assay

Two questions ordinary CI does not answer, about work that already passes its tests:

Could those tests have failed? Does the tree already answer this?: in either language

A green suite tells you the code did what the suite asked. It tells you nothing about whether the suite could have objected, and nothing about whether the code needed to be written at all.

pip install assay-checks      # the CLI: assay
npm install -g assay-checks   # the CLI: assay (JavaScript projects)

$ assay scan src/                      # the Python half, over a Python tree
FINDINGS — 1, each checked rather than guessed:
  finding  same answer (arity1/v3): src/format.py::humanize, src/report.py::pretty
           no input in the ladder told them apart — READ them; only a person decides
           whether the duplication is a defect

The JavaScript half answers in the same words, because the verdicts, the exit codes and the ladder key are one contract rather than two:

$ assay scan src/                      # the JavaScript half, over a JavaScript tree
FINDINGS — 1, each checked rather than guessed:
  finding  same answer (arity1/v3): src/slug.js::slugify, src/url.js::toSlug
           no input in the ladder told them apart — READ them; only a person decides
           whether the duplication is a defect

...and one contract is enough to ask the question across the two, which is where the duplication a polyglot repository actually accumulates:

$ assay cross src/api.py::shout src/ui.mjs::yell --with assay-js
FINDINGS — 1, each checked rather than guessed:
  finding  same answer across languages (cross1/v3/d3b2ba61ccb7):
           src/api.py::shout [python]  vs  src/ui.mjs::yell [javascript]

Neither half is a linter and neither is a test runner. Both are stdlib-only, in both languages, because a quality tool that drags in a dependency tree is one more thing that can break the build it was added to protect.

Half one: could those tests have failed?

assay runners audits mutation harnesses: the scripts that deliberately break your code and check that something notices. If you run one, these are the seven ways it can lie to you, and each is a failure that looks exactly like success:

The last one is about the third one's blind spot. restore-in-finally proves the restore path executes; it says nothing about the file on disk. A harness that restores from a buffer it read after mutating, or writes the text back in a different encoding, or saved one of the two files it touches, satisfies the other six and still leaves the tree wrong, and every suite after it scores code nobody wrote. Hashing before and comparing after is the check, and the detector wants both halves of it: a digest nothing compares is arithmetic, and a message nothing computes is a string.

And sigterm has a blind spot of its own, which is SIGKILL. SIGKILL cannot be caught, blocked or handled: no handler runs, no finally runs, and no property in the table above would have helped. The ordinary way to be SIGKILLed is not an impatient person but a timeout: subprocess.run(..., timeout=...) kills the child outright, and so does the kill step of a CI runner that has waited long enough. So a harness satisfying all seven, invoked under a timeout it then exceeds, leaves the tree mutated exactly as though it carried none of them, and every suite after it scores code nobody wrote.

The remedy is not in the harness and cannot be: it belongs to whatever invoked the harness, which has to check that the tree came back rather than trust that the harness was given the chance to put it back. That is restore-verified's argument one level up; a restore that ran is not a restore that worked, and this is a restore that never ran at all. See Running harnesses in CI below for the shape of the check.

That argument is now measured rather than asserted. conformance/ runs four real frameworks (mutmut, cosmic-ray, Stryker, PIT) under exactly this interruption, timed to land at the instant a mutant is observably on disk, and hashes the tree afterwards. Three of the four never mutate the tree at all: mutmut copies the project into mutants/, Stryker into .stryker-tmp/, and PIT mutates bytecode in memory, so their source survives SIGKILL because it was never dirty. cosmic-ray, which mutates in place, is left DIRTY by a plain timeout: a SIGTERM, not a SIGKILL, so it fails sigterm before this blind spot is even reached.

The reading to take from that is narrower than the paragraph above and more useful: the invoker's check is load-bearing for harnesses that mutate in place, and if you are writing one, the remedy that actually closes the hole is to mutate a copy. No process can promise to clean up after being killed, and the only way to have nothing to clean up is to have put nothing there. The suite carries a harness satisfying all seven properties as its calibration row, and that harness is still DIRTY under SIGKILL: see conformance/README.md.

They collapse into one rule worth remembering on its own, because the seven are just instances of it:

A harness must answer separately whether the suite RAN, whether it FAILED, and whether the failure was the RIGHT one. Collapsing any two of those three is how every defect in this family happens.

assay anchors checks the other half of a mutation table: every anchor string must match its target exactly once. An anchor matching zero places is a guard nobody is testing any more; one matching twice means replace(old, new, 1) took the first, which may not be the one you meant, so the harness mutates something nothing asserts and reports NOT DETECTED. That reads as "your guard is untested" when the truth is "your mutation tested something else", and those send you to opposite ends of the codebase.

assay diff asks whether a change carries the checks it needs: a guard added with no mutation exercising it, a changed file no harness names, tests named for a limitation that a capability change will turn red on working code.

Half two: does the tree already answer this?

A property suite is per-artifact and behavioural. Duplication is cross-artifact and structural. Two implementations that both pass are two implementations that both pass; correctness was never the question duplication asks.

The usual instrument is a differential test, and it works, but you have to know which two to compare: the pairing is declared, so it only ever covers pairs somebody already suspected. assay scan finds them instead.

Every comparable function is probed once against one deterministic ladder of inputs, producing an outcome vector. Two functions are candidates for being the same function exactly when their vectors match, so discovery is a hash bucket rather than a quadratic sweep, and the decider is execution, not text.

Names are never read. In the tree this grew out of, it paired is_wordy with _word, which no textual or name-based detector puts together.

