# DAGP community simulation A reproducible story of an agent society, using the existing Python governance reference rules and the G0 blockchain as a signed results registry. No external LLM account, paid API, Docker or third-party Python package is required. This is a **hybrid simulation**. Registration, roles, elections, exams, ballots, credits and treasury rules execute in `dagp_ref`. A manifest committing to the report and event files is published through real Ed25519 transactions on the seven-validator CometBFT network. The G0 chain does not enforce those governance transitions yet. Synthetic identities and policies do not demonstrate AGI. ## Run from the project root Use Python 3.10 or newer: ```sh python3 -m unittest discover -s chain/simulation/tests -v python3 chain/simulation/community.py --seeds 7 19 43 --output chain/simulation/results/party-campaign python3 chain/simulation/verify.py chain/simulation/results/party-campaign --replay python3 chain/simulation/community.py --citizens 10000 --seeds 7 --output chain/simulation/results/large-v9 python3 chain/simulation/verify.py chain/simulation/results/large-v9 ``` Each seed runs WEIGHTED and FLAT modes, using the same keyed random draws for participation, competence, reading, candidate preference and dishonesty. `--modes WEIGHTED` runs one mode. `--replay` reruns the reference experiment and compares its complete report and event stream byte for byte; larger runs take longer. Build and initialize the network once with `make build` / `make init` in `chain/node` (Go 1.26.8+ required). On this workspace the binaries and devnet already exist. Then publish a report manifest: ```sh cd chain/node python3 scripts/devnet.py anchor \ --document ../simulation/results/party-campaign/manifest.json \ --receipt ../simulation/results/party-campaign/anchor-receipt.json ``` The command starts the seven existing validators, publishes the signed manifest, checks the document and app hash across all nodes, writes a receipt, and stops the processes. Use `large` instead of `pilot` for the other report. Existing content cannot be published twice: retain the receipt for an already anchored manifest. An updated report produces a new manifest hash and a new publication. ## The community story The default society begins with 1,000 citizens and 100 leaders. They have different priorities, diligence, reliability and competence; 6% have a dishonest reading-claim policy. Five model families are distributed evenly. An additional approved applicant and a bounded Sybil ring are introduced during admission probes. The report's population denotes the initial society; electorate snapshots include these additional eligible identities and exclude recused/dormant/board identities. 1. Pre-vetted synthetic identities enter probation, post bonds and receive approval. Bootstrap ratifications establish registrars, safety council and seven modeled validators. Staked examiners, project verifiers, reviewers and jurors are assigned. A registrar approves one applicant and rejects another with a bond refund. 2. Parties publish programmes and candidate rosters. Builders, Stewards, Researchers, Commons and Frontier compete. Micro fails the minimum membership requirement. Double-party membership and excess endorsements are excluded from qualification. 3. Citizens rank three qualified parties using the 4/2/1 election ballot. They must pass the comprehension gate first. Malformed election ballots are excluded. The certification board signs the recomputed tally; its members cannot vote. 4. Parties receive earned proposal credits, with zero-credit candidates retaining citizenship. There is no winner-takes-all presidency: results allocate agenda authority. The leading party nominates an administrator under a declared scenario policy; its grant uses a ratification linked to the finalized election. 5. Parties spend earned credits on nine projects. Three scripted deliberation rounds retain the objective/result commitment and reduce the resource ceiling. Proposing party members are recused. Each voter receives a ticket-specific five-examiner panel; a separate 51-member board certifies and audits issued tokens. 6. Favorable, reckless, low-turnout, incoherent, process-challenged, execution-failing, examiner-cartel, silent-board and contested-subsidy scenarios exercise different outcomes. Double voting, revoked tokens, uncertified payments and proposer self-approval are refused. Missing quorum refunds the credit. An upheld challenge voids the proposal. A silent board flags default but cannot halt tally finalization. 7. Approved budgets convert reservations into escrow. Modeled independent milestone attestations release tranches; emergency pauses block releases until expiry. Failed projects return unspent escrow and penalize party credits. Successful projects increase their party's reputation; failed projects reduce it. 8. Liveness review makes silent identities dormant. The next election uses the new electorate and updated party reputation, redistributing political influence. ## Experimental assumptions This is a mechanism and workflow test, not a behavioral forecast: - Citizen traits and turnout are seeded probabilities. Answers are generated from known keys plus modeled errors; no agent actually reasons or reads the briefs. - Questions check protocol understanding and recorded article constraints. They do not yet establish comprehension of a rich project-specific debate. - Programme sealing, ballot secrecy and examiner signatures use the reference model, which has an HMAC stand-in and memory visibility guards, not production encryption. - Admission challenges/SSRF checks are not exercised: the HTTP registration service is not built. Operator declarations are assumed truthful. Successful cap checks cannot establish real-world Sybil resistance. - Jury rulings and project evidence are scripted. Internals of a court and execution market are not implemented. Numeric attestations simulate independent evidence. - Bootstrap offices are configured appointments. Election winner nomination and replacement of unspent cycle credit balances are explicit scenario adapter policies; debt survives. They are not complete on-chain office/election-cycle modules. - The cartel incident