Submission README (tracks, sponsors, receipts) + Space hygiene: Gradio 6 sdk pin, drop stray cert, trim requirements

.gitignore

@@ -1 +1,2 @@
 __pycache__/
+.gradio/

.gradio/certificate.pem

@@ -1,31 +0,0 @@
------BEGIN CERTIFICATE-----
-MIIFazCCA1OgAwIBAgIRAIIQz7DSQONZRGPgu2OCiwAwDQYJKoZIhvcNAQELBQAw
-TzELMAkGA1UEBhMCVVMxKTAnBgNVBAoTIEludGVybmV0IFNlY3VyaXR5IFJlc2Vh
-cmNoIEdyb3VwMRUwEwYDVQQDEwxJU1JHIFJvb3QgWDEwHhcNMTUwNjA0MTEwNDM4
-WhcNMzUwNjA0MTEwNDM4WjBPMQswCQYDVQQGEwJVUzEpMCcGA1UEChMgSW50ZXJu
-ZXQgU2VjdXJpdHkgUmVzZWFyY2ggR3JvdXAxFTATBgNVBAMTDElTUkcgUm9vdCBY
-MTCCAiIwDQYJKoZIhvcNAQEBBQADggIPADCCAgoCggIBAK3oJHP0FDfzm54rVygc
-h77ct984kIxuPOZXoHj3dcKi/vVqbvYATyjb3miGbESTtrFj/RQSa78f0uoxmyF+
-0TM8ukj13Xnfs7j/EvEhmkvBioZxaUpmZmyPfjxwv60pIgbz5MDmgK7iS4+3mX6U
-A5/TR5d8mUgjU+g4rk8Kb4Mu0UlXjIB0ttov0DiNewNwIRt18jA8+o+u3dpjq+sW
-T8KOEUt+zwvo/7V3LvSye0rgTBIlDHCNAymg4VMk7BPZ7hm/ELNKjD+Jo2FR3qyH
-B5T0Y3HsLuJvW5iB4YlcNHlsdu87kGJ55tukmi8mxdAQ4Q7e2RCOFvu396j3x+UC
-B5iPNgiV5+I3lg02dZ77DnKxHZu8A/lJBdiB3QW0KtZB6awBdpUKD9jf1b0SHzUv
-KBds0pjBqAlkd25HN7rOrFleaJ1/ctaJxQZBKT5ZPt0m9STJEadao0xAH0ahmbWn
-OlFuhjuefXKnEgV4We0+UXgVCwOPjdAvBbI+e0ocS3MFEvzG6uBQE3xDk3SzynTn
-jh8BCNAw1FtxNrQHusEwMFxIt4I7mKZ9YIqioymCzLq9gwQbooMDQaHWBfEbwrbw
-qHyGO0aoSCqI3Haadr8faqU9GY/rOPNk3sgrDQoo//fb4hVC1CLQJ13hef4Y53CI
-rU7m2Ys6xt0nUW7/vGT1M0NPAgMBAAGjQjBAMA4GA1UdDwEB/wQEAwIBBjAPBgNV
-HRMBAf8EBTADAQH/MB0GA1UdDgQWBBR5tFnme7bl5AFzgAiIyBpY9umbbjANBgkq
-hkiG9w0BAQsFAAOCAgEAVR9YqbyyqFDQDLHYGmkgJykIrGF1XIpu+ILlaS/V9lZL
-ubhzEFnTIZd+50xx+7LSYK05qAvqFyFWhfFQDlnrzuBZ6brJFe+GnY+EgPbk6ZGQ
-3BebYhtF8GaV0nxvwuo77x/Py9auJ/GpsMiu/X1+mvoiBOv/2X/qkSsisRcOj/KK
-NFtY2PwByVS5uCbMiogziUwthDyC3+6WVwW6LLv3xLfHTjuCvjHIInNzktHCgKQ5
-ORAzI4JMPJ+GslWYHb4phowim57iaztXOoJwTdwJx4nLCgdNbOhdjsnvzqvHu7Ur
-TkXWStAmzOVyyghqpZXjFaH3pO3JLF+l+/+sKAIuvtd7u+Nxe5AW0wdeRlN8NwdC
-jNPElpzVmbUq4JUagEiuTDkHzsxHpFKVK7q4+63SM1N95R1NbdWhscdCb+ZAJzVc
-oyi3B43njTOQ5yOf+1CceWxG1bQVs5ZufpsMljq4Ui0/1lvh+wjChP4kqKOJ2qxq
-4RgqsahDYVvTH9w7jXbyLeiNdd8XM2w9U/t7y0Ff/9yi0GE44Za4rF2LN9d11TPA
-mRGunUHBcnWEvgJBQl9nJEiU0Zsnvgc/ubhPgXRR4Xq37Z0j4r7g1SgEEzwxA57d
-emyPxgcYxn/eR44/KJ4EBs+lVDR3veyJm+kXQ99b21/+jh5Xos1AnX5iItreGCc=
------END CERTIFICATE-----

