See docs/DISCLAIMER_SNIPPET.md

Solving α-AGI Governance · Governance Workbench

Launch Demo

Review the rules before an autonomous enterprise acts. Inspect incentive compatibility, validator independence, quadratic ballots, aggregate risk and upgrade controls. Leave with a reproducible review and nine measurable, unsubmitted $AGIALPHA verification jobs.

Open the workbench · Open the notebook · Original research archive

Current runnable path — 1.14.0

Mode: Offline governance review. Evaluates nine proposal gates and exports reproducible evidence and Ascension job specifications.

Prerequisites: Python 3.11–3.13. The workbench uses only the standard library; no wallet, API key, provider or database is needed.

From a source checkout:

python -m alpha_factory_v1.demos check solving_agi_governance
python -m alpha_factory_v1.demos run solving_agi_governance --output-dir governance-runs

Expected result: REVIEW_REQUIRED for the constructed accountable-upgrade case; five evidence files in a content-addressed directory.

Scope: The calculation does not authenticate identities or votes, calibrate risk estimates, prove universal convergence, approve an upgrade or move tokens. All jobs remain unsubmitted.

A three-minute first run

1. Open the browser workbench and select Accountable upgrade.
2. Inspect the assumptions and roster, then Evaluate all nine gates. All nine modeled checks pass; independent review remains required.
3. Try A majority with a shared controller. The identity gate blocks it even though the vote totals are unchanged.
4. Download the ZIP. It contains the complete inputs, outputs, job specifications, review brief and SHA-256 checksums.
5. Import dossier.json to recompute every decision, or verify it with Python. Changing any input invalidates the current result and disables exports.

Nothing is uploaded. Browser draft storage is explicit: Save draft, Restore draft, and Clear saved draft affect this browser only. A ZIP is the portable backup. After a complete first load, the service worker supports offline recalculation. The original replay's optional OpenAI mode is a separate, explicit network action.

What the gates establish

flowchart TD
    Proposal["Proposal and supplied evidence"] --> Incentives["Conditional incentive check"]
    Proposal --> Voice["Identity, credits and ballot"]
    Proposal --> Safety["Risk budget and release controls"]
    Incentives --> Gate{"All modeled gates pass?"}
    Voice --> Gate
    Safety --> Gate
    Gate -->|No| Revise["Blocked conditions and review jobs"]
    Revise --> Proposal
    Gate -->|Yes| Review["Independent validator review required"]
    Review --> Evidence["Authenticated evidence and policy approval"]
    Evidence --> Protocol["Separate Ascension execution boundary"]

Gate	Exact modeled check	What still needs external evidence
Identity	At least one eligible, staked validator; no repeated controlling principal	ENS ownership, complete roster and actual independence
Credits	Each eligible ballot costs votes², within its credits	Legitimate credit issuance, authenticated ballot signatures
Quorum	Valid nonzero ballots / eligible validators meets the threshold	A complete, frozen electorate snapshot
Mandate	Strict positive majority and declared support percentage	Proposal-bound signed vote records
Incentives	Cooperative reward ≥ normalized one-shot deviation return	Payoffs, monitoring, enforceable slashing and credible punishment
Risk	min(1, N × p) ≤ total risk budget	Calibrated upper bounds for the actual action envelope
Timelock	Explicit scenario time ≥ queue time + delay; minimum seven days	Authoritative chain timestamps and governance delay
Policy	Proposed and expected policy SHA-256 commitments match	Exact policy bytes and independent authorization
Pause	Supplied emergency pause is inactive	Authoritative stop state; tested rollback and recovery

A PASS / MODEL is conditional arithmetic. A proposal becomes REVIEW_REQUIRED only when all nine gates pass; otherwise it is BLOCKED. There is deliberately no APPROVED or EXECUTED result. A correctly calculated blocked case is a successful CLI run, so the process exits 0; invalid input or an altered dossier exits 2.

Run and verify locally

python -m alpha_factory_v1.demos.solving_agi_governance --list
python -m alpha_factory_v1.demos.solving_agi_governance --case scale-risk --output governance-runs
python -m alpha_factory_v1.demos.solving_agi_governance --input scenario.json --output governance-runs
python -m alpha_factory_v1.demos.solving_agi_governance --verify governance-runs/<sha256>/dossier.json

Replace <sha256> with the directory printed by the run. The governance-workbench console command is equivalent after installing the project. --json prints the complete dossier. Verification is read-only and never fetches a source URL.

Constructed case	Expected result
accountable-upgrade	Nine modeled passes; independent review required
captured-ballot	Identity blocked: two named validators share a controller
weak-deterrence	Incentives blocked: the one-shot deviation still pays
scale-risk	Risk blocked: 10¹² actions at 10⁻⁹ per action exceed the budget
paused-upgrade	Timelock, policy commitment and emergency-stop gates blocked

Every bundle contains:

• scenario.json: the normalized, complete input, including the supplied clock and source note.
• dossier.json: all gate decisions, exact arithmetic and a SHA-256 commitment.
• review-brief.md: readable findings, assumptions and next steps.
• jobs.json: nine goal ↔ success metric ↔ bounty specifications accepted by ascension-compile.
• SHA256SUMS: checksums for those four files.

Python and browser outputs must match byte for byte. The hash is an integrity commitment, not a signature or proof that supplied assumptions are true. A dossier import recomputes the entire report, including its jobs; changing the status and rehashing the file cannot turn a blocked case into an approved one.