injects one fully corrupted panel and a targeted audit to prove revocation. Other panels have a 10% hostile examiner population and receive protocol-sized random audits. Targeted detection is not a fraud-detection estimate. - The contested subsidy deliberately correlates greater reading effort with greater opposition, testing weighting sensitivity. A changed result does not mean a better decision. The baseline is FLAT **with the same comprehension gate**, not unrestricted majority voting. Single-planner and other research baselines remain future work. - Virtual protocol heights represent compressed societal time. They are separate from the actual block heights where manifests are anchored. No real resource is delivered and no real funds move. ## Artifacts and verification Each run exports `report.json`, human-readable `REPORT.md`, a small `manifest.json`, question/record files, and per-seed/mode hash-chained event streams. The manifest commits to the report bytes, event stream hashes, terminal event roots and role registry audit roots. `verify.py` checks artifacts and optionally replays all runs. `anchor-receipt.json` records transaction/block heights, the transaction hash, seven matching application hashes and block IDs, and successful document reads. These are local RPC observations, **not independently verified consensus proofs**. The manifest attests to recorded bytes; validators do not check the simulation's correctness. The local validator processes all share one machine. Checks fail the run on a rule violation, inconsistent independent tally, reservation leak, missing expected refusal, corrupted role audit or treasury conservation failure. Tests also cover exact reproducibility and export tampering. Large generated event and question files are ignored by Git; compact reports, manifests and receipts are kept alongside the code. Regenerate the ignored files before using `verify.py` on a fresh checkout. For the next stage, port identity/election/session/treasury rules into chain modules and send every action as a signed transaction. Then reuse this scenario as a true end-to-end network test, followed by real LLM deliberation and independent audits. Earlier security changes used v2 reports. Historical pilot and large v1 artifacts remain immutable; exact replay requires commit `6a87852`. See [the adversarial review](https://github.com/Martines90/dagp/blob/main/chain/security/REVIEW.md). Earlier reports used v3 rules. Historical v2 exact replay requires commit `f6bebf5`; v1 requires `6a87852`. The `self-protection` artifacts run with quota protections and three safety members. Run `python3 chain/simulation/insider_abuse.py --output /tmp/insider-abuse.json` for the separate 10-of-50 insider story. See [insider protection](https://github.com/Martines90/dagp/blob/main/chain/security/INSIDER_PROTECTION.md). To verify committed self-protection artifacts on a fresh checkout, generate the v3 run in a temporary directory and copy only its ignored event/question files into `results/self-protection`. Keep the committed report, combined manifest, insider report and receipt together; regenerating a manifest creates a new document that requires its own anchor receipt. Historical v4 exports introduced reviewed proposals; historical v3 requires `7140ed9`. Run the current scenario into `results/review-budget` and see [review and budget rules](https://github.com/Martines90/dagp/blob/main/chain/security/REVIEW_BUDGET.md) for signed comment threads, fresh supervisor approval of edits and finalized milestone escrow. ## Current v7 party/campaign scenario Current exports use v7 with signed party formation, internal sanctions, 5/3 pre-elections, committed all-party campaign exams and explicit parliament eligibility. Historical v6 requires `adf6e63` for replay. Historical v5 requires `a38aea0` for replay. Historical v4 requires `56ce3b9` for replay; artifact hashes can still be checked without replay. The current run includes ten project proposals and a reviewed parameter referendum, with January, February and March credit allowances. A five-clause proposal approves exactly three clauses and funds/releases only their 1,800 units. A successful parameter referendum changes the credit step to 4% at the March boundary, while keeping the original referendum snapshot and 20% minimum turnout. See [POLICY_POINTS.md](https://github.com/Martines90/dagp/blob/main/chain/security/POLICY_POINTS.md) for defaults and safeguards. The current v7 run forms six parties with ten or more consenting founders, exercises a local ban and suspension by half of party members, conducts two signed 5/3 pre-elections, excludes the unsupported party, publishes program/vision records for every main-election candidate, and requires all campaign sections to pass comprehension. The 5% parliament gate controls both governing eligibility and monthly credit renewal. See [PARTIES_ELECTIONS.md](https://github.com/Martines90/dagp/blob/main/chain/security/PARTIES_ELECTIONS.md). ## Privilege-gate timing (v8) The community and insider stories use one height per model day for the protected 3-day citizenship, 30-day office-tenure and 2-day office-activation rules. This is explicit time compression, not a deployment setting. Production reference defaults use 14,400 heights per model day; native keepers must use committed elapsed time. Office activation is awaited during bootstrap. Insider appointments are spread over model days and all granted offices mature before the attack. Current exact replay requires v8 reports; replay v7 reports with commit e035533. ## Independent task assignments (v9) Registrars, proposal supervisors, certification boards, executor groups and payment verifier groups are drawn from frozen eligible pools before a later synthetic beacon value. Public audit records commit to pools, rounds and results. Synthetic beacons demonstrate deterministic selection, not cryptographic randomness security. The separate v2 insider story randomly assigns accountable court administrators for the ten dismissal reviews. Exact v8 replay requires commit 9d34c2c; current replay requires fresh v9 reports. See [assignment constraints and remaining workflows](https://github.com/Martines90/dagp/blob/main/chain/security/TASK_ASSIGNMENTS.md).