GEODE — Generalized Encoders for Open-Domain Expertise

GEODE is a network where anyone can contribute a capability and anyone can buy answers, priced in Ethereum at market rate. A capability is a frozen neural network — weights that no longer change — or a small deterministic program called a primitive. Capabilities are composable and routable: a contribution can be reused and built upon by anyone, every use pays the work beneath it, and payment follows the measured work that actually served the query. Every decision is deterministic, replayable, and recorded on a hash-chained ledger.

GEODE claims no new learning algorithm — its parts are old and named. What it claims is the architecture: frozen representation artifacts composed by declaration on a versioned code bus, upgraded without invalidation, and paid by measured downstream use — the composed-codes architecture. The wager: collaboration is the fastest path to advanced AI. Tasks break into smaller composable pieces that are cheaper to train and cheaper to run; competitors build on each other's work and share the rewards by measured use; a network anyone can improve compounds faster than central labs that fossilize under secrecy and intellectual-property protection.

For further information, read the whitepaper — the design, the measured results, and the honesty boundaries: docs/WHITEPAPER_GEODE.pdf (LaTeX source: docs/WHITEPAPER_GEODE.tex).

⚠ Development status — do not deploy yet. This repository is a research implementation under active development. The whitepaper describes the design; the code here is the working implementation of it, and the two are kept in line as issues are found and fixed. It is not known to be bug-free: the test suites pass (Python and EVM), but passing tests are not a proof of correctness. Do not run a GEODE network, post real funds to any contract, or rely on this code for anything in production. If you want to run it for experimentation, do so on a local or test network with test funds only, and expect breaking changes without notice.

The architecture in one idea

The name says it: generalized encoders. A registry of frozen encoders — each mapping one kind of input (image, audio, text, number series) into a code space — is the spine. A head declares a code manifest: an ordered list of the blocks it reads, concatenated into one design. Nothing beneath it is rebuilt: a closed-form fit on the composed codes is measured, priced, and routed on its own.

• Frozen. Every model is a frozen publisher checkpoint. Heads are closed-form solves — no epochs, no optimizer, no random seed.
• Composable. Codes concatenate additively; a new task enters as a measured addition, never a new foundation. Optional closed-form alignment bridges (Procrustes/CCA) are registered, measured, and priced per pair — never load-bearing.
• Open contribution surface. Learned representations — adapters, projections, feature maps, bridges — are registrable artifacts with an input contract, a measured utility, and a price, exactly like any arm. The free standard library holds code-defined transforms only; it never holds learned models.
• Replayable. The same declared task and input reproduce the same routing decisions and the same answer, re-runnable from the hash chain by anyone who holds the artifact.
• Paid by measured use. Payments flow to the contributors whose work actually served the query, vested over epochs, with a small development-fund share. Contribution is measured on held-out data, never self-reported.

What has been measured

Held-out measurements of an assembly of published parts, sealed with content hashes and reproducible from the recorded artifacts. Single-corpus and single-party — a working demonstration, not a competitive claim. The whitepaper reports the failures as loudly as the wins.

Axis	Measurement	Result
Routing	DomainNet: the router picks the right specialist / the answer the user receives	0.91 / 0.76
Code	Qwen2.5-Coder 7B, HumanEval pass@1 / pass@3	0.860 / 0.884
Speech	Whisper ladder, LibriSpeech test-clean WER	0.0296
Vision (scoped serving)	Open Images: 129 served classes, 472 refused	0.901 on the served subset
Fusion	clean cell: two frozen encoders concatenated vs one	0.548 vs 0.242
Nonlinearity	quickdraw wall (four frozen backbones) vs + hash-seeded random features	0.63 → 0.675
Text recipe	frozen BERT-base + closed-form ridge: SST-2 / IMDb	0.857 / 0.828
Audio recipe	frozen wav2vec2-base + ridge, Speech Commands v2	0.879
Number series	in-house temporal arms, Mackey-Glass one-step NRMSFE	0.0032

One honesty statement organizes the whole project: the protocol machinery is built and measured, and the generalized-encoder thesis is an open experimental program with measured boundaries — where a measurement fails, the failure is published.

Install

pip install .            # the product package (numpy + torch)
pip install '.[api,dev]' # plus the local HTTP API and the test tooling

Then python examples/hello_geode.py (the five-minute tour) and the geode CLI (geode route --fp 0.9,0.3,0.2,0.1). See archive/research_2026-08-28/docs/QUICKSTART.md. The API is local-only by design.

• Whitepaper: docs/WHITEPAPER_GEODE.pdf — the canonical whitepaper: the shipped system, the design decisions and why, the held-out measurements and metrics, the safety and economics, and the prior art — with every technical term explained at its first use. The LaTeX source ships beside it (docs/WHITEPAPER_GEODE.tex). A plain-language markdown copy existed through 24 Aug 2026; it was deleted from the public release rather than maintained stale, and the .tex/.pdf pair is the single source of truth.
• Deploying and using: archive/research_2026-08-28/docs/DEPLOYMENT.md (the deployment guide in plain language) and archive/research_2026-08-28/docs/QUICKSTART.md (the five-minute tour).
• Active plans: analysis/RESEARCH_IMPLEMENTATION_PLAN_v25.md (the sealed execution log; M215/M216 record the product architecture); the earlier phases live in archive/research_2026-08-28/analysis/.
• Architecture & testing: archive/research_2026-08-28/docs/ARCHITECTURE.md (layers and dependency rules) and archive/research_2026-08-28/docs/TESTING.md (the test pyramid).
• Lessons & limits: archive/research_2026-08-28/analysis/LESSONS_ARCHIVE_v22.md (epistemic discipline).

---

Repository map

docs/                     # the canonical whitepaper (.tex + built PDF)
analysis/                 # the sealed v25 execution log
archive/                  # research-phase material superseded by the
  research_2026-08-28/    #   public release (plans, threat reviews,
                          #   sealed logs, tools, docs)
  ...                     #   the earlier legacy archive (archive/README.md)
experiments/              # measurement harnesses and configs
  tier4/                  #   the sealed milestones' runners
  common/                 #   shared experiment utilities
  configs/                #   registered configurations
geode/                    # THE PRODUCT PACKAGE (public API in __init__.py)
  audit/                  #   replay, provenance, erasure
  core/                   #   descriptors, ontology, registry, router,
                          #   ledger, federation bus, economics
  attribution/            #   Shapley/beta-Shapley, incentives, pricing
  settlement/             #   the CreditLedger attribution wire
  privacy/                #   FHE serving, secret sharing, ZK arguments
  hashing.py              #   canonical JSON + payload hash
src/                      # runtime and geometric core modules
infrastructure/evm/       # EVM contracts + the local Hardhat harness
tests/                    # the test pyramid
  unit/                   #   single-concern, in-process
  integration/            #   multi-module flows
  system/                 #   the EVM harness gate
examples/hello_geode.py   # the five-minute tour
requirements.txt
pytest.ini

Rules of the road (from the evidence discipline):

• Register before measuring. Reproduce the anchor before any new number. A smoke run declares itself inadmissible. A void is not a negative. Everything frozen carries a payload hash.
• Layering: geode never imports experiments.*; the dependency direction is enforced by tests/unit/test_architecture_layering.py.
• $env:GEODE_CACHE_DIR points at the data/cache root (external to the repo); GPU runs use .venv-rocm with $env:HIP_VISIBLE_DEVICES="1".

Running a cell (example):

$env:GEODE_CACHE_DIR="F:\geode-ml\data\cache"; $env:HIP_VISIBLE_DEVICES="1"
& ".\.venv-rocm\Scripts\python.exe" -m experiments.tier4.eval_v23_m162_prune_retrain `
    --config experiments/configs/v23/m162_prune_retrain_smoke.json `
    --output logs/results/v23/m162_prune_retrain_smoke

Tests: python -m pytest runs the whole pyramid (unit / integration / system, marked per layer); -m unit selects one layer. The EVM harness runs as npx hardhat test in infrastructure/evm and is also covered by the system layer. See archive/research_2026-08-28/docs/TESTING.md. .venv-rocm is the only environment (the CPU-only .venv was deleted on 17 Aug 2026).

---

License

• Code (geode/, src/, experiments/, infrastructure/, tests/): GNU Affero General Public License v3.0 or later (LICENSE). Copying and deploying are permitted; anyone running a modified version as a public service must publish their modifications under the same license.
• Whitepaper and documentation: Creative Commons Attribution 4.0 International (LICENSE-DOCS).
• Publisher checkpoints are never redistributed; they are pulled from their publishers under the publishers' own terms.

---

Where the research phase went

The repository transitioned from research to development on 17 Aug 2026, and was prepared for public release on 28 Aug 2026. Everything superseded — historical plans, claim ledgers, research reports, threat reviews, sealed experiment logs, tools, docs, legacy tiers, and the legacy README — is preserved under archive/ (the post-17-Aug research material in archive/research_2026-08-28/, the earlier material in the rest of archive/; see archive/README.md).