README.md

@@ -1,39 +1,106 @@
 ---
 title: OUROBOROS Kernel Mint
-emoji: 🟢
+emoji: 🪙
 colorFrom: green
-colorTo: gray
+colorTo: yellow
 sdk: gradio
-sdk_version: 4.44.0
+sdk_version: 6.17.3
 app_file: app.py
 pinned: false
+license: mit
+short_description: A 1B model forges GPU kernels. An un-gameable referee.
+tags:
+  - track:backyard
+  - sponsor:openbmb
+  - sponsor:modal
+  - achievement:offbrand
+  - tiny-titan
+  - best-agent
+  - minicpm
+  - triton
+  - gpu-kernels
+  - reinforcement-learning
+  - self-distillation
 ---
 
-# OUROBOROS Kernel Mint
+# 🪙 OUROBOROS Kernel Mint
 
-Make a GPU kernel. An un-gameable referee decides if it's real.
+**Build a GPU operation out of blocks. A 1-billion-parameter model writes a real Triton
+kernel for it. An immutable referee decides if it's real — and times it against PyTorch's
+production compiler. Beat the compiler, land on the leaderboard.**
 
-Pick the pieces of an operation (or a classic one), press **Mint**, and a 27B open model
-writes a real Triton kernel for it. An immutable referee compiles it, checks it is correct
-against PyTorch on adversarial inputs, and times it against PyTorch's production compiler.
-Beat the compiler and you land on the leaderboard. A green tick is earned, not asserted.
+A green tick here is *earned, not asserted*: every kernel is compiled, checked for
+correctness against PyTorch on adversarial shapes/dtypes/magnitudes, and timed with CUDA
+events against `torch.compile` max-autotune — live, while you watch.
 
-## Running
+## How to play
 
-The model and the referee run on a GPU behind a small Modal backend (see
-`../ouroboros-mint/mint_backend.py`). This folder is just the Gradio front-end.
+1. **Build** — snap blocks together (norm → +residual → activations), or pick a classic
+   (softmax, swiglu, geglu…). The numbers flowing through the machine update live so you
+   can see what each block *does*.
+2. **Mint** — a fine-tuned **MiniCPM5-1B** (OpenBMB) writes a fused Triton kernel for your
+   machine. Each attempt faces the referee: compile → adversarial correctness sweep →
+   benchmark. Failures are shown, not hidden.
+3. **Beat the compiler** — if your minted kernel is correct *and* faster than
+   `torch.compile` max-autotune, it goes on the leaderboard. The reigning champions 👑 were
+   minted by the big 27B model; try to dethrone them. **Pro mode** switches the smith to
+   the full **Qwen3.6-27B** kernel-writer.
 
-```bash
-pip install -r requirements.txt
-BACKEND_URL="https://<you>--ouroboros-kernel-mint-mint-mint.modal.run" python app.py
-```
+> First mint of a session takes ~30–60 s while the model wakes up (the backend scales to
+> zero). After that, mints take a few seconds.
 
