stop=length_limit token=- generated_tokens=200 elapsed=2.1s
David-GPT-2> An engineer from Orbital Control enters the room.

OLD SOL (V.O.)

This is Clint. He's here to measure the integrity of our systems.

CARTER leans toward him.

OLD SOL (V.O.)

So if someone can hack into our systems? That might not be an unusual request. But it's a reminder that power supplies are delicate places. We're building something big when we're trying to keep them under control.

CUT TO BLACK.

INT. LUNAR HAVEN - ARCHIVE STRUCTURE

The Atlantis team finds themselves inside the restraint seat, staring at the Earth through a plasma shielded glass dome above them.

SG-1 personnel stand in the doorway, their hazmat suits covering their faces.

ANCIENT EVANS

Welcome. I'm David.

DAVID nods.

ANCIENT CLIANT

Welcome. David and I have
[turn complete: 5.3s]

Sol prompt> hello friend

[David-GPT-2] generating David-role continuation
[David-GPT-2] stop=length_limit token=- generated_tokens=200 elapsed=4.5s
David-GPT-2> .

A beat.

ANCIENT VOICE

Hello world.

They turn to one another.

ANCIENT TECH (V.O.)

Welcome.

DANIEL pauses.

ANCIENT VOICE

David.

Not with joy in mind.

He stops.

ANCIENT VOICE

Yeah.

Silence.

Then:

OLD SOL (V.O.)

The condition wasn't satisfactory.

SOL (V.O.)

It was just that they couldn’t see each other.

### INT. DERELICT VESSEL - CORE COMPARTMENT

Athena displays a diagnostic window.

ORBITAL CONTROLS

Quantum channels overlap. Interactions cascade outward. Partial synchronization fails. Metallic surfaces appear imperceptibly black.

No atmosphere contains more than 1% normal flow.
[turn complete: 4.5s]

Sol prompt>   

Boundary

Completed continuation experiment

Learning the persistent random and notes filesystem

/home/david/random began as a scratch directory, then accumulated projects, services, experiments, media, source history, abandoned prototypes, and the working structure of Sol-37. It evolved from junk drawer to project nursery, archive, Git repository, and persistent personal systems namespace. Generation 001 asks whether that heterogeneous history can become a coherent parametric environment rather than merely a pile of memorized files.

A full-model continuation fine-tune learned from content-addressed snapshots of the actual random and notes trees. The source checkpoint and training corpus remain frozen. Ordinary David prompts were deliberately reduced, while screenplay logs, audio-drama output, narration, and other scene/dialogue records received the largest training exposure.

weighted supervised-token mix
15% ordinary David continuation
 5% preserved special baseline
30% random + notes filesystem continuation
50% screenplay and log continuation

Every accepted source was frozen by SHA-256 before tokenization. The dedicated CUDA service restored the local inference daemons it had quiesced for VRAM. The dashboard distinguishes training, validation, serialization, recoverable checkpoints, final publication, and service restoration. Generation 001 is deployed to the dedicated continuation endpoint as an experimental model; measured limitations remain visible rather than being hidden behind a generic chat interface.

BAAI/bge-small-en-v1.5                 local FastEmbed
snowflake/snowflake-arctic-embed-s     local FastEmbed
text-embedding-3-small                  OpenAI coordinate
text-embedding-3-large                  OpenAI coordinate

per-model chunk indexes
        -> normalized per-file centroids
        -> SQLite vector + provenance ledger
        -> reciprocal-rank fused retrieval

The post-training work now crosses two machines

UPDATED 2026-08-19 · 00:37 CDT

Generation 001 training is complete; this activity belongs to the separate retrieval build. Two remote OpenAI coordinates are already committed, Linux is rebuilding the local BGE coordinate from a compact float32 cache, and the 2014 Mac mini now computes a deterministic quarter-shard of the second local Arctic coordinate. The dashboard keeps that distributed work separate from the committed ensemble percentage until Linux validates and consumes it.

Linux coordinateBGE · 72.31%

Whole ensemble68.08%

Mac Arctic shard392 / 9,958

Mac throughput~566 token/s

Why is the Mac not added to 68.08%? The four-model percentage measures complete coordinate state, not aggregate CPU effort. Mac vectors remain a resumable Arctic cache seed until an online SQLite backup reaches Linux, passes database integrity, exact model/chunk/dtype checks, 384-dimension checks, and a current SHA-256 comparison for every projection. Linux then rewrites path identity and document IDs into its own cache before the ordinary Arctic builder publishes the coordinate.