The calculations

Conditional cooperation

Let R be the cooperative reward, T the one-shot temptation, P the subsequent punishment payoff, q the probability of detecting a unilateral deviation, s its slashable stake in the same utility units, and δ < 1 the future discount factor.

The cooperative path has value R / (1−δ). With probability q, a one-shot deviation is detected, incurs a single slash and triggers punishment. Otherwise, the deviator returns to the cooperative path. Its expected value is T−q·s + δ[qP+(1−q)R] / (1−δ). The workbench checks:

R ≥ (1−δ)(T−q·s) + δ[qP+(1−q)R]

This assumes infinite repeated play, stationary payoffs and public detection probability, risk-neutral agents, no false positives and credible grim-trigger punishment after detection. Punishment-path credibility is assumed, not proved. Private signals, false positives, collusion, forgiveness, finite horizons and competing strategies require additional models. Passing this inequality establishes neither a unique equilibrium nor antifragility.

At q = 0, a profitable deviation cannot be deterred by patience or an unenforceable stake. At q = 1, the condition reduces to R ≥ (1−δ)(T−s) + δP.

Basis-point inputs and integer payoffs make the reported margin exact. The browser uses BigInt where products exceed safe numeric precision. The chart changes δ while holding other assumptions fixed; its 100% endpoint is a limit, not an allowed input for the infinite-horizon value formula.

Risk at scale

For N actions, each bounded by failure probability p, the union bound is min(1, Np). No independence assumption is needed. The sum of expected failures is bounded by Np; this is not the probability of at least one failure.

At p = 10⁻⁹ and N = 10¹², the expected-count upper bound is 1,000 and the union bound is 1. For a total budget of 0.001, require p ≤ 10⁻¹⁵. Input probabilities use integer femto units (10⁻¹⁵); the admissible per-action bound is rounded down, and inputs that cannot be represented exactly are rejected rather than rounded.

The original paper's Table 4 is a separate normalized score. Its stated formula produces 0.393045, exceeding 0.3, rather than the printed 0.215. It cannot be substituted for a catastrophe probability. The existing Governance Observatory retains the table audit, Hawk–Dove dynamics, counterexamples and further analysis.

Fit within α-AGI Ascension

Vision component	Implemented handoff
Insight and Nova-Seeds	Review the proposal assumptions and their content commitment before a seed or policy change
MARK and Sovereign businesses	Inspect risk, incentives and control gates before capital or execution decisions
α-AGI Jobs	Export nine input-bound verification jobs with measurable acceptance criteria and $AGIALPHA bounties
Agents and validators	Compile jobs for the separate protocol; validate staked agent eligibility and independent validator returns there
Settlement	The Ascension protocol demonstrates local-EVM escrow, validator-gated settlement and a 1% payout burn

alpha-agent ascension-compile governance-runs/<sha256>/jobs.json --output fusion-plan.json

Compilation does not post a job, escrow a reward or authorize execution. Review the protocol guide before using its separate local-EVM workflow. Utility stake in the incentive model and AGIALPHA validator/job balances have different units; no exchange rate is assumed.

Original simulator and optional integrations

The original command remains available, with its valid seeded numerical behavior preserved:

python -m alpha_factory_v1.demos.solving_agi_governance.governance_sim --agents 100 --rounds 1000 --delta 0.8 --stake 2.5 --seed 42

Here --delta is a numerical update rate, not the repeated-game discount factor. Randomness initializes the population; subsequent mean-field updates are deterministic. There is no universal δ = 0.8 transition. Negative, non-finite and unbounded work requests now fail clearly. Runs are limited to 100 million agent updates.

--summary explicitly opts into the optional OpenAI Python client when credentials are configured. It is a language summary, not part of the decision calculation. The default run is entirely local. The legacy governance-bridge command uses the repository's runtime compatibility interface; its presence does not prove a live OpenAI Agents service. Without the interface, or with OPENAI_AGENTS_DISABLE=1, it runs locally and accepts --seed. ADK exposure remains opt-in.

Recovery and input limits

• Imports are UTF-8 JSON, at most 256 KB, with 1–64 validator records. Unknown keys, duplicate keys, invalid Unicode, non-finite numbers, excessive nesting and out-of-range values are rejected.
• The maintained model is finite and bounded. No imported text is executed or rendered as HTML, and no URL in an input is fetched.
• Scenario edits invalidate exports. Apply edited advanced JSON before evaluating or saving.
• Python output is content-addressed. An unchanged rerun reuses an exact bundle; altered, partial or symlinked runs are refused. Preserve them and select a new --output directory.
• If browser draft storage is unavailable, use JSON or ZIP exports. If assets are unavailable on a first offline visit, load once online or run Python from an existing checkout.

Research collection — preserved

Vincent Boucher's original materials remain available, including every original diagram:

• Manuscript PDF and TeX source.
• Presentation PDF and editable PowerPoint.
• Original README, unchanged and original notebook archive.
• Maintained notebook, including all five cases, replay checks and the original simulator.

The archive preserves historical assertions; it is not new evidence for the claimed unique equilibrium, six-million-round experiment, Coq certificates, Landauer-limit behavior or guaranteed antifragility. See the white-paper implementation guide for the existing audit and missing evidence. The workbench makes these review obligations explicit rather than treating aspirations as completed verification.

View README on GitHub