What Counts as Evidence When the Run Gets Messy
Praxist connects a task's scientific contract to a generation's committed lifecycle, but those two facts do not make every output equally trustworthy. The practical question is how to classify a result, finding, report, or late artifact so that it can guide the right next decision without being promoted beyond what it proves.
Credible progress has two authorities
The first authority is scientific: the task project defines the objective, evaluator, metrics, protocol, constraints, and evidence standard. The second is lifecycle-based: Praxist records whether the resulting work passed through the ordered boundary that turns peer output into committed run state. A number can be meaningful under the task's protocol but still belong to an unfinished generation. Conversely, a generation can be fully committed even when its result is weak, negative, or insufficiently mature.
This distinction creates a useful chain of authority:
- The task says what question the experiment is allowed to answer.
- The peer records what it changed and why.
- The evaluator produces the result summary.
- Praxist materializes a finding that explains how later research may use the result.
- The generation boundary decides what is committed and what can influence inheritance.
Each step adds organization, not permission to strengthen the underlying claim. A planning agenda can recommend a test, but it is not a measurement. A frontier can retain a candidate for future work, but retention is not independent validation. A report can present a clear comparison, but it is a derived view of canonical state.
That separation is the key connection between the task boundary and the generation lifecycle. The task owns the meaning of “better.” The generation boundary owns the fact that a particular set of observations was reconciled at a particular point in the run. Praxist can preserve both without pretending that either one alone establishes scientific truth.
A result summary is not a finding
A result summary answers, “What did the evaluator measure?” It is produced by the task-owned evaluation path and should carry the metrics, protocol facts, stage or coverage information, caveats, and provenance needed to interpret that measurement. The summary is closest to the experiment itself.
A finding answers a different question: “How may later research interpret or use what was measured?” It can connect the result to a mechanism hypothesis, a failure mode, a trade-off, a next-step intent, a retention lane, or a parent decision. That interpretation is valuable because later peers need more than a number. They need to know whether the result is a credible improvement, a diagnostic control, a partial signal worth repairing, or a negative result that closes off an approach.
The distinction limits both objects. A finding must not invent a metric that the evaluator did not publish. A result summary does not become a theory merely because it contains a high score. Praxist's materialization layer carries evidence metadata into canonical findings, but it does not turn a peer's explanation into an independent measurement.
This is why a finding may preserve a caveat instead of hiding it. If the evaluation covered only a preliminary portion of the task, the finding can say so and remain available for validation. If the evaluator reports a protocol-integrity problem, the result can remain visible for diagnosis while being excluded from mature promotion. If a result arrives after the generation's evidence cutoff, it can be useful for follow-up without rewriting what the closed generation had actually seen.
The design favors preservation and calibration together. Discarding every imperfect output would erase information about dead ends and operational failures. Promoting every output would turn uncertainty into false lineage.