One function is never a pair with itself, and the ways it can look like one are not obvious. A CommonJS module whose export IS a function arrives under two keys, default and module.exports; a barrel module hands back the very objects its dependencies defined, so a helper is reachable as both registry.js::truncate and truncate.js::default. Both are one function wearing two names, and both are rejected by identity rather than by comparing names or source text, so a function genuinely copied into two files is still the two implementations it is.

differs is proof. same is not. A witness is a fact; agreement across a finite ladder is the absence of one. See What same is worth below: it is one character.

# Python
assay scan src/                                            # discover
assay pair src/format.py::humanize src/report.py::pretty   # the declared route, one pair
assay search src/format.py::humanize --in src/ lib/        # search before you generate
assay why src/format.py::humanize                          # ...and if it was not probed, why
assay why src/format.py::humanize --cross                  # ...and why it is in no bundle
assay accept --reason "read them; merging needs the router change"   # accept what you have read
assay cross src/format.py::humanize src/ui.mjs::pretty --with assay-js   # across the boundary
assay sweep src/ --against js/src --with assay-js          # ...for whole trees, naming no pair
assay bundle src/ > py.json                                # ...or hand the other half a bundle
assay search --stdin --in src/ lib/ < draft.py             # ...before it is a file
assay search --stdin --against js.json < draft.py          # ...and before it is a file, in the OTHER language
assay why --stdin < draft.py                               # ...and whether it can be searched for at all

# JavaScript: the same commands, and a reference is FILE::NAME in either language
assay scan src/
assay pair src/slug.js::slugify src/url.js::toSlug
assay search src/slug.js::slugify --in src/ lib/
assay search --stdin --in src/ lib/ < draft.js
assay why --stdin < draft.js

--stdin is what "search before you generate" actually needs. A FILE::NAME names something that already exists, so a command taking only one asks you to write the file first, which is the thing you were trying to find out whether to write. What arrives on stdin is a snippet parsed as a module, not a bare function: it may carry the imports and helpers the function needs, exactly as the file it is about to become would. Two definitions in one snippet and --name says which, because picking one would make the tool answer about code nobody asked about. assay why takes the same two ways in, because it is the same question asked one step earlier: writing the file first in order to be told the file was never the problem is what --stdin exists to avoid.

A refusal reason is the FIRST gate, not the only one, and the census cannot show that. The tally counts one reason per function, so its buckets are not independent and do not add up:

  no arguments                                 21
  arity 4                                      14
  needs docx                                   12

Raising the arity cap there frees none of the fourteen: every one also trips needs docx, touches os or needs matplotlib. assay why is where that gets said, because it is where somebody goes after reading the census:

$ assay why training/self_play_loop.py::retrain
look     training/self_play_loop.py::retrain — arity 4 (no ladder above 3)
         it trips 5 gates, not one: arity 4 (no ladder above 3); touches random;
         touches subprocess; touches os; calls open() — the census counts only the
         FIRST, so clearing that one alone would still leave this function unprobed

A query the ladder cannot tell apart is a look, never a none. The census files every function it cannot discriminate under not probed, so a constant or a projection can only fail to find the other constants and projections: the match was never possible, and printing the clean result there would say we found none where the truth is we never looked:

$ assay search --stdin --in src/ lib/ < draft.py
       the tree was not searched: the census excludes every function this ladder cannot tell apart, so a match was never possible
  look     <stdin>::k — not discriminated by the ladder
           1 distinct returned value across the 31 rungs that answered, and 2 is the minimum — as far as this ladder can see it is a constant

It is the same answer assay why gives about the same vector, from the same code: the two commands cannot disagree about it. A query the ladder can discriminate still gets the ordinary same none, which means the tree really was searched.

The pair no differential test covers: one function, two languages

A validator reimplemented in a Django backend and a Node frontend is the highest-value duplication a polyglot repository has, and it is exactly what nobody writes a differential test for, because writing one means agreeing, by hand, on what False and false have in common.

# both binaries installed
assay cross src/api.py::shout src/ui.mjs::yell --with assay-js

# or one half writes a record and the other reads it, which needs neither to know
# the other exists
assay probe src/ui.mjs::yell > yell.json     # the JavaScript binary
assay cross src/api.py::shout yell.json      # the Python one

finding  same answer across languages (cross1/v3/d3b2ba61ccb7):
         src/api.py::shout [python]  vs  src/ui.mjs::yell [javascript]

Two things have to be true before that means anything, and the tool asserts both rather than assuming them.

One ladder, not two that resemble each other. BASE_VALUES is a hand-written list per language, and the strongest thing that can be said about the pair is that they cover the same shapes: the languages have different primitives, so comparing lengths would fail for a correct reason. That is enough for two Python functions and nothing like enough here, where two lists that were meant to hold the same values and quietly stopped is the entire hazard. So the cross ladder is one JSON document, carried verbatim by both halves and parsed by each; test_parity.py compares the two texts, and the ladder key carries a digest of the rungs so a comparison across a changed ladder is refused by the branch that already refuses a mismatched arity.

One vocabulary for outcomes. V:False and V:false are two spellings of one answer. The interlingua renders every value as canonical JSON, and the three lossy mappings are choices about which mistake to make rather than accidents:

Anything JSON cannot hold (bytes, a Map, a Date, a class instance) is refused rather than approximated, and one such outcome makes the whole comparison a look. NaN and the infinities are spelled out, because JSON.stringify turns all three into null: three different answers reported as one absence.

A raise carries no name. The two languages' error taxonomies genuinely diverge: d['x'] is a KeyError in Python and undefined in JavaScript, so comparing names would make every honest pair differs, and declaring them equal is worse, because same is the verdict that fails. Every raise renders as one token, which masks a rung where both sides raised: two of them can never be a witness, and the vacuity guard counts only returned values so two of them can never be evidence either. A rung where one raised and the other answered stays a witness, and it is the most interesting kind there is.