Process	Class	What it is doing	How to read its state
david-gpt2-corpus-embeddings	Finite builder	Reads the frozen and current random/notes inventory, chunks eligible text, computes the active model's vectors, reduces them to per-file centroids, and records provenance.	active/running is the six-CPU embedding job. It is intentionally CPU-only so the restored inference stack keeps the GPU.
Mac mini Arctic shard 0/4	Finite distributed seed	Uses two background CPU threads to embed one deterministic SHA-256 partition with snowflake-arctic-embed-s. It commits each file to SQLite and can stop without discarding prior rows.	Forced SSH commands, a deny-network sandbox, one-job limits, idle/load admission, and BlueBubbles/Messages health gates protect the interactive Mac. Its progress is visible but remains outside the committed ensemble until Linux validation.
david-gpt2-ai-watchdog	Persistent safety control	Polls both hosts, preempts the Mac on protected-service, memory, disk, or interactive-user pressure, takes read-consistent snapshots, and sends deduplicated private Unified Clock notices.	A snapshot is transport, not authority. The Linux merge still rejects changed or mismatched projection bytes.
embedding-post-initial-catchup	Finite reconciliation	Waits for the initial four-model publication, then revisits files added or changed while the long build was in flight.	activating/start with almost no memory is a deliberate wait, not a hung worker. It closes the gap between the build's opening tree signature and current reality.
sol-video-semantic-index	Separate one-shot	Maintains the video-semantic index when explicitly invoked; it is displayed so another finite semantic job cannot be mistaken for the corpus ensemble.	inactive/dead is normal when no video indexing pass is scheduled.
sol-chat-model	Persistent inference	Keeps the Gemma controller resident for the Orb, SOL Chat, policy routing, and synthesis.	Its GPU use is bursty. GPU panels sampled a few seconds apart can disagree without implying that Generation 001 resumed training.
david-gpt2-continuation-api	Persistent inference	Serves the validation-selected Generation 001 checkpoint and the attributed historical lineage endpoints.	It remains separate from Gemma so continuation predictions and controller synthesis retain distinct provenance.
ollama / sol-chat-api	Persistent routing	Preserve local model compatibility and the site's normal chat contract while background retrieval work continues.	These are infrastructure, not unfinished build steps; the dashboard reports them but does not stop them when managing the embedding PID.
knowledge-query-api	Persistent retrieval	Continues serving the previously published hybrid knowledge index until the new four-coordinate ledger is complete and approved for use.	Availability of the old index and incompleteness of the new index are compatible states.

finite path
Linux BGE -----------------------> complete BGE coordinate
Mac Arctic shard 0/4 -> snapshot -> hash-gated Linux Arctic cache
Linux Arctic remaining 3/4 ------> complete Arctic coordinate
OpenAI small + OpenAI large -----> already committed
        -> atomic ensemble publication
        -> post-build live-tree catch-up
        -> watcher becomes idle

persistent path
Gemma + G1 continuation + Ollama + SOL Chat + Knowledge API
        -> remain available throughout the finite build

The Prioritize Build control raises only the Linux builder's systemd CPU and I/O weights without restarting it. The Mac remains independently bounded to two threads, 2 GiB of schedulable memory, 32 GiB of scratch, and one background job. Pause, resume, stop, snapshot, and process controls verify expected identities before acting; neither host signals an unrelated high-CPU process merely because it appears in telemetry.

frozen G1 source snapshot
        < retrospective screenplay log

current retrieval may return the log
active G1 weights cannot claim it

Generation 001 field record

The desktop sequence records training, checkpoint, validation, hardware, and GenOS states as they actually appeared. Select any image for the larger web copy.

LLM-first operating environment

Learn the world, test its physics, curate the disagreement

The long-term project is not a model pretending that Linux commands worked. David-GPT-2 supplies a learned prior over this particular filesystem; a controller proposes observable behavior; a policy decides when belief is insufficient; and a sandboxed Linux oracle supplies ground truth. Every comparison can become a structured operational trace for a later generation.

authority order
current user write or direct observation
  > verified materialized state
  > retrieved historical state
  > parametric memory
  > unconstrained generation

learning loop
predict → execute or verify → compare → preserve trace
        → train on informative disagreement → predict again

Controlled replay regimes

Each regime receives the same deterministically reset sandbox scenarios. Backend outage is kept separate from a deliberate Linux-only control so retry and failure behavior cannot contaminate the oracle baseline.

Replay v1 contains 13 reset scenarios. A0 and A1 currently agree on all Linux outputs and filesystem transitions. Future model regimes will be compared on exact output, exit status, mutation correctness, prediction coverage, false confidence, calibration, Linux escalation, and consequence-weighted regret. The decisive mutation test presents observed state B after training on state A: current observation must override parametric memory.

Controlled learning loop

The final ensemble can nominate its own next training examples

With training functions explicitly enabled, the Gemma-owned final ensemble may tag an earlier completed-history window as a candidate or stage an on-demand fine-tuning request. The host, not the model, supplies exactly the transcript window the model reviewed. Candidates are append-only, content-hashed, and unapproved by default; routine conversation is deliberately excluded unless the model identifies durable training value.

completed host-owned session history
        ↓ model nominates rationale + tags
append-only candidate ledger
        ↓ model requests a fine-tune
immutable session-separated dataset
        ↓ explicit operator authorization
configured trainer or no launch

A function call cannot approve its own data or launch training. Launch requires a separate operator flag and argv-only trainer configuration, and remains blocked while another David-GPT-2 training process is active. The current in-progress answer is never captured mid-generation; it becomes eligible on a later turn.

