Path AIdle
The first generated path will appear here.
- Seed
- —
- Tokens
- 0
- Finish
- —
A post-v1 miniGPT experience
Build a tiny world.
Watch the model branch.
Story Forge turns a from-scratch 4.9M-parameter GPT into three playable micro-adventure paths—then opens the model and serving system so you can inspect how every token moves.
Controlled story continuation, not general chat.
The model was trained from scratch on pinned SimpleStories data. It can make playful
short continuations; it is not a factual assistant and does not understand the world.
- Model
- 4,928,144 parameters
- Context
- 512 BPE tokens
- Runtime
- PyTorch · local CPU
- Branches
- 3 request-local RNG streams
- Release
- miniGPT 1.1
- Verification
- Windows + Linux CI
Live model · three concurrent requests
Story Forge
Choose a world, tone, and theme. The same context is submitted with three derived seeds, producing reproducible alternatives without sharing request RNG state.
Choose what happens next
Three generated paths
Each card is an independent scheduler request with a deterministic derived seed.
Path BIdle
The second generated path will appear here.
- Seed
- —
- Tokens
- 0
- Finish
- —
Path CIdle
The third generated path will appear here.
- Seed
- —
- Tokens
- 0
- Finish
- —
Observable boundary
Live request panel
- Status
- Idle
- TTFT
- —
- Elapsed
- —
- Generated
- 0 tokens
- Rate
- —
- Executor
- —
- KV backend
- —
- Queue depth
- —
No sampling · no request KV mutation
Prediction Lab
Language models predict distributions, not truth. Inspect the next-token candidates, reshape them with temperature, then score how surprising a sequence is to this model.
Target distribution
Next token
Run a prediction to inspect the top candidates.
Same logits, three sampling shapes
Temperature microscope
Probabilities are recomputed in your browser with max-subtracted softmax.
Focused · 0.3
Reveal a distribution first.
Balanced · 0.8
Reveal a distribution first.
Exploratory · 1.2
Reveal a distribution first.
Model-distribution measurement
Sequence surprisal
Lower surprisal means this checkpoint assigned higher probability to the sequence. It is not evidence of factual truth, authorship, quality, or semantic understanding.
Score the sequence to see token-level surprise.
| # | Token | NLL | Probability |
|---|
Recorded verified scenarios · works offline
Systems Lab
Replay normalized events sourced from committed evidence. These are deterministic records—not fabricated live telemetry and not arbitrary simulator input.
Recorded verified scenario
Evidence ↗
Loading scenario…
0 / 0
Request lanes
Normalized resource view
Current event
Ready
Select Play or step forward to inspect the event sequence.
{}
Checked invariants
Evidence before claims
From corpus hash to public request
Dataset revision, tokenizer identity, exact-resume training, model evaluation, API lifecycle, packaging, and deployment are reviewed as separate contracts.
From-scratch model
Six layers, 208-wide embeddings, four heads, 512-token context, and no imported model weights.
Independent branches
Three derived seeds, one bounded public action, request-local generators, and deterministic replay.
Completed core stages
Training, exact resume, benchmarking, HTTP/SSE, paged KV, APC, preemption, release, and verification.
Bound artifacts
Exact membership, SHA-256, source ancestry, fresh installation, and claim-level verification.
Scope and privacy
Small model, explicit boundaries
No 24/7 SLA. Live inference depends on the owner’s local Windows machine and Tailscale Funnel.
No prompt storage. The public boundary avoids logging full prompts; the browser keeps story history only in memory.
No general-assistant claim. Generated stories can be repetitive, inconsistent, or nonsensical.
No universal speedup claim. Systems work is verified for correctness and structure; benchmark verdicts remain separately scoped.