Proof-Carrying Research Loops

Status

This is a working paper draft. The general program claim is a HYPOTHESIS, the local pandas reproduction is a PROBE_MATCH, and paper/source intake remains SOURCE_LEAD. No new theorem, natural law, official submission, or publication acceptance is claimed here.

Promotion ladder

SOURCE_LEAD

A paper, issue, dataset, repository, or source body worth examining.

HYPOTHESIS

A plausible claim with a named falsifier or missing experiment.

IDENTITY

A mathematical or system identity with a proof sketch, derivation, or executable model.

PROBE_MATCH

A bounded local computation matched the expected behavior.

CRUCIBLE_MATCH

Crucible assessed a claim as matching a measurement packet.

UNVERIFIABLE

Evidence is insufficient, inaccessible, or out of scope.

LAW_CANDIDATE

A repeatable invariant with clear scope and falsification.

PROMOTED_LAW

Reserved for independently checked work. None are promoted here.

Method

The research loop is tool-native, not prose-native.

• Gather captures source candidates and keeps them as source leads until stronger evidence exists.
• Index bounds the workspace context and source references.
• Forum routes the intent and review vocabulary.
• Crucible compares claims to measurements and returns MATCH, DRIFT, or UNVERIFIABLE.
• Learn turns proof packets into durable skill acquisition and next-study loops.
• Telos packages the loop into action receipts, research packets, and public handoffs.

Probe

A current OSS proof-packet example.

• The source lead is pandas-dev/pandas#63458.
• A clean local virtualenv probe records Python 3.12.10, pandas 3.0.3, and pyarrow 24.0.0.
• The issue-authored expression fails with AttributeError: 'ArrowExtensionArray' object has no attribute 'sum'.
• The readiness packet still blocks because there is no patch, no passing test evidence, and no Crucible MATCH after-state.

Publication rule

Every official paper should carry a website copy and an archive-ready copy. Both must preserve the same status labels, evidence paths, missing-evidence boundaries, and recheck commands. A later stronger receipt revises a paper explicitly; it does not silently overwrite the earlier claim.