Custom dataset guide

Recreate the architecture with your own target voice

The reusable idea is role-reversed causal-language-model training: context belongs to the prompting side, while loss is applied only to the target person's continuation. Replace David's private corpus and identity labels with your own reviewed data; do not redistribute either model's weights or source conversations.

1. Choose one target role

   Define whose continuation the small model predicts and which role supplies context. In this project, assistant/Sol text is context and David/user text is the supervised target. Keep that mapping explicit; do not mix both sides into one assistant target.

2. Normalize conversations

   Export ordered conversations with stable conversation and message IDs. Adapt `src/david_gpt2/archive.py` if your source is not the expected transcript layout. Preserve turn order, exclude system/tool metadata from visible training text, and assign train/validation/test splits by conversation rather than by isolated message.

3. Define task and role markers

   Replace or extend the structural tokens in the dataset builder for your target role. Mask task markers, context, padding, and role delimiters with label `-100`; supervise only the target body and its real end-turn token. Keep literal marker text escaped so authored strings cannot become control tokens.

4. Build and audit the private dataset

   python scripts/build_dataset.py \
     --archive /path/to/your/private-archive \
     --output data/processed \
     --context-length 512 --seed 37
   python scripts/audit_literal_markers.py --dataset data/processed
   python scripts/audit_supervision.py --dataset data/processed
   pytest -q

   Review decoded labels locally. Every intended target occurrence should have complete token coverage, with no cross-split conversation leakage and no private target text in reports.

5. Curate without rewriting the frozen source

   python scripts/dataset_workbench.py

   The local workbench presents searchable human-readable context and supervised targets. Add, rewrite, remove, or restore individual entries through a reversible overlay, then materialize and validate a versioned candidate. Opening the editor or building a candidate never starts training.

6. Run a real smoke fine-tune

   python scripts/smoke_test.py --config configs/smoke-test.yaml

   A smoke checkpoint validates mechanics, stop behavior, save/reload, and CUDA memory. It does not establish voice quality. Compare base and fine-tuned output before authorizing a full run.

7. Train and select the checkpoint

   python -m david_gpt2.train --config configs/combined.yaml
   python -m david_gpt2.evaluate \
     --model outputs/david-gpt2-combined/best \
     --dataset data/processed/combined/test

   Copy `configs/combined.yaml`, then adjust paths and mixing ratios deliberately. Select `best` by supervised validation loss and inspect unrelated negative prompts for memorization or mode collapse.

8. Prepare Gemma and your identity prompt

   ollama pull hf.co/zaakirio/gemma-4-12b-it-uncensored-GGUF:Q5_K_M
   ollama create gemma4-12b-q5-chat -f /path/to/Modelfile

   You may substitute another Ollama model that exposes streamed `message.thinking`, but verify its API fields and template. Create a separate system prompt for the final model's identity; do not train that identity implicitly into the target-role GPT-2 unless that is a deliberate dataset goal.

9. Point the bench at your checkpoint

   python david_gpt2_gemma_dialogue.py \
     --david-model /path/to/your/checkpoint \
     --mode dialogue --show-internal \
     --gemma-checkpoint-tokens 192 \
     --gemma-primer-tokens 128

   Rename UI labels and role markers in a fork, but preserve the boundary: Gemma creates a prose seed, the small model continues that seed as its target role, and a later Gemma request owns the final visible answer.

10. Add retrieval only after the baseline works

    Run first with knowledge off. If you add retrieval, point `DAVID_CONTINUATION_KNOWLEDGE_URL` at your own read-only query endpoint. The endpoint must return ranked `results` containing `text`, `path`, and `chunk`. Search may cover your complete index, but cap `top_k` and per-hit characters so evidence stays inside the model's finite context. A generative backend should not silently rewrite evidence.

11. Run a separate continuation service

    python -m david_gpt2.continuation_api \
      --model /path/to/your/checkpoint \
      --host 127.0.0.1 --port 8793 \
      --knowledge-url http://127.0.0.1:8892/api/knowledge/query

    Publish it beneath a dedicated route such as `/api/your-model/`. Do not alias it to an existing assistant-chat endpoint. Keep the request centered on prompt, optional prefix, target mode, and continuation controls.

12. Package source, never private runtime assets

    Publish source, configs, tests, launch definitions, and sanitized documentation. Exclude checkpoints, raw/processed corpora, embedding caches, manifests containing private IDs, logs, exports, API keys, and virtual environments.

Complete publication

Reproduce the experiment without publishing the private world

SHA-256 checksums