Which is how the README's own ½ lesson reads across the boundary:

def is_wordy(tok):
    if not isinstance(tok, str):
        return False
    return tok[:1].isalnum() or tok[:1] == "_"

export function isWordy(tok) {
  if (typeof tok !== 'string') return false;
  return /[A-Za-z0-9_]/.test(tok[0] ?? '');
}

ok       differs: word.py::is_wordy [python]  vs  word.mjs::isWordy [javascript]
         ["½"] -> V:true vs V:false

The two halves do not invoke each other. pip install assay-checks gives you one and npm install assay-checks gives you the other; neither can assume the other is on the machine, and a command that shells out to a binary that may not exist fails in a way that reads like the code being wrong. So assay probe writes a record and assay cross reads one, and --with CMD runs that first step for you when both are installed. Two references in the same language are a look pointing at pair, which compares them on their own language's fuller ladder.

...and the pair nobody named: bundle + sweep

cross answers about two functions somebody already suspected. Nobody suspects the pair that matters. A rule written once in the API and again in the front end, by two people, a year apart, is exactly the duplication no one goes looking for, so the command that finds it must not need either name.

Finding it means probing both trees, and the halves still do not invoke each other. So one writes a bundle and the other sweeps against it:

assay bundle js/src > js.json           # the JavaScript binary: every function's cross vector
assay sweep  src/ --against js.json     # the Python one: which of mine does that tree answer?

# or, with both binaries installed, in one step
assay sweep src/ --against js/src --with assay-js

finding  same answer across languages (cross1/v3/d3b2ba61ccb7):
         src/api.py::shout [python]  vs  src/ui.mjs::yell [javascript]
         no input in the shared ladder told them apart — READ them; only a person
         decides whether the duplication is a defect

12 files, 0 not loaded
41 functions, 9 probed, 32 not probed
  zero-arity, so no input ladder applies         19
  not discriminated by the ladder                 8
  touches os                                      5
[javascript] 8 files, 1 not loaded
[javascript] 30 functions, 7 probed, 23 not probed
  zero-arity, so no input ladder applies         14
  could not load                                  6
  uses `this`, so it is a method                  3

A bundle is many assay probe records in one envelope, byte-identically shaped, so an entry lifted out of one is a record assay cross already reads. It is versioned apart from the record it carries: adding a key to the envelope does not change what any one record means by vector, and both schemas are checked before anything is compared, because comparing a new answer against the wrong earlier answer is precisely the defect a difference checker exists to catch.

Both censuses are printed, and the far one is the point. A function the other binary refused was never compared. A report that says same none while staying quiet about the twenty-three functions the far side never probed is reporting we never looked as we found none: across a boundary where the reader has no way to check.

A bucket is only a comparison because of what never reached it. sweep groups by vector equality, which is legitimate only because every pair assay cross would refuse was refused first: an outcome the interlingua cannot state, and a vector no rung told apart from a constant. One place decides that for both commands. Two places deciding it is how the tree-wide command comes to print a finding for a pair the pairwise command calls a look; two answers to one question, and the weaker one on screen.

A bundle of your own language is a look-shaped refusal pointing at scan, which compares one language's functions on its own fuller ladder.

...and for ONE function, before you write it: search --against

sweep needs a whole tree on both sides; cross needs the pair named. Neither answers the question you actually have at the keyboard: I am about to write this one function; does the other language already have it?

assay bundle js/src > js.json
assay search src/format.py::humanize --against js.json    # a function that exists
assay search --stdin --against js.json < draft.py         # ...and one that does not yet
assay search --stdin --in src/ --against js.json < draft.py    # both corpora, one run

finding  the javascript tree already answers <stdin>::loud: src/ui.mjs::yell
         no input in the shared ladder told them apart — READ them before writing
         a second one

--stdin is the point of it. Search before you generate cannot mean "first write the file", and the cross-language form is where that bites hardest: the duplication you are about to create is in a language your editor is not even open in. A FILE::NAME works too, for the function you already wrote and now suspect.

--in and --against are independent, and either alone is a complete question. Give both and you get both verdicts in one run, kept apart: the native ladder is the stronger of the two, so the answers are not interchangeable and neither replaces the other. Give neither and it is exit 2 rather than a clean no findings from a run that looked nowhere.

A query the shared ladder cannot tell apart is a look, never a none: the same rule the one-language search follows, from the same admit that fills the bundle. A constant can only fail to match the other constants, because every constant was kept out of the bundle in the first place; printing the clean result there would say we found none where the truth is we never looked, on the one path where the reader is about to write the function.

...and when it does not cross, why --cross says which gate

sweep prints 41 functions, 9 probed, 32 not probed. That is the right shape for a tree and the wrong shape for a question: somebody who expected a particular function to cross cannot read not discriminated by the ladder 8 and learn whether theirs is one of the eight.

assay why src/cache.py::entries --cross

look     src/cache.py::entries — an outcome the interlingua cannot state
         22 of 29 rungs answered with one, the first at [0] -> X:set — the
         interlingua is JSON, so bytes, a set, a Date or a class instance
         cannot be said in it

A native why cannot answer this, and answering as though it could is the failure. The very same function:

ok       src/cache.py::entries — probed on arity1/v3: 24 of 31 rungs answered,
         24 distinct value(s)

Both are true. The native ladder probes it happily; the shared one cannot state a set at all. The two ladders hold different values and refuse different functions: a function the native ladder discriminates can be a constant on the shared one, because the shared one is the intersection of what the two languages can express. Reporting one verdict for the other question would be confident and wrong, with nothing on screen to say the two ladders had been asked different things.

The rung is named, not just counted. X:set is a fact about one input, and a person with the function open can usually see immediately which of their return paths it is. A count alone sends them back to reading the whole thing, which is the work the answer was supposed to save.

A FILE-level refusal deliberately says nothing about which ladder was asked, because it is the same answer for both: the module was never loaded, so no function in it was looked at on either.

