Megapixel99's assay tool flags duplicate functions by running each on the same fixed inputs

Differential testing already checks whether two functions agree, but only for a pair somebody declared in advance ^[3]. assay scan skips the declaration. It walks each comparable function through one fixed, ordered list of probe values and records the return value or the exception type at every position ^[4]. Functions whose outcome vectors match land in the same hash bucket, so finding candidates takes one pass over the functions with no sweep over every pair ^[5]. It never reads function names ^[5].

The ladder is where the method is weakest, and the author, who publishes the code as Megapixel99/assay-checks ^[1], words the verdicts to match. A "differs" verdict comes with a witness input a stranger can replay ^[9]. A "same" verdict means only that no rung of a finite ladder told the two functions apart ^[9]. The tool fails the run on "same" for that reason: it is the weaker claim, and a person has to read the pair ^[9]. The author wrote that the lesson is worth stealing whatever you test with: "ask what characters your inputs never contain, then add them." ^[10]

Most of the engineering is in the guard against false matches. The author wrote that every clause exists because of a mistake it now blocks ^[11]. Two functions that raise TypeError on every input agree perfectly ^[11]. So do two that return the same constant ^[11]. Without a guard, the scan would call every one-argument function everyone else's twin ^[11], a result with perfect recall and no use.

Counting distinct outcomes did not fix it ^[12]. One returned value plus one exception counts as two outcomes, and that rewards a probe that found a function's type errors and never reached its behaviour ^[12]. Comparing whole vectors against the identity failed too. A transform whose vocabulary the ladder lacks is the identity wherever it answers and raises everywhere else, so the comparison has to be about the positions where it answered ^[12]. Then two unrelated query-parameter transforms matched on every rung. The ladder held no key either one recognised, and both copied the object through ^[13]. A copy is now rejected alongside the identity ^[13].

Aliasing is the other route to a false pair. A CommonJS module whose export is a function arrives under two keys, and a barrel module hands back the objects its dependencies defined ^[14]. One helper then appears as registry.js::truncate and as truncate.js::default ^[14]. Object identity is the correct test for this, and assay uses it to drop those ^[14]. Names and source text play no part, so a function genuinely copied into two files is still two objects and still counts as two implementations ^[14].

The author's coverage figure ^[16] leaves about nine functions in ten outside every verdict ^[1]. That figure comes from someone else's workload. It carries over only to a codebase with a similar share of functions that pass the probe rules ^[15]. I'd expect a class-heavy or I/O-heavy tree to come in lower. I think the census is the best decision in the tool, and the author calls it the design position they would defend hardest ^[16]. Each scan ends with counts of what it refused and why, files and functions as separate populations, with probed plus not probed equal to the function count ^[16].