Skip to content

Project notice

Strategy Succession — plan, adapt and repair

Prepare a release plan, change its requirements, then repair that same project while preserving its committed planning prefix. The connected interface runs reviewed fixed methods on your explicitly configured local executor. It retains checked results, reusable candidate methods, costs and recovery evidence in the existing Alpha Journal.

This is a bounded planning product. It does not execute physical jobs, run arbitrary repository code, operate a desktop, fund a job or activate the held Harness/Fleet integrations. The free flagship browser experience remains local and requires no account, wallet, installation or API key. A static tab is not an unattended executor.

Install the supported connected subset

The companion is agi-alpha-strategy-succession-engineering 0.1.0rc1. Core release 1.44 distributes it, while its qualified runtime remains the separately installed public Alpha 1.42.0 wheel. Other runtime versions and product combinations remain unqualified. Keep existing accepted Alpha, Jobs and Manager installations intact.

Use Ubuntu 24.04 x86_64/glibc 2.39, CPython 3.12.14 and Node 24.22.1. The installation profile pins all 26 core dependency versions, including SciPy 1.17.0, NumPy 2.4.6 and patched Starlette 1.7.0. The launcher checks parent and actual child environments before opening operator state. It refuses unknown combinations.

Download the companion wheel, its operator guide, compatibility JSON and SHA256SUMS from the 1.44 release. Before publication this is a proposed destination, not an assertion that the assets exist. Download the Alpha 1.42 wheel from its preserved release. Verify each asset against the checksum list belonging to that release. The exact Alpha runtime wheel is 5,498,450 bytes, SHA256 fff8c6c8a8dbeeebf1f91cc8e9a3fae17113baac93b6ed0d6096ecb59d072554.

Create a separate virtual environment. From a neutral directory outside any Alpha checkout:

python3.12 -m venv strategy-env
strategy-env/bin/python -m pip install ./alpha_factory_v1-1.42.0-py3-none-any.whl ./agi_alpha_strategy_succession_engineering-0.1.0rc1-py3-none-any.whl
strategy-env/bin/python -m pip check
strategy-env/bin/strategy-succession-engineering --home /absolute/path/to/existing-approved-home --native

The executor binds 127.0.0.1:8768 and serves its interface there. It requires an existing approved operator home and does not create new persistent credentials. The operator supplies that home's existing access token. Enter it in the password field; the field clears and the token stays only in that tab's memory. Credentials are excluded from exports. No model API key, account or wallet is required for these fixed algorithms.

One explicit start

  1. Start prepares the exact input, executor, source, expiry and separate budgets. It does not begin execution.
  2. Review the immutable input and approve that exact finite scope. Editing the future-input field does not change it.
  3. Approve and start work runs planning, changed planning, construction comparisons and repair without further phase prompts.
  4. Review the three checked results and their committed receipts. Native specialists remain available at every phase.
  5. Keep downloads inspectable candidate methods, with executable AST, dependencies, provenance, limits and cost history. These records contain no inherited proof or production authority. Improve is available only for an admitted separate proposal.
  6. Export downloads the signed local evidence package. Its signature establishes local attribution, not truth or authority.

The default request caps are 500,000 declared service units, 3,000,000 improvement units, six native invocation attempts and twenty solve-allowance seconds. Every native attempt charges its two-second allowance before starting, including startup or import failures; actual physical wall/CPU measurements stay separate. The request expires within 120 seconds, approval within 600 seconds, with one active request and finite installation-wide totals. There is no default recurrence daemon.

What the comparisons mean

The authored example first yields makespans 15 and 17. Reducing capacity during slots 12–14 makes the actual preceding plan invalid at slots 12/13. Repair preserves the starts already committed at the supplied planning time and produces a valid makespan-19 schedule. These are planning classifications; physical job execution is not observed.

Parent construction and repair usefulness are separate tests. Inherited, rebuilt and equivalent parents have matched total and child allowances. Retrieval, reconstruction, failed proposals, checks, integration and retention are charged. Inherited construction can cost less than exact reconstruction while ordinary equivalent reuse is cheaper still.

On the binding disruption, the constructed repair is valid with movement 9; the existing matching native repair achieves movement 7. The system selects the stronger native method and retains the candidate's LOSS / HOLD. Earlier no-change and tied cases remain historical evidence. These observations establish neither capability growth nor economic superiority.

Correctness does not imply economic admission. Human effort, model-assisted authorship, hardware/currency cost, Python HTTP/journal/cryptographic overhead and whole QA/development cost remain partly unknown. Model inference calls in the fixed executor are zero. Forecast maintenance is an assumption. Work units, CPU seconds, monetary value and revenue are different quantities; no synthetic saving is profit.

Pause, stop and recover

Pause fences the next bounded checkpoint. Stop is permanent for that installed operating envelope. Closing a tab does not cancel an already approved request; reconnect and select the retained mission. An approved but unstarted plan requires fresh consent to resume. Running or uncertain work never retries automatically.

After a crash, the previous process may be gone while its journal still records a lease. The interface says it is awaiting evidence, preserves checked partials and holds the full unknown reservation. Inspect effects and export the record before reconciliation. An unexpired lease cannot be cleared. Reconciliation retains unknown ceilings without a refund and does not resume work. The operator procedure in the packaged guide covers source changes, expiry and renewal.

Actual supervisor-death checks include a bounded native child and an unchanged server restart. SIGKILL cannot run Python cleanup instantly; stdin EOF fences the next phase after any already bounded call. Descendants do not inherit proof, grants or a serving release. Use the last accepted installation for rollback; do not move tags, rewrite old evidence or reset a stopped journal to recover authority.

Release evidence and boundaries

The compatibility matrix and capability register distinguish implemented, executed, independently checked, published, installed and operationally authorized features. Source review, fresh hosted Chromium, exact public package installation and the existing resulting-main release gates are required. Earlier failed audits, missing fixtures and the corrected virtualenv invocation defect remain preserved by source epoch.

The original research edition under studies/agi-alpha-flywheel-strategy-succession-v1 retains its immutable protocol, negative strong-comparison findings and all 239 source/evidence bodies. The engineering example is visible common-custody qualification, not a new blinded scientific study. Preferred Edition 02, original Harness/Fleet restrictions, historical URLs and separate unfunded AGIALPHA/USDC routes remain unchanged. This subset is not a complete four-product installation.