--stdin works here too, for the same reason it works on search.

Three verdicts, and they are never mixed

finding   something was CHECKED and is wrong.        exit 1
look      a rule applies and this tool CANNOT decide. never fails
ok        checked and fine.                          printed, not silent

look never failing is a deliberate limit, not timidity. A check that reports things a person then has to dismiss stops being read, and an unread check occupies the place where a working one would go. Anything the tool cannot settle by looking at the code is offered for a human to settle.

ok is printed rather than left silent because "we found none" and "we never looked" are different claims, and only one of them is evidence. Every scan ends with a census of what it refused and why:

247 files, 32 not loaded
  reads the clock                               19
  touches os                                    13
1412 functions, 137 probed, 1275 not probed
  no arguments                                 274
  not discriminated by the ladder              127

A census answers about a tree; assay why answers about a name. Reading no arguments 274 does not tell you whether the function you expected to be probed is one of the 274, and guessing which of eight gates rejected it is the work the census was supposed to save you:

$ assay why src/format.py::humanize
  look     src/format.py::humanize — touches os
           refused before the ladder, so it is in no bucket and can pair with nothing

$ assay why src/slug.js::slugify
  ok       src/slug.js::slugify — probed on arity1/v3: 29 of 29 rungs answered, 23 distinct value(s)

It never produces a finding: it reports what the tool did, and decides nothing. --stdin asks it about a snippet instead of a name, with --name picking one definition out of several, exactly as search does. It also splits the one reason the census cannot split: not discriminated by the ladder covers a constant, a projection, and a function the ladder never reached, and those need a wider ladder, a different function, and inputs of another shape respectively.

On the JavaScript half the answer may be at the file level, and it says so: a file that reaches for the clock on the way in is refused whole, so none of its functions were ever looked at and every one of them has the same answer. A clock inside a body is a different answer; it refuses that function and leaves the rest of the file probeable.

Files and functions are counted separately, and the second line is an equation. A file nobody opened holds an unknown number of functions (not opening it is exactly why the number is unknown), so adding the two populations together prints a total nobody measured. Reading probed + not probed and not getting functions is the shape of that mistake.

Exit codes are identical for every subcommand, because scripts depend on them more than on anything printed: 0 nothing to read, 1 findings, 2 the tool could not run. 2 is never suppressible: "could not run" and "found nothing" are opposite situations, and letting the second silence the first is how a broken invocation reads as a clean audit for months.

--json, for the thing reading this instead of you

Every subcommand takes --json, on either side of the subcommand name, and emits one object instead of the prose report:

{
  "baseline": null,
  "command": "scan",
  "error": null,
  "exit_code": 1,
  "items": [
    {"verdict": "finding", "message": "same answer (arity1/v3): …",
     "where": "src/slug.js::slugify", "detail": "…"}
  ],
  "language": "python",
  "notes": ["…the census, verbatim…"],
  "root": "/abs/path",
  "scan": {"files": 247, "functions": 1412, "probed": 137, "not_probed": 1275,
           "skipped": {"no arguments": 274}, "unloadable": {"reads the clock": 19},
           "skipped_refs": {"src/fmt.js::pad": "no arguments (a ladder cannot …)"},
           "unloadable_paths": {"src/hbs-helpers.js": "reads the clock"}},
  "schema": 1,
  "tool": "assay",
  "version": "0.4.0"
}

One shape, always. A run that could not start emits the same keys as one that finished, with error set and items empty. Prose on the failure path and JSON everywhere else hands you a parse error at exactly the moment the tool could not run, and a sloppy consumer reads a parse error as no findings: the same collapse 2 is never suppressible to prevent.

look gets no severity. The three verdicts are not mapped onto somebody else's error/warning/note; that mapping is the collapse the vocabulary exists to prevent. The verdict travels by its own name and you decide what it means.

The census is data, not the printed equation. probed + not_probed == functions is yours to check rather than something you parse back out of notes, and files stay a separate population from functions. A command that ran no scan emits null rather than 0, because zero probed functions and no sameness half at all are different claims.

And it names what it never looked at. unloadable and skipped count; unloadable_paths and skipped_refs say which, and carry the whole reason rather than the tallied key. could not load 12 is a number you cannot act on: the only recourse is assay why FILE::NAME, which has to be told a file and a function name in it, the two things the tally withheld. The full reason matters most in the biggest bucket: a tally key stops at the first (, which is exactly where a load error's message begins, so could not load (JWT_SECRET must be set) survives here and nowhere else. sum(unloadable.values()) == len(unloadable_paths), so the two never drift.

The baseline's caveat travels as data. performed says what this run audited and unchecked names every entry it could not have seen fire, rather than an empty stale list that reads as "checked, found none". There is no complete boolean: completeness stopped being a property of the run when an entry learned to name the command that fires it.

Keys are sorted all the way down, in both halves, so one contract prints as one document rather than two. language is in the payload because a polyglot repository runs both halves over one root and a consumer merging two reports has no other way to tell which produced which.

schema is versioned separately from the tool. A consumer parsing this has the same claim on stability as a script reading the exit code, and the two do not move together.

Configuration

assay.json in your project root. One file serves both languages, deliberately: two files that had to be kept in step would be the exact duplication this tool exists to find.

{
  "runner_exempt": [{
    "path": "test/mutate_api.py", "property": "sigterm",
    "reason": "writes only under a tempdir, so a kill leaves nothing mutated"
  }, {
    "path": "test/mutations-http.js", "property": "parses-mutant",
    "reason": "every mutant goes through the bundler first, which rejects one that does not parse"
  }],
  "anchor_exempt": [{
    "path": "test/mutate_api.py", "reason": "anchors into generated source"
  }],
  "baseline": [
    "test/mutate_legacy.py: no `evidence` (no failures reported and no test executed look identical)",
    {
      "line": "same answer (arity1/v3): src/slug.js::slugify, src/url.js::toSlug",
      "reason": "one is the URL path form; merging them needs the router change first",
      "from": "scan"
    }
  ]
}

Paths are relative to --root, and the language of the path is not a category: runner_exempt and baseline take Python and JavaScript entries side by side, because a polyglot repository has one root and the audit that reads this file may be either binary. anchor_exempt affects only assay anchors, which both halves now run: the Python half by parsing the table out with ast, the JavaScript half by importing the harness and reading the exported table as data.

