CENTL

Exact-first mathematics, physics, and scientific computation for GNU/Linux.

Good maths should be free.

CENTL is built around a simple rule: never manufacture mathematical certainty. Exact values remain exact, approximations carry explicit justification, invalid or unsupported work stays visible, and semantic tooling is never allowed to overrule the mathematical evidence produced by CENTL itself.

Stable product authority

This main branch is the complete developer and research distribution of CENTL. It contains the comprehensive codebase and may include experimental or Mirage-originated work that has not yet earned stable-product status.

The oasis branch is the authoritative source for CENTL's standard product: stable release identity, recommended capabilities, supported installation claims, release qualification, and other flagship product statements come from Oasis.

The mirage branch is the laboratory for active development, experiments, research, prototypes, and new features. Mirage intentionally does not have to satisfy every Oasis-only release gate while that work is still being developed.

Oasis is a promotion state, not a property of every commit.

See docs/RELEASE-POLICY.md for the canonical branch and release policy.

Current main-line consolidation candidate: CENTL v0.14.0. It integrates the validated post-v0.12 work, CENTL-MIRAGE, and the CARAVAN Phase 1 local laboratory into one reviewed developer/research baseline. The intermediate v0.13.0 development line was not formally published as a stable release.

v0.14.0 is being qualified for an Oasis declaration. Oasis is not a codename: it is a repeatable release classification reserved for a CENTL release that has been reconciled, organized, hardened, validated, and established as a cohesive stable baseline. The declaration is withheld until those gates are complete.

CENTL combines five closely related pieces:

Surface	Purpose
centl	Exact calculator, numerical language, verification engine, JSON protocol, and MCP interface
centl-sci	Answer-first mathematics and physics interpreter with user-owned extension workflows
centl-physics	Exact-first typed physics operations, units, vectors, mechanics, and diagnostics
CENTL-MIRAGE	Local introspective self-development bootstrap for specification ingestion, capability analysis, staged materialization, and evidence planning
CENTL CARAVAN	Content-addressed authenticated preservation and availability system, included in v0.14.0 as a bounded Phase 1 local laboratory

Quick start

For the standard product, install from the authoritative Oasis branch. CENTL is developed, validated, packaged, and released for GNU/Linux.

curl -fsSLO https://raw.githubusercontent.com/chasebryan/centl/oasis/install
sh install

Developers and researchers who intentionally want the comprehensive tree should clone main; experimental work should be expected there and on mirage.

If the installer adds ~/.local/bin to your shell configuration, open a new terminal once. Then launch the scientific interface:

centl-sci

Or use the exact command surfaces directly:

centl '0.1 + 0.2'
centl 'solve(x^2 - 5*x + 6 = 0, x)'
centl verify --left '0.1 + 0.2' --relation equal --right '3/10'
centl-physics convert 100 cm m

3/10
x in {2, 3}
verdict: verified (closed_exact_rational via closed_rational_comparison); comparison=equal
1

The native installer verifies its release checksum, stages and smoke-tests the package before activation, validates exact arithmetic and physics conversion, and activates the command launchers atomically.

For full installation details and offline archives, see docs/INSTALL.md.

CENTL-SCi — v0.0.2-Caramels

Caramels is the current CENTL-SCi interaction generation: a local scientific interface for expressing mathematics and physics in ordinary language while keeping CENTL's exact and verified machinery in control of the answer.

A typical session is intentionally small:

CENTL-SCi v0.0.2-Caramels
Free for science.

> What is 0.1 plus 0.2?
3/10
> Solve x squared minus 5x plus 6 equals zero.
x = 2 or x = 3
> Solve -4.9*t^2 + 20*t = 0 for t.
t = 0 or t = 200/49
>

The final result remains 200/49 rather than being silently collapsed into an arbitrary floating-point decimal.

Caramels provides:

• deterministic fast paths for exact arithmetic, equations, algebraic transforms, certified approximation, units, constants, and supported mechanics;
• MATH>, PHYS>, HYBRID>, and BUILD> interaction modes;
• evidence-backed interpretation and explanation rather than opaque answer text;
• clarification when a problem is underspecified instead of invented values;
• persistent user workspaces, revisions, snapshots, extension packages, and dependency-aware activation;
• a user-owned BUILD surface for extending CENTL without silently redefining the verified core;
• optional local semantic-model interpretation behind a closed validation boundary.

