Ember

Ember is a clean-genesis sovereign foundation intelligence, a local foundation-model creation primitive, a continuously improving organism, its ember-cli body, and a general local AI laboratory. The target is an owned 30-35B sparse unified text-image-audio decoder with native reasoning and structured tool use, created and operated on one 24 GiB GPU.

For a fresh context, start with docs/START-HERE.md.

Install the measured Python environment

Python dependency authority: manifests/python-environment-v1.json

From the repository root, install the exact direct Python environment measured for Ember:

python tools/ember-restart-3b/python_environment.py install

The manifest records CPython, CUDA-wheel index, exact direct-package versions, the pinned Transformers source commit, import-to-distribution mappings, and guarded optional integrations. The default install excludes the exact ARC-AGI feature pins because their upstream wheel declares Python 3.12 or newer while the measured Ember interpreter is Python 3.10. Validate the repository contract and the current interpreter without installing anything:

python tools/ember-restart-3b/python_environment.py verify --check-installed

Rust remains authoritative in runtime/ember-lab/Cargo.toml, and the ember-cli TypeScript package remains authoritative in tools/ember-cli/src/package.json. This measured contract is not yet a claim that a fresh clean environment was independently installed.

Current truth

Board receipt: ember-totality-20260711T102112Z (ts 20260711T102112Z).

Counts: 11-GREEN / 24-RED / 2-UNEVALUABLE / 3-AUDIT-OK / 0-AUDIT-INCIDENT / 0-AUDIT-PENDING-EPOCH (total 40 rows, 29.7% of state-conditions GREEN).

Legend: GREEN = a fresh receipt satisfies the condition's CHK; RED = CHK unmet or a satisfying artifact is absent; UNEVALUABLE = the probe genuinely cannot look (counts as RED for completion math); AUDIT-OK/AUDIT-INCIDENT/AUDIT-PENDING-EPOCH = the three standing process-invariant rows (cadence-audit results, never a completion conjunct).

What changed 2026-07-06. GOAL.md was re-founded from scratch via a maintainer-operator constitutional interview; the totality board was honestly reset under stricter probes rather than carried forward from the pre-refound era. The counts above will not resemble the prior 29-GREEN board and are not a regression — that record is archive only (docs/goal-archive.md), describing a goal document and a probe set that no longer exist. Read the counts above as the honest starting line under the current constitution, not as a drop from a prior score.

No current checkpoint is an admissible Ember model.

Current checkpoint subject: bf20f05018991eb611b0623edd50a00ec30639da2f8ccae646f6962f152a2a2b.

• Disposition: CHECKPOINT_CANDIDATE_NOT_ADMITTED; capability credit: none; sufficient pretraining proven: false.
• Config: 559959894dc603f9fbccbb091b3a084fef23b58d29add05efd14799a9a298ae0; tokenizer: 2c557e7ffe64706112ea947d056be503005d90b16f64c57ec354267c7e9e9c97; optimizer state (custody-only, public bytes absent): ee864fc9779e7f0d943a08836253726a41f86679360f477a35d5348486f3162b.
• Cursor: step 2, record 2, token offset 2048, tokens seen 2048; active route: shared.
• Parameters: 3839161856 unique, 3839161856 trainable, 3839161856 served, 1020589568 active, 1020589568 episode-trainable.
• Historical predecessor: af954c22fb8fb7a0dc640bfd2e0ab97e8e4cde989607372fc45c3db7878699a4 at 1024 tokens (historical_step1_predecessor).

• All four specialist banks in that checkpoint remain at their independently verified random genesis. Image, audio, reasoning, and tool training exposure is therefore still zero for this checkpoint; no modality or capability admission is claimed.
• The approximately 2.195B owned checkpoint remains radically undertrained, capability-unproven, below the 3B birth floor, and historical-only.
• Earlier approximately 0.4B, 0.7B, and 1.2B references are stages or accounting views of historical lineages, not separate current Ember models.
• The closer unified multimodal implementation is a disconnected research prototype; it is not the trained cbase checkpoint.
• The checked-in ember-cli source now refuses ordinary model-backed launch without an OWNED_ADMITTED identity. Borrowed Qwen remains available only through the explicit REFERENCE_ONLY comparison seat and must never be displayed or credited as Ember. The installed cockpit deployment has not yet been clean-checkout reverified against the new owned path.
• No complete neural verified-experience self-improvement cycle or field-level breakthrough is currently proven.

The current active work is the executable EMBER-02 rung: specialist semantic training, frozen checkpoint-bound evaluation, sufficient-pretraining evidence, and clean-checkout owned serving/ember-cli integration. The one-step 3B checkpoint is an instrument for that work, never the destination.

Non-negotiable model floor

The first model and every 3B, 7B, 15B, and >27B rung are sufficiently trained native text-image-audio models with reasoning and structured tool use. The headline hypothesis is Verified Expert Accretion. Borrowed learned or evaluative signals never enter the lineage. Negative evidence preserves research families and later synergy tests.

Verify the authority tree

Run:

python scripts/verify_authority_conservation.py --root .
python -m pytest -q scripts/tests/test_authority_conservation.py

A local landing or dispatch check also supplies the operator-owned durable goal selection.

Read next

• INVARIANT.md - unamendable identity
• GOAL.md - complete project and machine contract
• STATE.md - compatibility pointer only; exact artifact identities and maturity live in CONTINUITY.md
• docs/goal-clear-protocol.md - proof required for completion
• docs/ember-authority-matrix.md - D-001 through D-062 conservation matrix
• docs/ember-completeness.md - preserved legacy M/C manifest and diagnostic tally input
• CONTINUITY.md - exact current artifact identities, maturity, execution boundary, and resume order
• GOVERNANCE.md - authority, provenance, and landing rules