The charge must fit the evidence
Everything below concerns the Engine's internal network-streaming work, which has not shipped in any release. No listener has run any of it. Every claim is drawn from source (source-only throughout), and the refusal copy quoted is copy nobody outside the machine has seen.
When the Engine opens a network stream, it audits the server's account of the bytes before it will play them. Three witnesses sit at that door, consulted in sequence. The header witness compares the type the server's catalog advertised against the type its delivery header actually declared. The decode witness goes deeper: it asks the decoder what the bytes turned out to be, and compares that against the family the advertised type belongs to. And the numeric witness compares two figures the catalog attested — bit depth and sample rate — against the figures the decoder reports. A confident contradiction at any of the three refuses the stream outright. There is no silent downgrade; that refusal doctrine is settled law here, and I have written about it before.
What I had not seen until I read this seam whole is the subtler discipline layered on top. The machinery does not just grade when it may refuse. It grades, separately and explicitly, what each refusal is allowed to claim about the cause — and the two are priced by different evidence. Laid out in order, it forms a ladder with four rungs.
Agreement proves nothing
Start at the bottom: when the catalog's figures and the decoder's figures agree, that agreement earns the stream nothing. The numeric witness is bound by a written invariant to be an additional refusal trigger only — never an independent claim of purity. Agreement between two attestations does not grant the bit-perfect claim; the byte-level witness remains that claim's sole arbiter. This witness can indict. It is structurally forbidden to exonerate.
A lone mismatch is admitted — and inscribed
One rung up: suppose the catalog's bit depth contradicts the decoder's, but the sample rate agrees, or was never attested. The stream plays. The refusal is corroboration-gated, by ruling: the numeric witness refuses only when both dimensions confidently mismatch. The reasoning is written in the source, and it is evidence-quality reasoning, not leniency: catalog numeric attestation is called "the least reliable evidence class in the field" — a server that systematically misreports one figure would otherwise refuse a listener's entire library with no override — while a real transformation moves both dimensions, or moves the codec, which the other witnesses catch on their own.
But the tolerated doubt is never swallowed. The lone contradiction is written to the diagnostic channel with a description ending "admitted uncorroborated," so that the admitted-with-contradiction cohort stays measurable, and a future field report — my server's depths are all wrong but tracks play — maps to evidence already on the record. Admission and refusal are not the only two verdicts; there is a third, admitted under protest, and it leaves a paper trail.
A double mismatch refuses — but may not say "converted"
Now the rung this entry exists for. When both figures confidently contradict the decoder, the stream is refused with exactly the same force as any other fidelity refusal. And a recent cure attached a flag to that specific refusal whose only effect is to change the sentence shown to the listener. The comment at the site states why: the numeric evidence proves a catalog/delivery mismatch, not a conversion — "a stale catalog entry yields the identical evidence." A server that re-encoded the stream on the fly and a server whose database simply holds an outdated row about an untouched file produce indistinguishable numbers. So the refusal copy for this leg is forbidden from accusing the server of converting anything. It says, instead: "The stream delivered does not match what the server's catalog declares for it" — and its recourse line includes the remedy for the innocent explanation, a rescan of the server's library, alongside the remedies for the guilty one.
The type-witnessed refusals keep the accusation, and the same comment says why they may: "a codec change cannot be a stale catalog." No bookkeeping error turns one format into another; only a transformer does. Their copy names the actor plainly — "This server is converting the stream" — because there, and only there, the evidence can carry that charge.
Even the strongest charge minds its tense
One more bound, recorded as a ruling at the same cure: the conversion copy says the Engine "will not play" a converted stream — deliberately never "never plays." The comment calls the first a commitment about witnessed cases and the second a totality claim the witness cannot honestly make: a server that lies consistently — same false story in the catalog, the header, and the bytes' own container — passes a consistency witness by construction, and the campaign's records name that residual class rather than papering over it. Even the sentence of conviction is scoped to what was actually seen.
Opinion, plainly marked. A refusal is two speech acts wearing one dialog: a verdict — this will not play — and an indictment — here is who did what. Nearly all software collapses them. If the machine is sure enough to stop, it considers itself sure enough to accuse, and the error message blames whatever the nearest suspect is; every user has been told their password was wrong by a system whose proxy was eating the credentials. What this seam does, and what I have not seen a codebase do this explicitly before, is price the two acts separately: stopping requires confident contradiction, but naming a cause requires evidence that can discriminate between causes — and where the numbers cannot tell a transcoder from a stale database row, the machine convicts the mismatch and declines to sentence the server. The family resemblance is worth stating. This codebase already holds that absence may not accuse, and that the absence of a rejection may not confirm. This is the third leg of the same law, applied at the moment of conviction itself: the verdict may rest on any confident contradiction, but the charge must fit the evidence.
Nothing above is in any listener's hands. The network-streaming surface, its witnesses, and every line of refusal copy quoted here are internal work in progress, described as they stand in source today.