Research Integrity (M1-M8)

This skill is a prompt-level discipline, not an MCP server or a CLI. It is the agent's own pre-flight checklist for catching eight recurring AI research failure modes before a computation begins, before the work crosses an approval boundary, or before it gets folded into durable project artifacts.

The skill does not replace the machine-enforced gates that already live in the control plane (nullius A1–A5 approval gates, HARNESS_INVOCATION_REQUIRED anchor gate, quality_compile, quality_originality, and convergence gates). It is the agent-side check that runs before those gates fire — so the gate hearing is fair and the durable record actually reflects work that was done.

When to use

Run M1–M7 immediately before any of these moments. These are observable file events — none of them requires the engine's approval flow to be active, so a project that never calls nullius approve still hits every trigger below:

• Folding a new number, claim, or result into research_plan.md#Current Status or the research_contract.md claims table.
• Checking off a task-board item whose output later work will build on — the item that makes a claim, interface, or result durable, not each sub-step or remediation retry inside it; sub-steps inherit the walk owed at their parent boundary, and a walk already recorded for a boundary is not re-owed for intermediate patches that create no new durable claim.
• A milestone closeout commit.
• Promoting or superseding the current main research report.
• Marking a research-team cycle as converged, or acting on a convergence gate that reports the cycle converged.
• Assembling submission material, or handing off to another agent or human via the research-harness "Fold Results Back" step.
• Requesting nullius final-conclusions on a run.
• Posting a draft to research-writer or invoking a referee-review pass.
• Calling nullius approve <approval_id> for an A1, A2, A3, A4, or A5 gate, in projects that use the engine's approval flow.

M8 has a distinct trigger: run it before starting each new heavy or production computation, or hand-building a reusable numerical method, solver, quadrature, or robust primitive. Unlike M1–M7, this trigger fires before the work, not at a durable boundary: re-scan the upstream toolkit and every sibling or related project for a reusable implementation first.

You may run a smaller subset earlier (for example M2/M4 during a literature pull, the Extraction / transcription fidelity check when you transcribe a source into a deep-read / extraction note, or the Reference-reproduction fidelity check when a result starts claiming to match a published value) and a fuller pass at the boundary. The M1–M7 boundary pass is non-optional. M8 is separately non-optional whenever its pre-computation trigger fires.

Boundary walk and focused reading

At every M1–M7 boundary, judge each mode for applicability; read and perform every applicable mode's minimum disconfirming checks before crossing. If applicability is unclear, read that mode rather than skipping it. The table routes the walk; it does not replace the checks or turn mandatory gates into optional advice.

An active team cycle narrows only the M5 sweep/branch and M6 method-binding checks explicitly covered by its capsule. M5b production-configuration precondition checks have no in-cycle exemption. A smaller-scale pass, clean reproduction or convergence declaration cannot establish a production precondition.

• Before recording, approving or closing a boundary, read boundary-recording.md. Record modes, actual disconfirming evidence and per-mode inapplicability reasons; consolidated wording is allowed. Failed checks stop the boundary, then only affected modes are re-walked after the fix. Keep narrative detail and the required machine receipt distinct; the current machine receipt accepts M1–M7 only, so record M8 in narrative. Approval-flow and conversational/file-mode decisions keep their specified receipts.
• Before notebook closeout, read notebook-closeout.md. Structural links and accepted current full-text content review are separate requirements. Missing, partial, unavailable, stale or capacity-unverified evidence does not pass. The M1–M7 self-check cannot issue its own content-review verdict, and semantic review does not establish scientific truth.
• For scope, provider routing or overlap with other gates, read scope-and-integration.md and adjacent-skills-and-domain-routing.md. This discipline is generic and supplements existing gates. Route evidence by the cited work and quantity, not by a tool package name; unsupported provider coverage is not a failed scientific check.

Read references/transcription-fidelity.md when you transcribe a source into a deep-read or extraction note, restate a source's equations, values or conventions, or review such a note (the Extraction / transcription fidelity section and its failure checklist, moved there verbatim).

Read references/numerical-claim-checks.md when a result starts claiming to match a published value, an upstream fix changes results already recorded, a claim is defended as validated, or a heuristic fast path is accepted (the Reference-reproduction fidelity, Post-correction impact accounting, Validation-chain validity and Accelerated / heuristic fast-path honesty sections, moved there verbatim).