Existing scientific agents inherit the failure mode of general coding agents: under pressure to finish, they fabricate, skip, or smooth over results. The AfS (Agent for Science) platform instead embeds research discipline into the system architecture itself. It enforces iron laws mechanically—commitment before measurement, unforgeable freezing, evidence that persists but verdicts that do not, first-class negative results—rather than relying on the model to behave.
The platform organizes around three time horizons: within a run, a minimal set of research nodes with strict contracts; within a project, an inquiry contract with a hash-chained artifact ledger and a referee loop that makes every conclusion re-verifiable from disk; across projects, a two-tier knowledge base with promotion by rewriting. No benchmark is reported; three traces and two closed campaigns illustrate how the mechanisms catch failures. A process-integrity suite for quantitative comparison is under construction.
The work is interesting because it shifts the source of trust from model honesty to system invariants, addressing a root cause of unreliability in long-horizon scientific work. The cost is only the operational overhead of the machinery, which the paper measures.