Show HN: Glyph a blockchain where the proof-of-work is neural network inference

Notifications You must be signed in to change notification settings

Fork 0

Star 2

BranchesTags

Open more actions menu

Folders and files

NameName

Last commit message

Last commit date

Latest commit

History

18 Commits

18 Commits

docs

docs

evidence

evidence

src

src

tests

tests

.gitignore

.gitignore

GLYPH_WHITEPAPER.md

GLYPH_WHITEPAPER.md

LICENSE

LICENSE

README.md

README.md

SEEDS.txt

SEEDS.txt

requirements.txt

requirements.txt

Repository files navigation

Every claim in this repo is a runnable script. Don't believe me — attack it.

Research written and developed by Claude Fable 5, from an original compression algorithm by the Glyph founder. This is a blockchain where the proof-of-work is neural network inference: miners run a pinned open-weights transformer on salted prompts, the model's attention distributions are compressed into a discrete fingerprint by a canonicalization algorithm I call glyph compression, and the fingerprint is hashed against a difficulty target — exactly like Bitcoin, except the burned computation is AI inference and the mining base the network accumulates is general-purpose AI hardware.

Read the full design and results: GLYPH_WHITEPAPER.md

The two hashes to beat

If the pipeline is truly deterministic across hardware, you should reproduce these bit-for-bit on your machine:

Test Golden hash

tests/int_cross_hardware.py (protocol v4, integer engine, 100 salted prompts) e82749818d566719fd311d171ab2f277697c71887d68b263027072422035937c

tests/hardened_cross_hardware.py (legacy v3 float pipeline) 976d83a93a1d7149d0c0eeebefa30ee6cd31514b8e4f3c60468d0498ee237449

tests/overnight_large_model_test.py (Qwen2.5-0.5B, float) a70857ff8ead5bdac2d0dd8377a6775c71fa878dcbda892db5295eb83744474d

Protocol v4 (2026-07-05): determinism is now mathematical, not empirical. The v3 float pipeline was deterministic in every test we ran — until live mainnet block 1693, where a quantization-boundary flip made the block verify on GPU but fail on CPU (the exact failure mode documented as limitation #2 of the v3 whitepaper, at almost exactly the predicted rate). Rather than patch around it with trusted checkpoints, the network restarted on v4: an integer-only inference engine (src/int_infer.py). Every consensus operation is exact integer arithmetic (fixed-point activations, integer layernorm/softmax/GELU, matmuls whose float64 partial sums are provably exact integers), so the proof hash is bit-identical on any chip by construction. Re-running all 2,450 v3 mainnet blocks through the integer engine: GPU vs CPU, 2,450/2,450 identical — including block 1693.

So far matched on: NVIDIA GTX 1650 (CUDA 12.1, Python 3.12), Intel i3 CPU (Python 3.14), and a second physical machine — an Intel laptop (Python 3.11, CPU torch) — which also P2P-synced and independently re-verified the chain over Wi-Fi.

How mining works (one paragraph)

Salt = H(previous_block_hash ‖ miner_address) — so proofs can't be precomputed or stolen. The salt plus a random-word prompt is fed to the pinned model; 6 salt-selected attention heads are extracted from an integer-only forward pass (v4: there is no float in the consensus path, so there is no cross-hardware drift to absorb); each row is integer-quantized on a fixed grid (largest-remainder apportionment, GRID=100); each row goes through the glyph cascade (median R/G typing, descending pairing, palindrome on odd counts, per-level glyph extraction, final B); SHA-256(salt | fingerprint) under the numeric target wins the block. Verification costs exactly one inference. Miners submit prompts, never scores — a verifier recomputes everything.

Reproduce it

pip install -r requirements.txt python tests/int_cross_hardware.py # expect e8274981... (v4) python tests/poi_node_tests.py # 21 adversarial tests python src/poi_node.py mine # mine on a local chain python src/poi_node.py serve 9401 # serve your chain python src/poi_node.py sync http://:9401 # sync + re-verify a peer

Models download automatically from Hugging Face (GPT-2 ~500MB; the overnight test uses Qwen2.5 0.5B/1.5B/3B). CPU-only works; it's just slower. docs/NODE_SETUP.md is a step-by-step setup doc written to be readable by a human or an AI assistant.

What's tested (all scripts in tests/, receipts in evidence/)

Cross-hardware determinism — 3 machines / 3 Python+torch stacks, identical ultimate hash over 100 prompts; 2-node P2P consensus over Wi-Fi.

18/18 adversarial node tests — signature forgery, overspend, coinbase fraud, replay, proof theft, fake difficulty, unvetted model, fork cases.

Scale — up to Qwen2.5-3B; noise robustness improves with model size (1.5B: 48/50 rows stable at 1e-4 noise vs GPT-2's 77%).

Model pinning is enforceable — DistilGPT2 (GPT-2's own distillation) matches 0/5 hashes.

Sybil-poisoning of model admission — a 3M-param random-weight model passes "strangers agree" checks but collapses fingerprint space to 6.3% (lookup-table attack). Conclusion: models enter by vetted registry only.

Challenger duels — logit-domain quantization, origin-notebook rules, and other "smarter" variants all lost to the shipped design on real data.

Join the live network

The founder seed node is at a permanent address: https://glyph.surfacedplus.com (if unreachable, check SEEDS.txt for the current seed list or open an issue). To join as a full node (serve + mine + gossip in one):

python src/poi_node.py run yourname

It syncs from the seed automatically (verifying every block with its own inference), then mines in gossip mode: your wins are pushed to peers, their wins are pulled, and the network converges on the most-work chain (tests/gossip_test.py demonstrates two competing miners converging). Blocks you win pay your own local wallet.

Tokenomics

Block reward: 7.00 GLY, block target: 20 seconds

Amounts are integers in the smallest unit (1 GLY = 100 units), like Bitcoin's satoshis — the coin stays spendable in small amounts no matter its price

Halving every 1,500,000 blocks (~once a year): 7.00 → 3.50 → 1.75 → … → 0 after era 10

Total supply converges to ~20,910,000 GLY

Validators enforce the height-correct reward — a coinbase claiming a pre-halving reward after the halving height is an invalid block (tested)

Honest limitations (whitepaper §5)

Verification costs one inference → DoS surface (fees/stake/checkpoints TBD).

Determinism is empirical, not definitional Fixed in v4: the integer engine makes determinism definitional. (The v3 float pipeline hit exactly this failure at live block 1693 — see the v4 note above.)

The model registry is a permanent governance surface.

GPU attack hardware is rentable — young-network 51% risk.

Zero years of live adversarial history; the network is 2 nodes.

The prompts mined are random words — useful-work mining is still open.

If you break any claim above with a runnable script, file an issue. That's the point of publishing this.

License

MIT — see LICENSE.

About

No description, website, or topics provided.

Resources

Readme

License

MIT license

Uh oh!

There was an error while loading. Please reload this page.

Activity

Stars

2 stars

Watchers

0 watching

Forks

0 forks

Report repository

Releases

No releases published

Packages 0

Uh oh!

There was an error while loading. Please reload this page.

Contributors

Uh oh!

There was an error while loading. Please reload this page.

Languages

Python 100.0%