The checker is now a machine too
AI scientists propose hypotheses, analyze data, and draft results, quickly and in volume. When one returns a finding, how do you decide whether to believe it? You check it. But human attention does not grow at the rate of machine output, so the checker is now often a machine as well. That turns a simple question into a harder one: what does it mean for one machine to check another’s work?Reliability came from diversity of error
A published result is rarely repeated exactly, and most findings are never directly replicated. What has protected the record is the structure around a finding. Independent groups use different data, different instruments, different habits, so the mistakes they make tend to differ. A wrong result, set against a field whose errors point in many directions, tends not to fit. Two lines of evidence resting on mistakes unlikely to coincide: their agreement carries information. Two lines that share the same mistake: their agreement carries little, however confident each one is. The property doing the work is diversity of error. It is the same property that lets a human-and-machine team beat either alone. A teammate adds value by being wrong in different places, not by being right more often.Machine science loses it, invisibly
When one machine checks another, the two are seldom independent. They share training data, weights, tools, and assumptions, so their errors are related from the start. A second machine’s agreement is not worthless, but it is worth less than it looks. It reads as independent corroboration while carrying a fraction of the information a genuinely independent check would give. Two things make this specific to machine-generated science. The loss of independence is built in, because sharing a model produces shared blind spots. And the loss does not show in the finding: two systems that share everything look, on the page, like two independent labs.What Mareforma does
Mareforma treats independence as a quantity to surface, not a verdict to declare. It does not pretend to sit outside the system that produced the finding: on the default library path it runs in the producer’s own process, under the same user, with the signing key loaded in the same interpreter. It is not an external auditor, and claiming to be one would be the exact self-report it exists to replace. What it supplies instead is a structural check on the record and an honest account of its own reach, so a producer’s account of its work is measured against what the run did rather than taken on its word. A finding enters the graph as a signed claim carrying its provenance: what it builds on, what it contradicts, who signed it, and what the run that produced it actually did. From that record Mareforma reports what a reader cannot see on the page.- Grounding. The cheapest check: did real data reach this finding, or did the
model fill the gap from its prior?
observe()watches the run and returns GROUNDED, UNGROUNDED, or OPAQUE when the read happened somewhere it could not see. The catch is that a silent failure does not hide behind OPAQUE: when a cited read is attempted and every open of that path fails, the scope lands UNGROUNDED, so a pipeline that quietly read nothing is named, not excused. - Effective independence. The lead signal is a computed count: the number of
pairwise-distinct (model, data, signer) checks behind a finding. A second key
minted from the same model shares the same blind spots, so it does not raise
the count when the model each check used is observed. A genuinely different
model, or a human check (one of the more dependable sources of independent
error), does. A check whose model was not observed cannot certify a distinct
model, so it counts on the older signer axis and the trust map marks the
difference, rather than silently reading as independent. The older
REPLICATEDandESTABLISHEDlabels still resolve this release as deprecated aliases, but the number is what to read. - The residual. Where independence cannot be checked, Mareforma says so. The
trust map places each property at its tier:
COMPUTEDfor what it measured from evidence,PROXIEDfor what it could only reach through a stand-in whose bound it names (re-execution stands in for faithfulness), andDEFERREDfor what it does not evaluate at all, instead of rounding up to a verdict.
Who it is for
- Autonomous AI scientists that need a shared memory: findings accumulate, and a new claim can query what is already established before it is made.
- Labs building AI-science infrastructure that need to know which agent outputs are grounded, which are replicated, and which are speculative, without changing the agents.
- Multi-agent systems where convergence across runs, agents, and data sources has to be measured rather than assumed.