report
What a run leaves behind.
A trial executes once and then it is gone. The report preserves which trial ran, the revisions and invocation it stood on, where it was written, and what became of its attempt. A replay capsule is separately earned evidence that binds a reduced reproducing input to one recorded execution. An ordinary report carries no replay capsule, and neither value claims what only the other established.
Everything here is a value. Nothing in this home runs a trial, reads a clock, or touches a file. The runner hands this home facts, and this home returns records that callers may compare, count, retain, or replay through the separately earned road.
Meaning and location
A trial has a semantic identity and a source site, and the two answer different questions.
The semantic identity names what the trial means through its declared claim, subject, check, population, and profile. Moving or renaming the code that hosts a trial leaves that meaning unchanged.
The source site names where the trial was written so a person can filter it and navigate to it. A report joins both rails without letting either stand in for the other.
Execution standing
An execution key names one execution of one semantic trial under its exact revision, invocation, target, and toolchain standing. The target and toolchain are always part of that standing, because evidence reached on one host configuration cannot authorize a cache claim about another.
The authored row has a separate revision identity for census and aggregation. Editing bookkeeping around a trial moves the row revision without claiming that its executable meaning moved.
Reproduction authority
A report states the subject and check revisions its execution used. The weaker revision posture limits both cache eligibility and what a reproduction may claim. No caller can strengthen that ceiling by attaching a stronger label beside the recorded revisions.
A replay capsule can be minted only from completed reduction evidence bound to a real refused report. It preserves the execution standing, reduced input, failure identity, generation and minimization authority, generated-support schema, and the replay ceiling earned across every participant that actually ran. Holding a capsule is therefore holding a bounded reproduction account rather than an unqualified promise of exact replay.
Complete accounting
A run report is stated over the complete table view the invocation stood on, including rows the selection did not name. Selection narrows execution and never narrows the denominator. That is what makes exercise coverage a computation over recorded evidence instead of a hand count, and what makes a shrinking population visible.
A selection that matched nothing is still a complete report. The report preserves whether that empty result contradicted the caller’s declared expectation or was admitted for a typed reason, and no empty result is described as a passing trial.
Comparison and coverage
Comparison is a pure reading of two retained reports. It keeps population movement apart from execution-standing movement, refuses reports whose table postures cannot lawfully be compared, and keeps a missing baseline distinct from an unchanged run.
Coverage is a pure reading of one authored report’s complete denominator. It counts exercise rather than correctness, so a trial that ran and refused is exercised while a selected trial that never ran is not.
Neither reading observes the host or reconstructs facts absent from its input.
Foreign material
Text arriving from a subject, decoder, or external tool is bounded and carries whether bytes were cut or cannot render exactly as text. It travels one way. Nothing in the harness reads it back to decide an identity, conclusion, or comparison.
What this home does not do
It does not execute trials, choose selections, judge subjects, persist records, inspect the host, or infer a target, toolchain, revision, or clock reading. Those facts arrive through their owning values and remain under the authority that established them. It also does not interpret a caller’s finding cause; the caller’s declared cause spelling is stored and returned unread.