prebuilt llama-cpp wheel index + field notes

FIELD_NOTES.md

@@ -0,0 +1,83 @@
+# Field Notes: GODSEED — the god grants one wish at a time
+
+*Build log for [GODSEED](https://huggingface.co/spaces/build-small-hackathon/godseed) — Build Small Hackathon, Thousand Token Wood track.*
+
+## The idea
+
+Type a wish. A tiny Nemotron god reads it aloud — you watch its thoughts stream — then
+terraforms **one persistent WebGL planet shared by every visitor**. No level select, no
+reset button. Your wish lands on the same small world as everyone else's, forever.
+
+The pitch fits in one line: **3.2B active parameters. One shared world.**
+
+## Architecture: the model owns nothing
+
+The design rule I refused to break: *the deterministic engine owns every fact; the model
+only interprets, narrates, and chooses.* The god picks from a DSL of exactly 8 tools
+(`raise_terrain`, `spawn_flora`, `place_water`, `set_sky`…). The engine validates every
+call, clamps every number, seeds every feature, and feeds terse observations back:
+
+```
+{"thought":"The tower. The lamp. The long patience.",
+ "call":{"tool":"place_structure","args":{"kind":"lighthouse","lat":26,"lon":-64,...}}}
+→ ok: lighthouse placed at (26,-64)
+```
+
+Because every feature carries its seed, the whole planet is a pure function of an
+append-only feature list. That one decision bought three things for free: the landing
+page renders the entire world client-side with **zero GPU**, any wish in history can be
+**replayed deterministically** in the Genesis Log, and the agent traces published to the
+[dataset](https://huggingface.co/datasets/AndresCarreon/godseed-world) are complete
+evidence of every plan-act-observe loop.
+
+## Things that went wrong (the good part)
+
+**The planet was a black void.** My first renderer pass had gorgeous typography, a
+glowing atmosphere rim, golden inscription rings — wrapped around a planet rendered at
+5% luminance with no light rig. Obsidian-dark albedo × almost-no-light = a hole in
+space. The fix was cinematography, not brightness: a palette-tinted key sun, an
+atmosphere-colored rim light behind the dark limb, and a low bounce from below-front so
+the shadow side reads as deep violet instead of dead pixels. Dark worlds still need
+three-point lighting.
+
+**The god got stuck raising the same mountain six times.** Wish: *"raise a mountain
+crowned with a shrine, where embers drift like prayers."* The 4B model raised terrain at
+(38,-102)… then again… six times, before finally placing the shrine. The trace showed
+why: every turn got back the **identical observation** — `ok: mountains risen at
+(38,-102)` — so the context never changed, and a small model in an unchanging context
+loops. I didn't fix it with prompt engineering. The engine now refuses an identical
+consecutive call with a new observation: `rejected: already done; change something, or
+say done`. The context changes, the loop breaks, deterministically. Small models don't
+need bigger prompts; they need environments that push back.
+
+**The free CPU tier dream died with a stopwatch.** Plan A was llama.cpp + the 4B GGUF on
+the free CPU tier — a god that runs on two vCPUs, fully off-grid. Then I timed it: 183
+seconds per wish *on an M-series Mac*. The Space's CPU would be several times slower.
+You can call queue latency "liturgical pacing" — and I do, the god grants one wish at a
+time — but ten minutes is not liturgy, it's an outage. The live Space runs the same
+grammar-constrained loop on ZeroGPU; the llama.cpp mode survives as the documented local
+config (`GODSEED_BACKEND=llamacpp`), where the whole god is a single ~2.5GB GGUF.
+
+**The deploy stack fought back.** Three build failures in an hour: the Spaces image
+force-installs its own gradio and rejects your pin (`sdk_version` in the README is the
+real control); `mamba-ssm` won't source-build under pip's build isolation (its setup.py
+imports torch — use the prebuilt cu12 wheels from GitHub releases); and the
+transformers-v4 line caps `huggingface-hub<1.0` while new gradio wants ≥1.0, so the two
+must be version-matched deliberately. If you're shipping a NemotronH model to ZeroGPU
+this week: `sdk_version: 6.16.0`, `transformers>=4.56,<5`, wheel-pinned mamba kernels.
+
+## Moderation, because one bad forest ends everything
+
+A shared persistent world during judging week is an invitation. Wishes pass three layers
+before a single tool runs: charset/length limits, a normalized wordlist (NFKC,
+confusables, leetspeak folding), and an LLM judgment with **default-deny on any parse
+failure**. Rejections are final and polite: *the god declines this wish.*
+
+## What I'd tell you to try
+
+Open the Space, read the planet's history in the Genesis Log, then wish for something
+the world doesn't have yet. Watch the thoughts stream. Watch the camera glide to the
+exact latitude the god chose. That moment — the plan becoming terrain in front of you —
+is the whole project.
+
+*Speak, and it will be made.*

README.md

@@ -122,7 +122,7 @@ guardrails, geometry, and archive.
 - **Llama Champion** — the shipped backend is llama.cpp (`llama-cpp-python`).
 - **Off-Brand** — fully custom WebGL frontend; Gradio serves the altar console.
 - **Tiny Titan** — the 4B GGUF configuration runs the whole experience.
-- **Field Notes** — build log linked below with the submission links.
+- **Field Notes** — build log: [FIELD_NOTES.md](./FIELD_NOTES.md) (org blog post coming).
 - **Sharing is Caring** — live agent traces published as a public dataset.
 
 ## 📦 Submission links

requirements.txt

@@ -6,8 +6,9 @@ gradio>=5
 pydantic>=2.7
 huggingface_hub>=0.30
 # llama.cpp runtime for the Nemotron-3-Nano-4B GGUF backend (GODSEED_BACKEND=llamacpp).
-# Builds from source on the Space (cmake ships in the image); first build takes a few
-# minutes, then it's cached.
+# Source build KILLS the Spaces builder (June 12: died mid-compile) — use the
+# project's prebuilt CPU wheel index instead.
+--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
 llama-cpp-python>=0.3.8
 
 # ---------------------------------------------------------------------------