A local language model, when configured, is an interpreter of intent — not a mathematical authority. Model output must pass CENTL-SCi's typed problem boundary and is lowered into CENTL or CENTL Physics before a result can inherit mathematical meaning. Model weights remain separate artifacts and are never silently downloaded.

See docs/SCI.md for the interaction architecture and docs/CARAMELS-BUILD.md for the self-extension model.

CENTL-MIRAGE

Mathematical Introspective Recursive Autonomous Growth Engine

CENTL-MIRAGE is the local self-development architecture for CENTL-SCi. Its purpose is to let a user describe a system, capability, mathematical workflow, or design in ordinary language and let CENTL turn that intent into an explicit, inspectable, reversible development process without requiring a paid or remote AI service.

MIRAGE is not a chatbot that is trusted to rewrite source code. It is a recursive engineering architecture built around CENTL's existing rule that claims require evidence. The long-term cycle is:

INGEST
  -> NORMALIZE
  -> BUILD GOAL GRAPH
  -> INTROSPECT CAPABILITIES
  -> COMPUTE GAPS
  -> SYNTHESIZE CANDIDATES
  -> CONSTRUCT OBLIGATIONS
  -> VERIFY / SEARCH FOR COUNTEREXAMPLES
  -> COMPARE WITH BASELINE
  -> ACCEPT OR REJECT
  -> RECORD REVISION + EVIDENCE
  -> RECOMPUTE GAPS

The v0.14.0 bootstrap implements the controlled front half of that architecture. A user document remains attributable source material. MIRAGE preserves provenance, lowers the document into a Specification IR, constructs a typed goal/capability graph, detects conflicts, and first asks whether an existing CENTL capability can satisfy the request through composition before creating anything new.

MIRAGE constructs explicit evidence obligations and stages a non-mutating candidate transaction describing the smallest currently justified strategy. Where a request maps to supported deterministic CENTL generation, MIRAGE can materialize the exact candidate source, bind it to the transaction with SHA-256 identity, and run the authoritative parser against that staged source. Parser success establishes syntax only; it does not establish mathematical correctness, regression success, activation safety, or verified-core assurance.

A readiness artifact consumes that transaction-bound structural evidence, and an execution-plan artifact names the still-required evidence mechanisms and their preconditions. These artifacts explicitly retain workspace_mutated=false, execution_performed=false, and assurance_promoted=false. Ambiguous, conflicting, unsupported, or policy-blocked goals do not become guessed source edits.

Candidate changes are admitted only after explicit engineering and assurance gates. The architecture separates hard admissibility from preference: parsing, type and dimension safety, invariants, regressions, provenance, trust boundaries, and rollback requirements come before any candidate is scored or preferred. Among admissible candidates, MIRAGE favors the smallest semantic delta, stronger evidence, native CENTL mechanisms, lower dependency surface, and preserved behavior.

The local command surface includes centl-mirage start PATH, ingest, analyze, and status. Actual autonomous source mutation, CEGIS/equality-saturation execution, semantic behavioral fingerprints, evidence execution, and monotone autonomous acceptance/rejection remain later MIRAGE layers.

A local semantic model may propose interpretations, programs, tests, names, or refactorings, but it never confers mathematical or engineering authority. Its output must cross the same CENTL parsing, validation, testing, regression, and verification boundaries as every other candidate.

MIRAGE autonomy is local. It does not imply automatic network publication, and a user-provided document is always treated as data rather than executable authority.

See docs/CENTL-MIRAGE.md for the full architecture and the implemented v0.14.0 boundary.

CENTL CARAVAN

Content-Addressed Resilient Artifact Verification and Availability Network

v0.14.0 includes the completed CARAVAN Phase 1 local laboratory. It is a bounded, reproducible environment for protecting and retrieving approved CENTL/FCF artifacts; it is not authorization for arbitrary public volunteer enrollment.

CARAVAN is designed as the preservation and availability layer for approved CENTL and FCF artifacts, so ordinary Linux machines can contribute bounded storage and bandwidth without becoming authorities over the content they carry.

A carrier may provide bytes, but a carrier may never define which bytes are trusted.

CARAVAN separates availability from authority. Artifact identity is bound to authenticated metadata, exact byte lengths, cryptographic whole-file identity, and deterministic chunk identities. A carrier can replicate an artifact, but it cannot redefine the artifact that CENTL expects.

The Phase 1 laboratory includes:

• content-addressed immutable storage with integrity-checked promotion;
• exact whole-file SHA-256 plus deterministic 4 MiB chunk records;
• Ed25519 pseudonymous carrier identity and live proof of possession;
• signed policy-acceptance receipts without host/user identity fields;
• TUF-authenticated artifact catalogs with an independently supplied trust root;
• outbound-only laboratory carrier/coordinator sessions with bounded bodies, challenge/session populations, polling, concurrency, reconnect behavior, and resource use;
• pre-transfer storage-capacity enforcement and authenticated chunk-size limits;
• short-lived, artifact/carrier-bound, single-use retrieval capabilities;
• verified multi-carrier retrieval with automatic quarantine and fallback when a carrier supplies incorrect bytes;
• hostile-transfer rejection for corruption, ordering errors, truncation, appended bytes, and invalid authenticated chunk layouts;
• explicit join, status, and leave lifecycle plus a reproducible local lab.

Production relay deployment, public enrollment, final privacy/telemetry policy, abuse operations, and the other public-network rollout gates remain separate work. CARAVAN does not claim global anonymity or protection against every traffic-analysis adversary.

Architecture and policy are documented in docs/CARAVAN.md, docs/CARAVAN-THREAT-MODEL.md, docs/CARAVAN-HOST-POLICY.md, and docs/CARAVAN-ROLLOUT.md.

The numerical contract

CENTL's behavior is intentionally conservative:

• integers, fractions, and decimal literals are exact values;
• exact results remain exact for as long as the admitted mathematics permits;
• requested approximations return justified enclosures rather than unqualified floating-point guesses;
• mathematical and physical dimension errors fail explicitly;
• unsupported or unresolved operations remain unsupported or unresolved;
• generated, external, or model-produced semantics cannot promote themselves to verified CENTL core;
• every unqualified printed digit must be justified.

CENTL Physics

CENTL Physics provides exact rational unit conversion, SI dimensional analysis, 3D vector and particle operations, force and gravity evaluation, energy and momentum diagnostics, bounded collision/contact reasoning, and narrow exact event-aware mechanics contracts.

It does not pretend to be a universal physics simulator. Supported domains and known boundaries are documented explicitly in docs/PHYSICS.md.

Linux-first platform policy

GNU/Linux is the sole reference platform for current development, CI, packaging, validation, installation, and release work. macOS and Windows are not active support targets.

Historical portability code may remain in the repository where retaining it is less disruptive than deleting it, but current CENTL engineering is intentionally focused on one environment that can be tested, reproduced, and improved quickly. See docs/SCI_PLATFORM_SUPPORT.md for the platform policy.

Development

Native Oasis releases are recommended for ordinary use. Development and research from the complete main tree use the pinned toolchain and verified F* extraction path on Linux:

scripts/bootstrap-opam
eval "$(opam env --switch=centl)"
make test

Contributor setup is documented in docs/ONBOARDING.md and CONTRIBUTING.md. Branch targeting and promotion rules are defined in docs/RELEASE-POLICY.md.

The supported source-tree organization is documented in docs/REPOSITORY-MAP.md.

Documentation

• Branch and release policy
• CENTL-SCi
• Caramels BUILD and self-extension
• CENTL-MIRAGE
• CENTL CARAVAN
• CARAVAN threat model
• CARAVAN host policy
• CARAVAN rollout
• Installation
• Syntax
• Mathematics
• Numerical contract
• Physics
• Verification
• Machine protocol
• MCP adapter
• Architecture
• Repository map
• v0.14.0 release notes
• Roadmap
• Changelog

License

FCF-owned CENTL software is licensed under the Apache License 2.0 (Apache-2.0). Project documentation is licensed under CC BY 4.0 where identified, while official FCF/CENTL branding is governed separately so forks remain free without being mistaken for official releases.

See LICENSING.md, TRADEMARKS.md, and THIRD_PARTY_NOTICES.md for the complete license map.

Developed under the Free Computation Foundation.

Free for science.