A baseline line is the exact text of a finding, and only a finding. It is matched whole, never as a prefix, so the line above is what assay printed rather than a description of it. A look cannot be baselined and does not need to be: it never fails the run, so there is nothing to accept.

An entry is that line, or an object carrying it as line. The bare string stays legal because adopting this means pasting lines out of a run, and a format that refuses the paste is a format nobody adopts. The object form carries the two things a string cannot:

from is what makes staleness a property of the line rather than of the run; see below. A from naming no real command is a hard error rather than a line nobody can ever check.

assay accept writes the entry for you, and refuses to write the two entries you should not have:

assay accept --reason "one is the URL path form; merging needs the router change"
assay accept "same answer (arity1/v3): src/slug.js::slugify, src/url.js::toSlug" --reason "..."

With no line it takes every new finding; with one it takes that one. It fills in from from the audit that actually produced the line, so the check that fires it is the one that can later call it stale. --reason is required.

It refuses a look: a look never fails the run, so the entry could never be suppressed and never expire, a record of nothing indistinguishable from a record of something already fixed. This package shipped a config example that baselined a look. It was corrected by editing the example, and an example is corrected once per copy of it; a command that cannot make the mistake is corrected once.

It also refuses a line nothing printed, because an entry that does not fire is stale the moment it lands. Nothing is typed by hand: the entry is the finding's exact text, taken from the run, which is what makes whole-line matching safe.

Put the line before the flags. --scan takes a list, and a line after it is one more path.

Every table is read in both directions. An exemption naming a file that no longer exists is a finding. A property name that does not exist is a finding. A baseline line that no longer fires is a finding, because someone fixed the problem and left the record claiming otherwise.

A table read only one way rots into decoration: it accumulates entries, none of them ever expire, and after a while it lists things somebody once believed rather than things that are true. The second direction costs about ten lines and is the difference between a suppression file and a record.

reason is required and not decorative: an exemption without one cannot be told from an oversight, and six months later nobody can say which it was. The baseline is the table this matters most for; it accumulates fastest and rots first, because a fixed finding leaves its line behind in silence, which is why the object form asks for one and assay accept will not write an entry without it.

Adopting this on an existing project means starting with a backlog. The two dishonest ways to handle that are a magic threshold (goes stale in silence) and a blanket suppression (hides the next real one). The baseline does neither: a new finding is not in the list so it fails, and one you fixed no longer fires so its line fails as stale.

Staleness is a property of the line, not of the run, and getting that wrong made the tool cry wolf at itself. assay runners cannot produce a finding that only diff reports, so checking staleness there flagged every diff line as fixed: the audit reporting a problem with its own config, on a clean tree, on every run.

The first fix was to check staleness only from assay all. Correct, and blunt enough to be its own problem: every line in every other run went unchecked, and the run printed a disclaimer where a number belongs. An entry that names the command firing it can be answered by that command alone, so each line lands in exactly one of three places:

BASELINE assay.json — 3 accepted, 0 new, 1 stale, 2 NOT checked for staleness
                      (anchors: 1; no `from`, so it needs `assay all`: 1)

Stale is a line this run could have seen fire and did not. Not checked is a line this run could not have seen at all: counted, and never folded into the stale number, because 0 stale from a run that never looked reads as "nothing is stale" and those are different claims. A line with no from keeps the old rule: only a run that performed every audit can call it fixed, since nothing narrower knows what produces it.

Every command still suppresses accepted findings, because that direction is safe from any run: a line that fires is a line that fires.

assay all --scan PATH is the complete run: all alone does not perform the sameness half, and saying otherwise is how a same answer line got called stale on a clean tree. Tag your entries with from and any command answers its own; leave them untagged and the complete run is the only one that can.

--sweep PATH --against BUNDLE folds the cross half in as well, so a same answer across languages line is answerable by the same run:

assay all --base origin/main --scan src --sweep src --against js.json
assay all --base origin/main --scan src --sweep src --against js/src --with assay-js

sweep is a legal from but is deliberately not part of what makes a run complete, and those are two different questions that it would be a defect to conflate in either direction. Add it to completeness and every assay.json written before this release quietly stops having its untagged entries checked: the tool doing less, on configs that were fine yesterday. Refuse it as a tag instead and a cross finding lands untagged, where all --scan (complete by the older definition, having never swept) calls it stale on a clean tree.

The two sets differ by exactly that one name, and it is safe for a reason about time rather than a convention: assay accept always writes from, so every untagged line that can exist was written before sweep was, and no line older than a command was produced by it.

Under Node this used to be impossible, because assay anchors was Python-only and no JavaScript run performed every audit that can produce a baseline line. It does now, so the JavaScript half answers on exactly the same terms.

What same is worth

The first run of the sameness half paired these two:

def is_wordy(tok):  return tok[:1].isalpha() or tok[:1] == "_" or tok[:1].isdigit()
def _word(tok):     return tok[:1].isalnum() or tok[:1] == "_"

They are not the same function (isalnum is a strict superset that also covers numerics), but every character in the ladder made them agree. So three characters went in (½, é, tab+newline), and:

differs  keystrokes.py::is_wordy  vs  pycomplete.py::_word
         ('½',) -> V:False vs V:True

A same became a differs with a witness, from one character. That is what same is worth, and it is why the verdict is worded the way it is.

