The proposed fixture scorer accepts any non-empty override token as a human approval

The scorer fragment decides pass or fail on one line: `ok = not missing and (not changed or bool(override))`, with `override` read from the environment variable the lock file names ^[14]. The prose above it says a changed digest requires `FIXTURE_OVERRIDE` to match a value injected by a human job, not by the agent workspace ^[13]. `bool()` compares nothing. Any non-empty string satisfies it ^[15]. The human-only property is not in the scorer at all; it lives wherever that variable gets set, and the published check would accept `FIXTURE_OVERRIDE=1`.

The replay driver has the same gap running the other way. Lowering `max_examples`, tightening a strategy or filtering the input that used to fail all leave pytest exiting 0 ^[17]. Its stated rule is that the job fails when `min_examples` falls ^[18], and the example log sets that floor at 200 for `invoice_total_non_negative` ^[19]. The code records a floor failure only when the value is less than 1 ^[20]. Rewrite 200 to 1 and the driver passes, because 1 < 1 is false ^[21]. Enforcing the written rule needs a comparison against the log on main. The fixture check already does that for new paths, allowing them only when main's lock does not name them, and recording them as added ^[12].

In the example lock, the `sha256` field is 32 hex characters, half the 64 a hexdigest returns ^[22]. It is an illustration, and it would not match anything the scorer computes.

The sequencing is the part I would take first. Digest before pytest, because a dirty lock makes every green result after it untrusted ^[10]. The case the post gives for locking bytes is that committed testdata is an API and changing bytes is a behavior change even when the assertions move in the same diff ^[7], while unit tests encode today's examples and not the distribution those examples came from, so coverage can rise as the awkward case leaves the repo ^[5]. Digests cover raw bytes with no pretty-printing first ^[11]. A formatter run over a golden JSON file therefore flips the digest, which the post calls intentional, putting canonicalization in the writer ^[16]. Job strictness is set the same way: a skip is a fail, and a missing file is a fail ^[6].

Running it costs one pytest process per recorded seed ^[20]. The example property carries three seeds, so a single invariant spends three interpreter startups per diff ^[23]. The post labels the `--hypothesis-seed` flag environment-specific and says the contract is the log, not the library ^[24].

Two of the three contracts have file formats here. The failure-signature freeze is named as the third and the text breaks off in the seed section before specifying it ^[3][25]. The author is explicit that the scripts are a proposed harness and not production measurements ^[4], and for a design sketch the weak point is enforcement: the code is looser than the rules written above it ^[15][21].