REI v3 now runs the native-modality architecture described in
Docs/architecture/REI_NATIVE_COMPOSITION_ARCHITECTURE.md. Racio, Emocio, and
Instinkt form independent native conclusions before communication or character
governance. Racio remains the only conscious decision type; Ego is an
append-only composition across cycles, not a fourth agent.
See CURRENT.md for the exact active boundary and
Docs/evals/rei_native_architecture_acceptance_2026-07-13.md for the final B14
verification record.
app/backend/rei/app/gui/scripts/run_rei_native_cycle.pyscripts/run_rei_native_profile_matrix.pyknowledge/canon_v2/tests/rei/ plus archive/cutover guards in tests/Run one reproducible cycle:
python scripts/run_rei_native_cycle.py `
--runs-root output/runs `
--ego-traces-root output/ego_traces
Run the canonical profile matrix:
python scripts/run_rei_native_profile_matrix.py `
--output output/reports/rei_native_profile_matrix.json
Run the active workbench:
python -m uvicorn app.gui.server:app --host 127.0.0.1 --port 8765
The default cycle and GUI paths are deterministic and do not silently call a model, renderer, image generator, or training pipeline. Model-backed adapters must declare exact model provenance, revision, seed, and call records.
Use a repository-local temporary root on this Windows machine because the global pytest temp directory may not be readable:
$base = Join-Path (Resolve-Path tmp) ("pytest-" + [guid]::NewGuid())
app\backend\.venv\Scripts\python.exe -m pytest -q --basetemp $base
The cutover also has permanent guards for archive hashes, archive isolation, removed legacy entrypoints, and absence of transitional package references.
The former textual three-LLM EgoResultant runtime, its 156-case Granite
profile matrix, prompt/dataset GUI, dataset tooling, and duplicate reference
tests are preserved at:
archive/rei_v3_text_llm_baseline_2026-07-13/
That snapshot is sourced from commit
05996b2b4a34cf6dd654e032d5dbc26bb5373ef0, carries a complete SHA-256
inventory, and is not imported by active code. Historical datasets and eval
reports may remain in the repository as research records; they are not active
runtime or automatic training inputs.
This software is a research simulator. It does not diagnose people, determine the character of real persons, or treat its outputs as objective psychological truth. The current upgrade does not train QLoRA or diffusion adapters and does not turn Ego or Življenje into an autonomous agent.
138 commits
Python
96.3%
JavaScript
1.3%
TypeScript
1.3%
REI v3 now runs the native-modality architecture described in
Docs/architecture/REI_NATIVE_COMPOSITION_ARCHITECTURE.md. Racio, Emocio, and
Instinkt form independent native conclusions before communication or character
governance. Racio remains the only conscious decision type; Ego is an
append-only composition across cycles, not a fourth agent.
See CURRENT.md for the exact active boundary and
Docs/evals/rei_native_architecture_acceptance_2026-07-13.md for the final B14
verification record.
app/backend/rei/app/gui/scripts/run_rei_native_cycle.pyscripts/run_rei_native_profile_matrix.pyknowledge/canon_v2/tests/rei/ plus archive/cutover guards in tests/Run one reproducible cycle:
python scripts/run_rei_native_cycle.py `
--runs-root output/runs `
--ego-traces-root output/ego_traces
Run the canonical profile matrix:
python scripts/run_rei_native_profile_matrix.py `
--output output/reports/rei_native_profile_matrix.json
Run the active workbench:
python -m uvicorn app.gui.server:app --host 127.0.0.1 --port 8765
The default cycle and GUI paths are deterministic and do not silently call a model, renderer, image generator, or training pipeline. Model-backed adapters must declare exact model provenance, revision, seed, and call records.
Use a repository-local temporary root on this Windows machine because the global pytest temp directory may not be readable:
$base = Join-Path (Resolve-Path tmp) ("pytest-" + [guid]::NewGuid())
app\backend\.venv\Scripts\python.exe -m pytest -q --basetemp $base
The cutover also has permanent guards for archive hashes, archive isolation, removed legacy entrypoints, and absence of transitional package references.
The former textual three-LLM EgoResultant runtime, its 156-case Granite
profile matrix, prompt/dataset GUI, dataset tooling, and duplicate reference
tests are preserved at:
archive/rei_v3_text_llm_baseline_2026-07-13/
That snapshot is sourced from commit
05996b2b4a34cf6dd654e032d5dbc26bb5373ef0, carries a complete SHA-256
inventory, and is not imported by active code. Historical datasets and eval
reports may remain in the repository as research records; they are not active
runtime or automatic training inputs.
This software is a research simulator. It does not diagnose people, determine the character of real persons, or treat its outputs as objective psychological truth. The current upgrade does not train QLoRA or diffusion adapters and does not turn Ego or Življenje into an autonomous agent.
138 commits
Python
96.3%
JavaScript
1.3%
TypeScript
1.3%