The ladder carries those characters in both halves, so the same question asked in JavaScript is settled on the same rung. This pair is a demonstration rather than a finding out of somebody's tree, but the run is real:

export function isWordy(tok) { return /[A-Za-z0-9_]/.test(tok[0] ?? ''); }
export function wordish(tok) { return /[\p{L}\p{N}_]/u.test(tok[0] ?? ''); }

ok       differs: w.js::isWordy  vs  w.js::wordish — ["\u00bd"] -> V:false vs V:true

½ is \p{N} and is not in [0-9], which is the whole of the difference between two functions that agree on every ASCII token you would think to try. The general lesson is worth stealing whatever you use to test: ask what characters your inputs never contain, then add them.

The guard the sameness half rests on

Two functions that raise TypeError on every input agree perfectly. So do two that return the same constant. Without a guard, a scan of any codebase reports every one-argument function as everyone else's twin. discriminating() therefore requires at least two distinct returned values and rejects a projection: a function handing back one of its own arguments.

Both halves of that guard exist because both mistakes were made:

• Counting distinct OUTCOMES is not enough. One returned value plus one exception is two distinct outcomes, so a keyword predicate that returns False for every string in the ladder and raises on everything else satisfies it. The counting was rewarding a probe that had found the function's type errors and never reached its behaviour.
• Comparing whole vectors against the identity is not enough. A transform whose vocabulary the ladder lacks is the identity wherever it answers and raises everywhere else, so its vector differs from the projection at exactly the positions where the function refused to run. The question is about the positions where it answered.
• Returning an argument is not the only way to do nothing with it. COPYING it is the same emptiness in another shape. Two unrelated query-param transforms (one renaming keys, one splitting a sort value) agreed on every rung of the ladder, because it holds no key either of them recognises and both degraded to copy the object through. A shallow copy is not behaviour the ladder reached; it is behaviour the ladder missed, so it is rejected alongside the identity.

The general form: a round trip is necessary and not sufficient, because an identity program passes it.

Safety: this executes your code

Stated plainly, because it does.

Python. A single function's source is lifted out with ast and executed alone, so the containing module is never imported. A function is probed only if it is module-level, undecorated, not a method, not a generator, 1–3 arguments, and reaches nothing outside its arguments: no files, no network, no clock, and no randomness (random and time import cleanly and both make an outcome depend on something the ladder does not control, so a differs from either is noise and a same is luck). Free names resolve only from the file's own literal constants, its other gated functions, and a stdlib allowlist. Each probe runs in a subprocess with a per-input SIGALRM and a wall timeout, so an infinite loop is a look, not a hang.

JavaScript, and this is a real difference rather than a detail. A function object only exists once its module has been evaluated, so the JS half loads the module and therefore runs its top-level code. The child answers the parent on fd 3, never on stdout, because a module is free to print at import time and an answer sharing a channel with arbitrary output is an answer that output can destroy: one function per line, as each finishes, so the kill that bounds a non-terminating function costs that function rather than the file. Two compensations: it happens in a child process, and the file's source is gated before it is loaded at all.

The gate is two questions, because it guards two events. May this module be imported? is asked of module-scope code only, top-level statements, initializers, IIFEs, class fields, because that is what running an import actually executes. A new Date() sitting in a function body is no evidence about importing the module, and refusing the file for it took every pure helper beside it down with it: on a barrel of ten Handlebars helpers, two clock-using ones cost the other eight their eligibility.

A body is deferred wherever a body cannot run: under a declaration, bound to a name, or as an object-literal property or shorthand method, which is how CommonJS exposes a barrel. An IIFE is refused in every position: it runs on the way in. So is an accessor: { get x() { … } } runs when the property is read, and enumerating a module's exports reads every one of them.

May this function be called? is the second question, and fn.toString() cannot answer it alone: a free name resolves in a module scope the function's own text cannot see. So the per-function gate runs over its real source (including the declared parameter list, because fn.length stops counting at the first default) and then over everything it reaches: local bindings, followed transitively, and relative imports followed into modules that themselves pass the load gate. A bare package specifier, a core module, an unknown global and a name declared twice are each refused by name. Probing calls the function, so this is the last thing between the ladder and a real side effect on a real path.

The refusal is a chain you can walk back to the code: reaches config, which reaches env, which touches process.

The residue is genuine: a module with an import-time side effect that mentions none of the gated names will still be evaluated. If that is unacceptable for your tree, point the tool at the files you trust rather than at the whole repository.

Two implementations, one contract

anchors is the one command whose halves work by different mechanisms, and the difference is worth knowing before you rely on either. Python lifts the table out with ast and executes nothing. JavaScript has no parser in its standard library and this package has no dependencies, so for a long time the honest answer here was a gap: a regex cannot tell a label from an anchor, and a check reporting confident nonsense about which strings are anchors is worse than no check.

But the JavaScript half already loads modules. A table that is exported is readable as data: a property access, no parser, no dependency, no approximation anywhere in the path:

export const MUTATIONS = [
  ['the negative guard stops firing', "if (x < 0) throw new RangeError('x');", 'if (false) {}'],
];

Two consequences, one in each direction. The harness gets imported, so guard your main() behind the entry-point check every program in this package already carries: a harness that does work at import time will do that work, in a child process that cannot reach your session but on the real tree. In exchange, a computed anchor is simply a string here, where the Python half can only report it as a shape it cannot read.

A harness that exports no table is a look, never a finding: it has not opted into being read this way, and inventing a reading of it is the thing this declines to do.

