abductio-core

ABDUCTIO MVP core engine: deterministic scheduling, ledger updates, and a strict application boundary. You supply evaluator/decomposer/audit ports; the engine handles orchestration and invariant enforcement.

What ABDUCTIO does: Evaluates a MECE (mutually exclusive, collectively exhaustive) set of hypotheses about a question/phenomenon/case (the "scope") under credit constraints, with permutation-invariant scheduling and mandatory template-based decomposition to prevent asymmetric scrutiny.

Project status: core engine library. The HTTP API lives in abductio-service. See architecture.md and docs/white_paper.org for the spec.

Install

Requires Python 3.11+.

From PyPI:

pip install abductio-core

Optional OpenAI adapter dependency (used by abductio_core.adapters.openai_llm):

pip install abductio-core[e2e]

Local path install (for use inside another codebase):

pip install /path/to/abductio-core

Editable install for local development:

pip install -e /path/to/abductio-core

If you host this repo, you can also install directly from VCS:

pip install git+ssh://<your-host>/<org>/abductio-core.git

Quickstart

from dataclasses import dataclass, field
from typing import Any, Dict, List

from abductio_core import RootSpec, SessionConfig, SessionRequest, run_session
from abductio_core.application.ports import RunSessionDeps
from abductio_core.domain.audit import AuditEvent

@dataclass
class MemAudit:
    events: List[AuditEvent] = field(default_factory=list)
    def append(self, event: AuditEvent) -> None:
        self.events.append(event)

@dataclass
class NoChildrenDecomposer:
    def decompose(self, target_id: str) -> Dict[str, Any]:
        if ":" in target_id:
            return {"ok": True, "type": "AND", "coupling": 0.80, "children": []}
        return {"ok": True, "feasibility_statement": f"{target_id} feasible"}

@dataclass
class SimpleEvaluator:
    def evaluate(self, node_key: str) -> Dict[str, Any]:
        return {"p": 0.8, "A": 1, "B": 1, "C": 1, "D": 1, "evidence_refs": "ref1"}

# Define the question/phenomenon you're evaluating (the "scope")
# and the competing hypotheses
request = SessionRequest(
    scope="Germany's GDP trajectory in 2026",  # What the hypotheses are about
    roots=[
        RootSpec("H_grow", "Economy will grow", "Not flat or shrink"),
        RootSpec("H_flat", "Economy will be flat", "Not grow or shrink"),
        RootSpec("H_shrink", "Economy will shrink", "Not grow or flat"),
    ],
    config=SessionConfig(
        tau=0.70,
        epsilon=0.05,
        gamma=0.20,
        alpha=0.40,
        beta=1.0,
        W=3.0,
        lambda_voi=0.10,
        world_mode="open",
    ),
    credits=10,
    required_slots=[{"slot_key": "feasibility", "role": "NEC"}],
)

result = run_session(
    request,
    RunSessionDeps(
        evaluator=SimpleEvaluator(),
        decomposer=NoChildrenDecomposer(),
        audit_sink=MemAudit(),
    ),
)

print(f"Ledger: {result.ledger}")
print(f"Stop reason: {result.stop_reason}")

Core Concepts

Scope vs Hypotheses

• Scope: The question, phenomenon, or case being evaluated (e.g., "Germany's 2026 economy")
• Hypotheses: The MECE set of alternative explanations/outcomes (e.g., {grow, flat, shrink, other})

The scope is the frame for the evaluation session. It's not itself a hypothesis - it's what the hypotheses are competing to explain or predict. This distinction is functional: it tells the engine what partition you're evaluating, enabling proper MECE accounting and H_other handling.

Permutation Invariance

Given identical inputs (scope, hypotheses, evidence, credit budget), ABDUCTIO produces identical outputs regardless of:

• Which hypothesis is listed first
• Which hypothesis the user is "focused on"
• The order operations are logged

This is achieved through canonical ordering (hash-based), frontier-only scheduling, and symmetric obligation templates.

Obligation Templates

Every hypothesis must be evaluated through the same template of required slots (default: feasibility, availability, fit, defeater_resistance). This prevents asymmetric scrutiny where one hypothesis gets loaded with requirements while rivals remain vague.

API surface

Public imports from abductio_core:

• RootSpec, SessionConfig, SessionRequest
• SessionResult, StopReason
• run_session, replay_session

Ports (implement in your app):

• EvaluatorPort, DecomposerPort, AuditSinkPort
• RunSessionDeps

Development

pip install -e ".[dev]"
pytest

Environment variables:

• Copy .env.example to .env and set OPENAI_API_KEY to run E2E tests that hit the OpenAI API.

Docs

Narrative/source docs live in docs/ as Org/HTML/PDF assets. These are treated as source material, not a hosted docs site.

Versioning

Version is stored in pyproject.toml and tagged as vX.Y.Z on release. The scripts/release.sh helper updates the version, commits, and tags.

Release (PyPI)

Recommended: automated publish on tags via GitHub Actions + PyPI trusted publishing.

One-time setup (PyPI):

1. Go to https://pypi.org/manage/account/publishing/
2. Add a trusted publisher for Promise-Foundation/abductio-core
3. Select workflow: .github/workflows/publish.yml

Release flow:

1. Update version in pyproject.toml
2. Ensure tests pass: pytest
3. Commit changes
4. Tag + push: git tag vX.Y.Z && git push --tags

The GitHub Action publishes the build to PyPI on tag push.

Helper script (recommended):

scripts/release.sh X.Y.Z
git push
git push --tags

Manual fallback:

python -m pip install --upgrade build twine
python -m build
twine upload dist/*

Token-based auth:

export TWINE_USERNAME=__token__
export TWINE_PASSWORD=pypi-<your-token>

License

MIT. See LICENSE.