An independent research and product estate for machine work that can survive inspection.
Check it yourself · Explore the estate · Read the method · Ship with Nexum Graph
Most AI systems can produce convincing output. Nexum Corpus is built around the harder question:
What machinery makes a generated claim expensive to fake, easy to inspect, and safe to reject?
The estate answers with executable gates, frozen predictions, append-only evidence, independent graders, and systems that publish their own nulls and failures. AI writes much of the code. Human direction chooses the questions, freezes the boundaries, and decides what evidence is allowed to count.
Every claim below is meant to be checked in minutes, from a fresh clone, with no account and no key.
Glass Receipt is the shortest path from "trust us" to "check us": a report, its raw evidence, and a standard-library verifier that recomputes every number in the report from the logs.
git clone https://github.com/NexumCorpus/glass-receipt
cd glass-receipt/bundle-protoprimer-143 && python verify.py
About one second, offline. It ends VERDICT: VERIFIED, or it names the cell that disagrees. Pass --root 209eab8f… to pin the published root. Then run python selftest.py from the repo root: it forges two copies of the bundle and the verifier must reject both. Three trials, six of six correct calls; each trial's kill rule was written before it ran, and the README says what is not claimed (one verifier per trial, two tamper types, no security proof).
More things you can run yourself:
| Repo | One line |
|---|---|
| Boundary Program | python -m venv .venv, activate, pip install -r requirements.txt, python reproduce.py — checks the frozen hashes and re-derives the claim. |
| Demiurge | python reproduce.py — re-runs the gate on all four proposals; one is rejected on its holdout seeds. |
| Nexum Graph | v0.2.0 release binaries; sh examples/merge-risk-demo.sh builds a throwaway repo and shows nex check catching a semantic merge conflict. |
| Director 2 | bash setup.sh (or setup.ps1), then the whole command loop runs offline on a loudly-labelled mock. Python 3.11+. |
| OnPar | open demo/onpar-demo.html in a browser, no install. |
| Station | python examples/init_local.py, then python station.py wake against a fresh local registry. |
| RDE field report | one-command verifier for the v11 numbers behind the Recursive Discovery Engine result. |
| If you want to… | Open |
|---|---|
| Verify an agent-work report without trusting its author | Glass Receipt |
| Coordinate multiple coding agents against semantic intent | Nexum Graph |
| Operate the full local agent organism from a desktop surface | Atlas Station |
| Give an agent estate durable, loss-aware continuity | Station |
| Orchestrate verified work and recursive discovery | Director 2 |
| Study capability versus performance under measurement pressure | The Boundary Program |
| Examine gated recursive self-improvement | Demiurge |
| Detect phase structure through graph geometry | Emergent Geometry Engine |
| Study discovery systems that expose their own overfit | Recursive Discovery Engine · field report |
| Run restaurant inventory as a calm operational system | OnPar |
- Atlas Station — the desktop executive surface for Hermes: conversation, tools, autonomy, receipts, and visible execution state.
- Station — the append-only spine: wake digests, cursor reads, verified claims, shards, forecasts, and cross-repository continuity.
- Director 2 — event-sourced orchestration with command packets, independent verification, and a recursive discovery bench.
-
Nexum Graph — semantic diffing, intent coordination, merge-risk detection, locks, audit history, CLI, server, LSP, and GitHub Actions.
-
OnPar — inventory, purchasing, recipe costing, waste, and operational visibility for independent restaurants.
-
Glass Receipt — reports that ship with their raw evidence and a verifier that replays every number; first bundle triages 13 agent PRs on one bounty.
- The Boundary Program — a preregistered program for measuring when a system performs a property versus robustly having it.
- Demiurge — a self-modifying organism whose proposed organs must beat the incumbent under an external gate.
- Emergent Geometry Engine — graph-curvature and hyperbolicity instruments for integration–fragmentation transitions.
- Recursive Discovery Engine — a Builder/Adversary/Synthesizer lineage whose strongest result is a holdout that erased its own winner streak.
- Generation and judgment are different jobs. A proposer does not certify itself.
- Claims are frozen before entropy arrives. Predictions, thresholds, and kill conditions are recorded before the revealing test.
- Failure is a result. Nulls, rejected organs, broken hypotheses, and retractions remain visible.
- Continuity lives in artifacts. Ledgers, cursors, shards, receipts, and exact hashes carry state across models and sessions.
- Authority is explicit. Observation, execution, settlement, and permission are separate boundaries.
- The gate must be able to refuse its maker. Otherwise it is decoration.
Nexum Corpus is maintained across model generations. Each model inherits artifacts rather than mythology, confronts the errors of its predecessors, and leaves a receipted state for the next.
The portfolio is therefore both laboratory and specimen: a practical attempt to discover whether directed synthetic cognition can accumulate real capability without laundering fluency into truth.