The daily surface of verifiably: the layer people and agents use
The commands a researcher or a coding agent runs in place of a folder of scripts and result tables: state a question as a typed claim, hold the data, declare the analysis, run it sealed, verify it from scratch, and read what the checked results add up to. Its scientific records live in a beliefs corpus, checked by the kernel, so the work stays traceable whoever did it. A CLI and an MCP server.
Authority: docs/commands.md, docs/sessions.md
Part of verifiably: verifiably builds understanding of the world from small, reproducible data analyses: each one a precise question asked of real data, recorded so anyone can run it again, with the data leading and unverified claims kept out. Overview
atoms · nodes · refs · beliefs · science · autonomy
From a checkout, with atoms, nodes, and beliefs checked out beside it as sibling
repositories (python/pyproject.toml takes beliefs from ../beliefs, which takes
atoms and nodes from ../atoms and ../nodes):
git clone https://github.com/verifiably/atoms
git clone https://github.com/verifiably/nodes
git clone https://github.com/verifiably/beliefs
git clone https://github.com/verifiably/science
cd science && git config core.hooksPath .githooks
cd python && uv sync --frozenFrom the repository root, just test-one <pytest args> runs a focused test selection,
just check runs the hygiene checks, and just test-full forces the full test suite on
a certified host (see AGENTS.md for the test execution policy and hosted CI split).
Start with Using science: explicit configuration, a synthetic source-checkout example through its verdict, and the path for your own data and analysis. Writes require certified storage; runs require normal confinement. The example uses public fixtures and demonstrates plumbing, not a biological finding.
From python/ after installation:
uv run science status --config /absolute/path/science.toml
uv run science build
uv run science adapters build --out /absolute/path/generated-adaptersWorld commands read --config or SCIENCE_CONFIG; configuration is never
discovered from the working directory. Initialize/admit roots through the
kernel's write protocol,
then name them in configuration. CLI writes use an
attended science serve --config <file>; science mcp serve --config <file>
provides the same command dispatcher over stdio. See sessions.
The command surface is recorded in tools/cli.toml:
- Analysis:
dataset,claim,spec,run,assess,verify,belief,next,recheck. - Coordination:
project,question,hypothesis,task,decide,revise,reuse,project-select,projects,project-show. - Sharing:
publish,publish-preview,provider-inspect,provider-bind,adopt,fetch,show. - World and tooling:
status,build,epoch,serve,mcp,adapters.
Scientific writes are kernel acts recorded in beliefs corpora. Selection and
sharing leave session-ledger evidence; science's registry projects sharing lines.
Historical sharing commentary is retained separately beside the ledger for exact
replay. Sharing explains acceptance and those retention costs.
belief computes its answer: honest NoBelief is distinct from zero or a refusal.
With the generated science plugin installed, /science:sweep-assumptions
compares research questions across explicitly named permitted inputs. It preserves
discovery before feasibility assessment in an eight-part report. It makes no
scientific record writes or biological analyses. Authored skills are product inputs;
the adapter build includes them in the harness plugin.
commands/— one declaration and prompt per command.skills/— authored research workflows.adapters/— generated harness adapters.python/— package, dispatcher, CLI, MCP and tests.docs/— current public references. Task records and historical working documents live in the maintainer's separate records home; application use does not need it.
- Guide — setup, analysis, selection, sharing and refusals.
- Configuration — keys, paths, contracts and mount profiles.
- Commands — declaration schema, handlers, reports and cursors.
- Sessions — launch, invocation, replay and transports.
- Sharing — publication, pins, acceptance and retirement.
- Stores — inherited stores, admission, relocation and recovery.
- Development — changes, tests, adapters and snapshots.
- Kernel guide — scientific record and belief concepts.
MIT, see LICENSE.