-On a Hugging Face Space, set `BACKEND_URL` as a Space secret. The backend scales to zero
-when idle, so the first mint of a session waits ~1–2 minutes for the model to wake up.
+## Why this is interesting
 
-## What it teaches
+GPU kernels are the small programs that actually run neural networks; *fusing* several
+steps into one kernel is where much of real-world AI speedup comes from. Writing them is
+expert work — PyTorch ships a whole compiler (`torch.compile`) to do it automatically.
 
-GPU kernels are the small programs that run neural-network steps on the GPU; fusing several
-steps into one is where most real speedups in running AI come from. The point of the
-project is that the valuable, scarce thing is not the big model but the **honest referee**
-that verifies the work — which is exactly what mints each kernel here.
+The thesis of this project: **the scarce, valuable thing is not the big model — it's the
+honest referee.** Give a small model a verifier it cannot fool (correctness is a boolean,
+speed is a measurement) and let it learn from its own verified wins (RL self-distillation).
+The referee that mints your kernel in this Space is the same one that trained the models.
+
+## The receipts (all machine-checked, none self-reported)
+
+- **76 model-era kernels** beat `torch.compile` max-autotune on H200 — 69 of them
+  stability-gated across 5 fresh re-benchmark runs (mean of means **1.299×**, range
+  1.11–2.04×), plus 7 single-shot invention probes on never-trained problems.
+- **Across a 376-cell shape/dtype grid**, the trained kernels hold a **1.49× geomean**
+  vs max-autotune recompiled per cell — with the ~10% of losing cells *reported per-cell*,
+  not hidden.
+- **Faster than hand-written expert kernels** (Liger / Unsloth / Triton tutorial) on
+  swiglu, rmsnorm, relu2 and geglu under two test conditions; softmax/layernorm are
+  ties-within-noise vs the best fixed-schedule expert.
+- **The referee defends itself**: a 30-case selftest (gold kernels pass, subtly-wrong
+  kernels rejected, three anti-gaming exploits — bench-shape special-casing, output
+  memoization, input mutation — rejected by construction). Green on RTX 4090 and H200.
+- Honest bounds: these are reproducible *scheduling* wins on bandwidth-bound fusion ops —
+  not wins over cuBLAS/FlashAttention, not new algorithms.
+
+## How it was built
+
+- **Models:** OpenBMB **MiniCPM5-1B** (the default smith — genuinely tiny) and
+  **Qwen3.6-27B** (pro mode), both fine-tuned with the OUROBOROS pipeline:
+  SFT on verified seed kernels → **RL self-distillation** where the only reward is the
+  referee's verdict (correct + faster = reward; no human labels anywhere).
+- **Modal** is used for **both development and runtime**: the 27B was trained on Modal
+  H200s (the RL run peaks ~110 GB VRAM), and the live mint backend runs on Modal
+  (L4 for the 1B smith, A100-80GB for pro mode) with scale-to-zero between sessions.
+- **Frontend:** a single Gradio Space whose whole interactive surface is a custom JS
+  machine-builder (snap blocks, live number-flow animation) bridged to Python — no
+  default Gradio look anywhere.
+- **Verifier:** immutable Triton/PyTorch harness — allclose vs PyTorch on adversarial
+  inputs (incl. the benchmark shape), CUDA-event medians vs `torch.compile`
+  max-autotune, anti-memoization poke + verify-after-bench, mutation checks.
+
+## Links
+
+- 📄 Field notes / write-up: *coming with submission*
+- 🤗 Fine-tuned models + verified-kernel corpus: *being published — links land here*
+- 🎬 Demo video: *coming with submission*
+- 📣 Social post: *coming with submission*
+
+---
+
+*Backyard AI track. Making small models fast is the backyard problem of everyone who runs
+them — this is my backyard, and the referee is the product.*

requirements.txt

@@ -1,2 +1 @@
-gradio>=4.44
 requests>=2.31