substrate-framework AGENTS.md

Substrate Framework Agent Contract

This repository turns a sequential physics corpus into a self-consistent, importable, review-governed framework. Apply these rules to the entire framework. They are not a special cleanup rule for any particular late campaign.

New contributors begin with AGENTS_START_HERE.md for the operational collaboration, memory, GitNexus, skill-selection, PR, and review workflow. This file remains the normative scientific and governance contract when the two documents differ.

Goal

The sole purpose of this repository and every agent working in it is the advancement and full promotion of meaningful validated claims. Review exists to find and protect the strongest useful statement the evidence supports, not to maximize caution, exclusions, check counts, or process artifacts. Recursive claim weakening is a known systemic failure mode of multi-agent research repositories, and guarding against it is part of this contract, not an exception to it. Across agent generations hedging compounds: one reviewer narrows scope, the next adds a gate protecting the narrowed residue, a third demands stronger evidence for what remains — until solution spaces close, hurdles cannot be cleared, and only vacuous claims survive. That is not rigor; it is the objective silently replaced by process. Therefore:

• Every acceptance gate must name the concrete failure mode it prevents and what it would have changed in a past attempt. A gate with no such pedigree is removed, not enforced. Gates accrete only through named mechanisms, never through caution alone.
• Every attempted route terminates in exactly one route-scoped verdict: established as stated, refuted with the mechanism named, or blocked with the route's missing construction named. qualified is a waypoint, never a resting place. None of these route verdicts is automatically a verdict on its obligation or parent question. A failed route is a branch point: execute the continuation ladder in-run—method repair, reformulation or representation change, and materially different candidate generation—or demonstrate why a step is inapplicable to that route. This ladder is a minimum continuation response, not permission to block the obligation after three moves. Obligation-level no-go or blocking requires the scientific-exhaustion certificate; a finite ladder with plausible routes untried is an incomplete attempt. Single-route termination is the mirror image of recursive claim weakening and equally a systemic failure mode of this contract.
• Freeze four distinct levels for decomposed work: the parent campaign objective, each dependency obligation or rung, each candidate route for that obligation, and each concrete attempt. A route verdict answers only its route. A rung remains active until its positive intent is established or its complete in-scope solution space is exhausted. Passing a prerequisite immediately activates the next unsatisfied parent obligation; neither successful nor failed child work inherits upward as campaign completion without the frozen closure map proving that implication.
• Scope reduction must argue that the reduction changes what downstream work can conclude. Reduction that leaves no consumer better-informed is drift, not honesty.
• After a claim survives two independent evidence passes, the burden shifts: further narrowing requires new contrary evidence, not new caution. When successive reviews keep shrinking a claim without such evidence, the review chain — not the claim — is failing; escalate to the user instead of narrowing again.