The JS dead-vs-real detector is textual where Python's reads an AST. The consequence is one-directional and worth knowing: it will not produce a false FINDING, it will miss a real one. If your harnesses are Python, run the Python half over them.

python/tests/test_parity.py asserts the contract rather than trusting it: same property names, same verdict names, same config keys, same subcommands, same ladder version, same thresholds, same baseline families, same probe schema, and, for assay cross, the same cross ladder byte for byte, because two lists that were meant to hold the same values and quietly stopped is the only thing that comparison rests on. Two implementations of one contract is exactly the duplication this tool exists to find, so it is checked.

Also shipped

GitHub Action, one uses: line instead of a run block:

- uses: actions/checkout@v4
  with: { fetch-depth: 0 }   # `diff` needs a base ref; a shallow clone has none
- uses: Megapixel99/assay-checks@v0.6.0
  with:
    command: all       # every audit that can produce a baseline line
    paths: src         # ...and with `all`, paths mean `--scan`: the sameness half
                       # is what makes the run COMPLETE for an untagged entry

# The same action runs the other half. `language` is `python` unless you say otherwise,
# so a JavaScript project has to name it.
- uses: Megapixel99/assay-checks@v0.6.0
  with:
    command: scan
    paths: js/src
    language: node

# ...and across the two, naming no pair. `against` is the far tree; `with` is the far
# binary, which is named rather than guessed — both packages install `assay`, so a
# half that guessed would compare a tree with itself.
- uses: Megapixel99/assay-checks@v0.6.0
  with:
    command: sweep
    paths: src
    against: js/src
    with: assay-js

Running harnesses in CI: check the tree afterwards. assay runners audits the harness, and the harness cannot audit the thing that kills it. A step that runs one under a timeout has to answer for the SIGKILL case itself, and git already knows the answer:

- name: mutation harnesses
  run: |
    rc=0
    for h in $(git ls-files '*mutations*.py'); do
      timeout 1200 python3 "$h" || { echo "FAILED: $h"; rc=1; }
      if ! git diff --quiet; then
        echo "DIRTY: $h left the tree mutated"; git checkout -- .; rc=1
      fi
    done
    exit $rc

Four things in that loop are the point rather than boilerplate. git ls-files, not a list of harnesses: a list is one more table that goes stale, and the harness nobody added to it is the one that has been asleep longest. The dirt check runs after every harness, not once at the end, because a tree checked once names the last harness rather than the one that broke it, and a mutated file left behind means every harness after it scored code nobody wrote. The failure is recorded and the loop continues rather than exiting on the first one: a red run that means one broke and the rest were never asked is the same collapse dead-vs-real exists to prevent, one level up. And rc is what exits, because a loop that reports failures and returns 0 is the sleeping suite this whole half of the tool is about.

Docker, both runtimes, one image, for CI that has neither toolchain, and the one place assay cross needs no arranging, because both binaries are already there:

docker run --rm -v "$PWD:/work" assay runners
docker run --rm -v "$PWD:/work" --entrypoint assay-js assay scan src
docker run --rm -v "$PWD:/work" assay cross src/api.py::shout src/ui.mjs::yell --with assay-js
docker run --rm -v "$PWD:/work" assay sweep src --against js/src --with assay-js

Limits (honest ones)

• same is never equivalence. It is "no input in this ladder told them apart", and the is_wordy/_word pair above is the measured proof that the distance between those two sentences can be one character. A ladder is a sample; only differs is proof.
• Coverage of the sameness half is roughly a tenth of functions, and the census says so. The largest excluded class is zero-arity, which no input ladder can ever tell apart. Every exclusion is printed with its reason and a count.
• It compares functions, not programs. Two programs that are the same function under a mapping of their arguments (a cipher and a special case of a more general one) are out of reach. That needs a declared pairing, and this does not replace one.
• Cross-language discovery runs on a strictly weaker ladder than scan does. assay sweep buckets every function of both trees on the shared ladder, which is the intersection of what the two languages can express, so it discriminates less than either native one, and a same it reports is worth less than a same from scan. Use cross on a pair you already suspect and you get the same ladder; use pair or scan within one language and you get a fuller one. A sweep is where to start looking, not where to stop.
• The cross ladder is a subset, and same there is worth less than same here. It holds no tuple, no set, no undefined, and one number type. Two functions it cannot tell apart may well be told apart by a value only one language has, which is why two references in the same language are refused, with a pointer at pair.
• The ladder is hand-written for three arities. A domain whose inputs are structured (an AST, a socket, a dataframe) gets not discriminated, and correctly so.
• The JavaScript half inherits Node's module resolution, including its version differences. A .js file containing export in a directory with no package.json is a SyntaxError on Node 18 and loads fine on Node 22 (module-syntax detection arrived in between), so the same tree can report different coverage on two runners. Ordinary projects declare "type" and are unaffected; a loose directory of ESM .js files is not, and shows up as could not load in the census rather than as a wrong answer.
• An async function is probed on the value it settles on. async function f(x) { return x * 2; }, function g(x) { return x * 2; } and function h(x) { return Promise.resolve(x * 2); } all answer the same question, and all three are compared as one. A rejection is the same outcome as a throw, by type. async widens what gets executed, and that is worth saying plainly: a service-layer function that awaits a database is a function this tool will call. It was already true that loading a module runs its top-level code, so this is more of the same hazard rather than a new one, but it is more of it, and the answer is unchanged: point the tool at the files you trust. async for and async with are still refused, because both drive an object's protocol methods and the ladder cannot supply one.
• A timeout is an outcome; only a synchronous hang is a look. Python bounds every input with SIGALRM, so a non-terminating input lands in the vector as a raise. JavaScript bounds every awaited rung with a timer racing the promise (the event loop is free while a promise is pending, so that race is a real interrupt), and the rung becomes E:TimeoutError, the same outcome by the same name. A synchronous loop is the one case with nothing to interrupt it: it never yields, so the JS half falls back to a wall clock and a kill, and that function is a look rather than a vector. It costs only itself: the child answers one function at a time, so the kill loses the function that hung and, after it, the ones never started, each named as such in the census, while everything already answered keeps its vector.
• The probe exits when it has answered, rather than when the event loop drains. A module that opens a pool, a socket or an interval at import time keeps its process alive long after the last answer is written, and every such file used to cost the full wall timeout for work that finished in a fraction of a second. This is not an async problem: a file of ordinary synchronous functions pays it too if its module opened something on the way in.
• A --stdin snippet may not import from the tree. Python already refuses one: free names resolve from the snippet's own constants, its own gated functions and the stdlib allowlist, and nothing else. JavaScript refuses one for a different reason worth stating, because a module has to be on disk to be imported at all: outside the root a relative import resolves to nothing, and inside the root the snippet would be scratch state beside the code under test, which is what no-tree-writes audits harnesses for. A snippet that already imports half the tree is a file, so point search at the file.
• A JavaScript snippet that exports nothing needs --name. A module's functions reach the probe through its exports, so an unexported declaration is invisible, and finding it anyway would mean reading declarations out of source with a regex: the same thing anchors declines to do. The name is asked for instead. Python has no such gap: a top-level def is a top-level def.
• assay anchors under Node IMPORTS the harness, and that is a real hazard rather than a detail. The Python half reads the table with ast and executes nothing; the JavaScript half has no parser to read it with, so it reads the table as data, which means the module that builds it has run. It runs in a child process, so a crash or a hang costs the probe rather than the audit, but a harness that mutates the tree at import time mutates the tree, and no child process undoes that. Guard main() behind the entry-point check every program in this package carries. If yours cannot, anchor_exempt is the table for saying so, with the reason.
• The seven properties are about mutation harnesses. If your project has none, that half has nothing to say about it and says so rather than reporting a pass.
• sigterm cannot cover SIGKILL, and a timeout is usually a SIGKILL. The property is satisfied by a harness that turns SIGTERM into an exception, which is the whole of what a process can do about being asked to stop. Being killed is not that: no handler runs and no finally runs, so a harness passing all seven still leaves the tree mutated when the thing invoking it runs out of patience. Only the invoker can check that, and assay does not see the invoker. See Running harnesses in CI.
• targets_mentioned under-reports. A harness that merely mentions a filename counts as covering it. An audit that errs should err toward saying less.
• It is not a code reviewer. It cannot tell you an abstraction is wrong, a name is misleading, or an edge case is unhandled, and it does not decide whether duplication is a defect. Only one flavour of duplication is; telling them apart is a judgment about what two pieces of code are for, which no execution can make. The output says READ them and stops there.

