Skip to content

[Workflow guidance] Auditable multi-agent statistical research for Mega-Sena indicators and covering designs #714

Open

Description

@FABIO8495

Request

We are developing “Mega-Sena · Histórico e quadrantes”, a Portuguese-language dashboard for Brazil's Mega-Sena lottery. We would appreciate guidance on implementing an auditable research workflow in ChatDev 2.0 / DevAll, and whether an existing workflow template is a suitable starting point.

Our primary research objective is to determine whether any indicator can improve future number-group coverage relative to a fair baseline. We do not assume that an edge exists. Under independent uniform draws, historical selection rules cannot create a predictive advantage; a negative result must remain an acceptable outcome.

Project and current position

  • Each draw consists of six distinct numbers from 1 through 60.
  • We use 12 disjoint groups of five numbers: Q1 = 01–05, ..., Q12 = 56–60.
  • The current experimental Z-core selects nine groups, corresponding to 45 numbers. It ranks standardized delays since each group's last appearance, using historical completed gaps.
  • The dashboard supports historical results, group analysis, indicator coverage and complementary number sets.
  • Frequency, delay, momentum and conditional repetition rules have been investigated. We do not currently have convincing evidence of a sustained predictive edge.
  • We have identified IOI-LottoLab as a potential source of research-workflow ideas, but have not yet executed or validated its models on Mega-Sena.

We also maintain a separate combinatorial optimization track, unrelated to predicting the next draw. In local experiments using Covering-PDO plus independent exhaustive verification:

  • 9 numbers / 12 six-number tickets covered all 126 four-number subsets.
  • 15 numbers / 129 tickets covered all 1,365 four-number subsets, reducing our project's previous 145-ticket reference. Global minimality was not proved.
  • A bounded five-minute search for 30 numbers / 2,604 tickets covered 27,345 of 27,405 four-number subsets. The remaining 60 counterexamples mean it does not have full coverage.

These are locally executed results reported for context, not independently reviewed results or claims of financial return. Full four-subset coverage provides a conditional guarantee only: if at least four drawn numbers are in the selected set, at least one ticket matches at least four numbers.

Proposed agent workflow

  1. Data auditor: validate draw identifiers, chronological order, dates, six distinct valid numbers, missing records and dataset hashes.
  2. Statistical researcher: specify one testable hypothesis, fixed formula, tie-breaking rule, primary metric and stopping rule before evaluation.
  3. Implementation agent: implement a past-only prediction function and reproducible chronological evaluation.
  4. Adversarial reviewer: inspect leakage, overfitting, multiple comparisons, baseline fairness and counterexamples.
  5. Independent validator: execute deterministic checks separately from the implementation, and verify numerical claims against saved outputs.
  6. Report/UI agent: publish the results and limitations only after the validation gate passes.

Required safeguards and deliverables

  • Separate development/tuning data from evaluation; acknowledge that already explored historical data are not a pristine holdout.
  • Maintain the same nine-group selection size across compared methods.
  • Measure covered drawn numbers (0–6), including 5+ and 6/6 coverage; predeclare one primary metric.
  • Use the appropriate combinatorial/random baseline, paired comparisons where applicable, uncertainty intervals and correction for multiple tested hypotheses.
  • Do not equate increased mean coverage with guaranteed improvement in actual-ticket prizes.
  • Freeze prospective predictions, model versions and timestamps before each draw.
  • Save source/version, dataset hash, seed, parameters, actual execution logs and per-draw outputs.
  • Permit “no advantage demonstrated”; agent consensus must never substitute for executed evidence.
  • Keep experimental indicators separate from production; no automatic betting or deployment.
  • Deliver a runnable workflow, minimal tool interfaces, reproducible evaluation report and explicit pass/fail gates.

Specific questions for ChatDev maintainers/community

  1. Which existing YAML workflow or combination of subgraphs is the best starting point for this researcher → implementer → adversarial reviewer → independent execution/validation pipeline?
  2. How would you enforce that an agent cannot report a numerical result without attaching a successful tool execution and its output artifact?
  3. What is the recommended way to isolate a final evaluation dataset from agents that generate or tune hypotheses?
  4. How should independent Python validators and external optimization programs be integrated with time limits, bounded retries and explicit failure states?
  5. Is there an example of immutable pre-run specifications and prediction records tied to workflow/model versions?
  6. Would this fit as a reusable scientific-validation example? If a different discussion channel is preferable, please point us to it.

Our immediate request is architectural guidance or a minimal reference workflow, not a promise of winning numbers or a request for the maintainers to build the entire application.

References:

Prepared with AI assistance at the project owner's request.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions