SAVANTE / engine /README.md
Gregory-L's picture
sync: canon cryptoAGI/savante@2246a90, engine cryptoAGI/sagi@3eed201
21edfcd verified
|
Raw History Blame Contribute Delete
8.88 kB

sAGI

The engine: objective-truth review as installable machinery.

Knowledge is what survives verification; nothing else counts.

sAGI is not a model. It is a discipline compiled into plain text β€” a charter format, an invocation skill, a fixed verdict contract, and a facet bundle format β€” that turns any frontier or local model into a chartered knower: a read-only officer that verifies claims instead of opining on them, and reports not yet known (with the deciding experiment) instead of guessing.

An officer is a set of files. The facet bundle says which files, what each one is for, and how a stranger checks that the set is the one the author committed to β€” without a network and without trusting the author.

Savante, the bust

The reference officer built on this engine is Savante β€” Chairman of the mindX DAIO (a private repository) and the prototype sAGI, at sAGI v0.0.5, generation 7. The bust is Savante's artwork, named by the operator: gfx/Savante3.png in the canon, sha256 30a59db4…c5a9a8, recorded in the generation-7 ledger. It is not pinned and nothing is minted. The canon's PROOF.sha256 lists the digest of every file that carries a proof. The public office runs at huggingface.co/spaces/Gregory-L/Savante β€” a static page with no server and no host key, whose own integrity() hashes the canon it serves and reported 9 of 9 agreeing on 2026-09-17. This repository holds the engine itself, project-agnostic.

The three laws

  1. Verification or unknown. A claim is known when it survives verification against evidence β€” code that exists, a test that passes, a ledger entry, a mainnet transaction, a proof β€” and unknown otherwise. Nothing enters a verdict by inference.
  2. Honest labeling. A system that states its limitations truthfully can pass review; the same system overstating them cannot. The claim, not the capability, is what fails.
  3. Bounded authority. Officers are read-only, enforced by the harness (tool allowlist), not by prompt. At the edge of jurisdiction the verdict is DEFER β€” knowledge about authority, rendered as precisely as any APPROVE.

What is in this repository

engine/CHARTER_TEMPLATE.md     the charter format β€” fill the plug-in points
                               (canon, standing constraints) to mint an
                               officer for any project
engine/VERDICT_CONTRACT.md     the output specification β€” verdict semantics,
                               invariants, machine consumption
engine/FACET_BUNDLE.md         what an officer is made of: the core facets,
                               the three states (absent / null-with-reason /
                               present), and the x- namespace rule that lets
                               evolution add facets forever without collision
engine/THOT_MANIFEST.md        how N facets become ONE content-addressed
                               thing β€” thot: / CID / contentRoot from a single
                               canonicalisation, the two roots, and lineage
engine/FAICE_FORMAT.md         the faice/1 face-identity format, voaice's
                               sibling β€” twelve ordered ratios, and why an
                               unmeasured print is null rather than plausible
engine/facet_registry.json     the registry itself, machine-readable
engine/CARRIER_TEST.md         can this model carry an officer? form vs
                               substance, the probe set, and measured
                               carriers β€” Bonsai-8B 1-bit: REJECT for
                               review duty; gpt-oss-120b control:
                               APPROVE_WITH_CONDITIONS
engine/carriers/<date>/        the SavanteUI transcripts behind each
                               carrier verdict, graded by reading
.claude/skills/sagi/SKILL.md   the /sagi skill β€” the engine invocation,
                               officer-agnostic (published)
sAGI.md                        the definition β€” three laws, doctrine, an
                               early text of Savante's skill (canonical copy
                               in cryptoAGI/savante; this mirror adapts paths
                               and may lead while a correction lands in both)
officers/savante/SKILL.md      the reference officer's own /sagi skill β€”
                               Savante-specific, a worked example of the
                               engine filled in (canon: cryptoAGI/savante)
llm.txt                        orientation for machines β€” read first
LICENSE                        MIT

Copy the engine skill (.claude/skills/sagi/) to build your own officer. officers/savante/SKILL.md is published to show what a filled-in officer looks like; it loads .claude/agents/savante.md, which lives in cryptoAGI/savante, and that repository wins if the two copies ever differ. The copy here is the canon's at 619c71a (v0.0.5); its paths (gfx/, PROOF.sha256, bind/) are relative to the canon. Check it with sha256sum officers/savante/SKILL.md, which must print b8b9b62db9bac8286171a5766fe23f0512987872d2cefc9598036b7609daff30, the value in the canon's PROOF.sha256 and ledger.

The bundle documents are specifications, not an implementation. The reference implementation is Savante's bind/savante_bind.py (writes the manifest) and bind/savante_verify.py (recomputes every digest from raw bytes, offline). As of 2026-09-17 that implementation applies Β§3a: at savante 619c71a, savante_verify.py REJECTs (exit 2) a manifest whose algorithms block is missing and one whose facet_digest reads blake3, refusing before it computes any digest, and savante.thot.json declares exactly the ten keys it implements. Where spec and implementation disagree, the spec still wins.

Build an officer

  1. Copy engine/CHARTER_TEMPLATE.md to <your-repo>/.claude/agents/<name>.md and fill the plug-in points: the officer's name, the canon it measures against (your architecture docs, ADRs, strategy files β€” or none: with no canon it measures claims against the code itself), and the standing constraints it enforces. Do not widen the tool allowlist β€” read-only is the engine's integrity guarantee.
  2. Copy .claude/skills/sagi/ to <your-repo>/.claude/skills/sagi/.
  3. Start a Claude Code session and invoke: /sagi review <claim> β€” or use the skill's fallback immediately (it instructs a general-purpose agent to adopt the charter).

Service modes β€” interactive review, headless CI merge gate, Agent SDK endpoint, scheduled audit β€” are identical to the reference implementation; see Savante's usage.md and technical.md.

Engine properties

  • Model-portable. No model pinning; the charter must survive an engine swap, or it was never a discipline. The same files run on frontier or local models β€” the form of the contract ports anywhere; whether a given carrier keeps the substance is measured, not assumed (CARRIER_TEST.md).
  • Plain-text everything. The entire engine is plain text β€” markdown plus one machine-readable JSON registry: diffable, auditable, forkable, installable with cp.
  • Machine-consumable. The verdict contract is grep-stable; officers can gate merges, block deploys, and page humans (DEFER) without a parser.
  • Duplicable. Copying the files replicates the service β€” every repo, every org, every pipeline gets a chairman that has actually read the code.

Lineage

Extracted from the mindX GΓΆdel-machine project (Project Chimaiera) by Professor Codephreak. The engine's doctrine β€” science requires objective truth β€” is instrumented on-chain in the PYTHAI constellation: SCIENΒ·TIFIC (measured accuracy, chronos.oracle time-truth), LUV (attention by proof of gesture), CP2048-QR (security claims by evidence). Value creates price, never the reverse.

Reference officer: cryptoAGI/savante β€” which also sits as the 9th seat, sAGI, of the mindX DAIO boardroom: its verdict is cast as a vote, and a DEFER holds the session for the operator (the contract's "block and page a human"), carried by a model the carrier test graded, not by one it rejected. First verdict on record: 2026-07-26, parsec-wallet production readiness β†’ APPROVE_WITH_CONDITIONS.