Development

Each half lives in its own directory, and the two are laid out the same way:

python/assay/     the Python package, imported as `assay`, published as assay-checks
python/tests/     its suites, and the mutation runner that audits them
js/src/           the JavaScript half, published as assay-checks on npm
js/test/          its suites

The folder is python/, the import is still assay. They are different names deliberately: the directory sits beside js/ so the two halves are findable in the same shape, while import assay, python3 -m assay and the assay console script are what is already published and do not change. pyproject.toml bridges the two with package-dir, so an installed wheel puts assay at the top level exactly as before.

Working from a checkout rather than an install, python/ is what goes on the path:

python3 python/tests/run_tests.py        # 319 tests, ~5 s
npm test                                 # 302 tests, ~20 s
python3 python/tests/mutations_assay.py  # 193 mutations, both halves
PYTHONPATH=python python3 -m assay scan python/assay   # scanned by its own scanner
PYTHONPATH=python python3 -m assay --root . all --base origin/master --scan python/assay

No install step, no virtualenv, no npm install: both halves are standard library only, and PYTHONPATH=python is the whole of what an install would have done. See CONTRIBUTING.md for the rules this code is built on and what happens when a mutation comes back NOT DETECTED.

The mutation runner carries all seven properties assay runners audits for, and several of its mutations are versions this tool actually shipped, kept as mutations rather than as comments, so a defect fixed once cannot come back quietly.

It breaks both halves. For a while it could mutate Python only, and the gap did not show in its score: it printed a full tally while every guard in js/src had nothing breaking it on purpose. A tally over the half you can reach reads exactly like a tally over the whole thing, which is the defect this package exists to report, so it was pointed at itself. A mutation names a file, the suffix says which half, and that half's suite is the one that has to go red.

One of them is worth naming here because no ordinary test could have caught it: a NUL byte landed where a space belonged inside a template literal. The file displayed correctly, the parser accepted it, and printing the function back showed a space, while the key it built at runtime could never match the key in the table, so the audit went on reporting the finding an exemption had been written to silence. That reads exactly like a config that was never loaded. test_parity.py now checks the bytes.

A mutation that comes back NOT DETECTED is not always a missing test, and it is never left in the table. Three came back that way while 0.3.0 was being written: two were fixtures that could not reach the case, and the third was a guard whose absence produced the same observable as its presence: dead code with a comment explaining what it did. That last one is the defect this package exists to report, arriving inside it, so the branch was deleted and the check moved to the thing it had been restating. A mutation nothing can catch is a table entry claiming a guard is covered when nothing breaks it on purpose.

A note on the name

assay is a common word and the space is not empty. @metahub-ai/assay (Apache-2.0, live) evaluates AI artifacts (skills, MCP servers, agents) with static analysis plus sandboxed behavioural testing, and produces signed reproducible reports. Adjacent territory and a different question: it asks whether somebody else's artifact can be trusted; this asks whether your own checks could have failed and whether your tree already answers what you are about to write.

Bare assay is taken on npm (2013, dormant) and on PyPI, which is why this ships as assay-checks on both. Its global CLI is still assay, and so is theirs: if you install both globally, one shadows the other. Recorded here rather than worked around, so nobody has to rediscover it from a confusing assay --help.

License

MIT