research-dossier-discipline

Note: the crate paths, documents, and measured numbers below are illustrations from the aurora-agent workspace where these methods were developed and tested. The discipline itself applies to any autonomous or semi-autonomous research write-up.

Research-dossier discipline

An autonomous run can produce real measurements and still produce a dishonest dossier, because the dishonesty lives in the write-up's altitude: measurements made at one level, claims stated at another. This skill fixes the altitude.

What a run may claim

An autonomous measurement run over synthetic worlds may claim, at most:

• Measurements over the worlds it actually ran — fixtures and seeded generator families — identified precisely enough to regenerate (spec, seed, grid), with numbers pinned to retained artifacts or tests.
• Observation- and evaluation-level statements: strategy X was admissible in N of M cells of this family; these two methods selected identical fact sets on every measured world; this postcondition failed on these instances.
• Properties of its own instruments, stated as such: this benchmark cannot discriminate these methods; this tie is overdetermined by this construction property.

Every such claim inherits the scope of its worlds. The workspace's findings say it plainly of their own headline table: "That is partly a property of the benchmark, not of the methods."

What a run must never claim

This block is the honesty frame the workspace applies wherever research output is described. Carry it into the limitations section verbatim wherever your dossier describes research:

This is autonomous measurement science over SYNTHETIC decision worlds (fixtures and seeded generators). It can never claim: biology or medicine, literature or prior-work coverage, external-world observation, or release-level claims from fixture evidence. Oracle review is a human gate.

Unpacked:

• No biology, no medicine. The evaluation contracts enforce this mechanically — the reproduction check refuses to convert a matched pipeline into a validity claim ("matching a pipeline cannot be promoted into biological validity", docs/EVALUATION_REPRODUCTION_CHECK.md), and the audit projections each end with the boundary that they do not establish biological, causal, or clinical validity. Where research is described at length, include the boundary sentence: "Research and developer infrastructure: it does not diagnose an individual, recommend treatment, triage care, enroll participants, or claim medical-device functionality."
• No literature coverage. A run that searched nothing read nothing. "Related work" and "novel" are claims about the world's literature, and a dossier built from local fixtures has zero evidence about it. Say "not compared against prior work" rather than implying a survey.
• No external-world observation. Synthetic worlds are constructed to have properties; finding those properties confirms the generator, not nature.
• No release-level claims from fixture evidence. "Works on the shipped fixtures" is a regression statement. Readiness, robustness, and generality are claims about worlds not yet seen; the workspace's own tie is the cautionary case — a method that matched the compiler on every world measured, on a family whose construction overdetermined the match.
• Oracle review is a human gate. Where a pipeline's judge is itself under evaluation, or a verdict feeds a consequential decision, a human review is part of the method, and the dossier records whether it happened — a structural success is never collapsed into approval.