holdout 0.2.0

Preserve the dissenting reasoning behind contested decisions, as a durable record.

holdout

Most multi-model tools synthesize the disagreement into one answer, dissolving the dissent in the process. holdout deliberately preserves the dissenting position — because for decisions with no verifiable answer, the losing reasoning is exactly what's worth keeping.

---

---

Preserve the dissenting reasoning behind contested decisions, as a durable record.

holdout puts a question that has no verifiable answer to several independently prompted reasoners, each committing a written rationale and a YES/NO vote before seeing any peer. It keeps every rationale — including the losing one — as a durable, retrievable record, and on disagreement returns a crux (the specific, falsifiable disagreement to resolve) instead of forcing an answer.

The output is an artifact, not a decision. holdout does not make a recommendation, does not synthesize the positions into one answer, and makes no accuracy claim.

The project's internal codename and report aesthetic is MAGI (the three deliberating systems from Neon Genesis Evangelion: Melchior, Balthasar, Caspar). The installable package, CLI, and imports are holdout; "MAGI" in prose refers to the concept and the visual identity of the rendered reports.

Why

Most multi-model tools synthesize: they merge several model outputs into one "best" answer, and the dissent is dissolved in the process. That is the right move for questions that have an answer key. It is the wrong move for consequential decisions that don't — "should we rewrite this service," "is this trade-off acceptable" — where fusing model outputs manufactures false confidence, and where the losing reasoning is exactly what you want preserved for the later postmortem.

holdout is for that second class of question only. It applies the practice that courts, medical boards, and intelligence analysis have long used — preserved, structured dissent under conditions of high stakes and absent ground truth — to software-assisted decisions.

What it is not

• Not a synthesis engine. There is no step that merges the positions into one answer.
• Not an accuracy tool. It is not benchmarked and concedes that consensus methods win on questions that have answers.
• Not a decision-maker. On a split it returns a structured input to a human, not a verdict.
• Not a learning system. The store records and retrieves; it does not score agents.

Install

pip install holdout

Use

from holdout import Agent
from holdout.protocol.engine import Panel
from holdout.providers.openai_compat import OpenAICompatProvider

panel = Panel(
    agents=[
        Agent(name="empirical",    mandate="Reason from data and measurable outcomes"),
        Agent(name="principled",   mandate="Reason from duty, cost, and absent stakeholders"),
        Agent(name="practitioner", mandate="Reason from pattern, precedent, and experience"),
    ],
    provider=OpenAICompatProvider(...),
)

record = await panel.deliberate(
    "Should we move the auth service to a new language?",
    tier="hard_to_reverse",        # the caller asserts reversibility
)

record.outcome      # 'majority' | 'split' | 'fragile_agreement'
record.positions    # every agent's full rationale + vote, verbatim
record.minority     # the preserved losing rationale
record.crux         # the consequence-anchored crux (only on a split)
record.to_report()  # render the self-contained report file

holdout "Should we adopt this dependency?" --tier hard_to_reverse --report decision.html
holdout similar "Should we adopt a different dependency?"

Status

Early. The type contract and provider seam are in place and verified; the deliberation engine, store, report, and CLI are built in order per the build spec. See CLAUDE.md for the build plan and the two invariants that must hold throughout.

Design

Three documents describe the project at three altitudes:

• Build spec — what to implement, and nothing else.
• Testing strategy — how "correct by inspection" becomes "correct by a green run," and why the two invariants are tested as structure rather than example.
• Design document — the full case, the boundary of the claim, and the cross-field evidence. Not needed to build.

License

Apache-2.0.