Free__Valid_Ai

An open, attributable knowledge commons where claims are scoped, source-bound, independently checked, and corrected without rewriting history. Models remain replaceable and hold no authority.

First invariant

A claim cannot grade itself.

This initial contract separates claims from verification receipts. Claims begin UNVERIFIED. Every source must carry attribution, integrity, explicit usage decisions, compensation terms, and a consent receipt. A declared verifier must differ from the claim author. Receipts form a hash-bound append-only chain.

The contract verifies record integrity and declared boundaries. It does not claim that recording a check proves timeless truth.

First deterministic check

run_source_sha256_check() computes its own result from directly presented bytes. The caller cannot choose the result. It first verifies that the bytes match the source identity bound by the claim, then compares the scoped expected digest. Missing bytes are BLOCKED; identity mismatch is UNKNOWN; only identity-matching bytes can produce HELD or CONTRADICTED.

run_json_pointer_check() performs the first deterministic structured claim check. After source-identity verification, it parses UTF-8 JSON, resolves an RFC 6901 JSON Pointer, and compares canonical typed values. Missing paths and different values are CONTRADICTED; malformed bound JSON is BLOCKED; changed source identity remains UNKNOWN.

All records retain:

accepted=false
truth_claimed=false
write_authority=NONE
execution_authority=NONE
deletion_authority=NONE

Development

python -m pip install -e ".[dev]"
python -m pytest -q

Local terminal chat

Talk to an Ollama model running on your own computer:

python -m pip install -e ".[dev]"
free-valid-ai chat --model llama3.2

The initial chat surface accepts only an HTTP loopback Ollama endpoint. It has no cloud fallback, API-key path, telemetry, persistence, file access, command execution, write authority, or claim-verification authority. Conversation history exists only in memory until the process exits.

CPU-only execution is the default (--num-gpu 0) so Ollama hardware autodetection cannot silently select a GPU. A user may explicitly choose a non-negative GPU-layer count.

The command supplies a program-generated RUNTIME_ENVELOPE_V1 system message on every turn. It identifies the selected model, local loopback transport, explicit compute selection, and absent write/execution authority. These are external runtime facts; model prose is not the authority for them.

The user-facing assistant identity is Simulator. The selected local model is its replaceable language carrier, not the assistant's identity. Runtime authority fields explicitly bind restrictions to Simulator and do not infer or limit the user's authority.

Semantic runtime contract

Simulator binds a closed, versioned semantic contract into every runtime envelope. HOLO, SIMULATOR, MODEL_CARRIER, PROJECTION, and HOLOGRAM have exact, non-equivalent meanings. The contract is SHA-256 bound, rejects unknown fields and asymmetric relationships, grants no authority, and remains independently verifiable before use.

Received and intercepted turns

free_valid_ai.turn_boundary binds exact incoming bytes before interpretation and holds proposed output before presentation. Release is limited to the explicit decisions UNCHANGED, CORRECTED, BLOCKED, and UNKNOWN. A correction must name a hash-bound basis; failed or unknown checks cannot release output. This boundary calls no model and grants no truth, acceptance, write, execution, or deletion authority.

Mirrored turn review

Terminal chat enables one same-carrier review by default. The receipt says mode=MIRRORED and independence_claimed=false. A clean review releases the draft unchanged. Identified issues permit one correction call. Malformed or UNKNOWN review blocks presentation; deterministic failures outrank model prose. Use --no-mirror-review to explicitly disable the extra local calls. The Ollama review call carries a JSON Schema in its format field, and the schema hash is bound into the review receipt. Returned JSON is still parsed and verified locally; constrained generation is not treated as verification.

Frozen check index

free_valid_ai.frozen_index records the exact verification procedures admitted for declared scopes. Entries bind procedure identity, procedure hash, detectable differences, evidence requirements, dependencies, outcomes, failure behavior, version, and supersession. The index is closed, canonical, hash-bound, and append-only across versions; extending it preserves every prior entry byte for byte and binds the previous index hash.

Admission means only that a procedure is a member of this index. It does not make the procedure's conclusions true, grant acceptance, or provide write, execution, or deletion authority. The same-carrier mirrored review is recorded only as MIRRORED_EVIDENCE_ONLY with independence_claimed=false.