v0.2.0b2 — Verified learning loop beta

v0.2.0b2 — First verify. Then learn.

v0.2.0b2 is a beta release of Morpheus's source-grounded truth layer and
local experimental learning path. It does not claim stable-release maturity or
evidence from broader real-repository use.

Highlights

• Reviewed active state now uses
  morpheus-active-state-review-authority/1: only exact reviewed
  candidate-to-claim and evidence bindings can authorize active-state datasets.
  Ordinary compile and wake receipts remain integrity records and cannot
  authorize learning.
• The adapter benchmark uses morpheus-benchmark-categories/1 with canonical
  coverage for product identity, commands and CLI behavior, architecture,
  safety rules, team conventions, stale-claim correction, and unsupported
  claim refusal.
• morpheus learn team-loop and POST /learning/team-loop share the
  morpheus-team-learning/2 reviewed-input contract for PR comments, rejected
  agent claims, human corrections, accepted review candidates, check results,
  and stale-claim corrections. New correction candidates remain pending.

Safety Boundaries

No accepted source span means no training example.

No eval pass means no adapter activation.

No rollback means no production activation.

Adapters never activate automatically. Pending, rejected, stale, inferred-only,
secret-like, and ignored claims do not become positive training facts. Cloud integrations remain opt-in. Adapter output is not the source of truth.

Install Or Upgrade

Pin the beta package explicitly:

uvx --from 'morpheus-wake==0.2.0b2' morpheus wake . --private

Or use pipx:

pipx run --spec 'morpheus-wake==0.2.0b2' morpheus wake .

Replace an earlier beta pin with morpheus-wake==0.2.0b2. Unpinned installers
may choose a stable release instead of this beta.

Local Gates

Run the local quality and provenance gates before relying on a project state:

ruff check .
pytest tests/ -q
morpheus wake . --private
morpheus verify --all

These gates verify repository state and receipt integrity; they do not authorize
adapter activation without the separate reviewed dataset, paired eval, readiness,
and rollback requirements.

v0.2.0b1 — First verify. Then learn.

v0.2.0b1 — First verify. Then learn.

Beta prerelease notes for the source-grounded truth layer and autonomous local
learning lab.

Highlights

• morpheus check verifies agent text against local source-backed project
  state and returns verified, stale, incorrect, or unknown.
• MCP truth tools expose local read-only checks, active state, evidence lookup,
  and WAKE fetches.
• Semantic review and strict lab acceptance keep pending, rejected,
  inferred-only, stale, ignored, and secret-like claims out of training data.
• morpheus learn lab runs an autonomous local learning experiment:
  strict accept -> dataset -> optional MLX LoRA -> base vs adapter eval -> report.
• morpheus learn status shows the effective trainable dataset across
  standalone reviewed datasets and lab datasets.
• morpheus learn train . --dry-run can plan from the latest trainable lab
  dataset without requiring GPU or downloading models in tests.
• Legacy raw training is blocked by default behind an explicit unsafe
  confirmation flag.

Live ML Core Result

Command:

morpheus learn lab . --dogfood --backend mlx --eval-limit 0 --repeat 2

Result:

• Stability verdict: ML_CORE_PASS
• Runs: 2
• Strict accepted candidates per run: 69
• Training examples per run: 290
• Eval coverage per run: 148/148
• Adapter pass rate per run: 0.9932
• Base pass rate per run: 0.7973
• Adapter delta per run: +0.1959
• Critical failures: 0
• Regressions: 0
• Adapter activation: not automatic

Scope

This beta is a local source-grounded project-truth system, not a real-world
truth oracle and not a production adapter activation claim. The source spans,
receipts, morpheus check, eval, and rollback gate remain the source of truth.

Try

morpheus wake . --private
morpheus check --input tests/fixtures/check_correct_input.txt --local
morpheus learn lab . --dogfood --no-train

v0.1.1 — launch polish

• Public AGENTS.md path cleanup
• PyPI description link/image fixes
• Dependency metadata polish
• Project URLs added

v0.1.0 — WAKE.md for AI agents

v0.1.0 — WAKE.md for AI agents

First public alpha of Morpheus, the Agent State Compiler.

Highlights

• Generate WAKE.md from project state.
• One-command morpheus wake . flow.
• Public/private WAKE modes.
• Verifiable provenance receipts.
• Agent handoff.
• Visual terminal demo in the README.
• MCP/A2A-style local interop.
• UI launchpad.
• Cache-backed integrations.
• Experimental training is explicitly not the core launch path.

Install

Install with uvx:

uvx --from morpheus-wake morpheus wake .

For private projects:

uvx --from morpheus-wake morpheus wake . --private

Notes

The PyPI distribution is morpheus-wake; the CLI command remains morpheus.
Private workspaces can keep generated state under .morpheus/.