Synthetic Beef

Public production record

Whole-run production methodology

Version: 1.0 (whole-run-v1)

Effective date: 29 August 2026

Editorial/audio implementation addendum: 5 September 2026; the whole-run-v1 protocol is unchanged.

Validation safeguards update: 6 September 2026. Named attribution sentences may be retained unchanged or deleted as dispensable recap, not reassigned. Post-production bridges are limited to reading handoffs to the actual next recorded speaker; they cannot introduce substantive questions. Release checks bind caption boundaries to the archived audio-unit timing record with a 10 ms rounding tolerance.

Owner: Kua SpA editorial team

Canonical URL: https://syntheticbeef.lab.kua.cl/legal/methodology

The central promise

The published panellist conversation comes from exactly one complete generation run. Humans do not assemble an episode by choosing preferred answers from multiple attempts. If Kua SpA declines a run, the next attempt begins the entire episode again and none of the rejected dialogue is carried forward.

The moderator's short version is:

The conversation you are about to hear comes from exactly one complete generation run,

which is kind of like the AI equivalent of live television.

That is an analogy about editorial selection, not a claim that the finished audio was broadcast live or recorded continuously in a physical room.

Before the run

Humans design and freeze the programme format, episode topic and questions, model/runtime cast, moderator direction, participation mechanics, output budgets, deterministic checks and publication criteria. These choices are editorial direction. They are not model answers.

There is no human control channel during a whole-run-v1 recording. Ideas for the moderator must be incorporated and frozen before the run begins.

During the run

Editorial direction and reply opportunities

An episode may freeze a private editorial brief before generation: a few destinations, written as claims to test rather than categories to cover. The live Moderator receives the brief as private text at each round boundary and decides for itself what the room has argued; nothing scores topics by keyword and nothing tracks coverage on its behalf. One round boundary per debate, chosen by recorded chance and never the last, invites the Moderator to ask the question a listener would ask rather than a debate question; it may decline.

The Moderator may let a productive exchange continue, clarify an obstructing ambiguity, ask one targeted follow-up, return to unfinished business or close. Code reserves the next reply opportunity before a named question is spoken. A reply opportunity does not certify a satisfactory answer. The same originating question receives at most one targeted follow-up, and no reply can be promised past the round ceiling. Invalid or repeated direction proposals are withheld before becoming dialogue; their raw outputs and decision warnings remain recorded. This does not authorise modification of panellist responses or extend the round budget. The Moderator may not put a third consecutive question to the same seat, and a second consecutive question must tell the other panellists why it is staying with that seat. The rule is enforced in code: the referee is shown its own tally of questioned seats at every round boundary, gets one re-ask when a verdict breaks the rule, and otherwise the question is dropped and the round continues; each outcome is recorded on the round decision. At the first Moderator decision point after two-thirds of the episode's declared length has been recorded, the live Moderator is asked once, privately, to name the weakest thing in the show so far and to decide one thing it will do differently in the last third. The decision is recorded in the run and attached, as the Moderator's own report, to every later host and referee direction. It is guidance for the last third, not an instruction for the next boundary: the Moderator does not interrupt a live exchange or change the subject for it, and may voice it in one sentence about the room at a natural seam, never as a review of the show.

Generation record

The recorder preserves the exact inputs, projected transcript, provider responses, model and runtime identifiers, timestamps, retries and hashes. Each non-empty successful panellist call must appear exactly once in the recorded conversation, with one predeclared exception: at most once in a live debate, the model-generated moderator may cut a panellist turn at a recorded complete-sentence boundary. No panellist words are rewritten. The complete response is archived; only its exact prefix enters the conversation seen by later models. The omitted continuation is published with the episode as "generated but not aired; not seen by any later panellist." A run may use no interruption. Chance may decide among predeclared eligible speakers or readers; those rolls are recorded.

Moderator lines containing a prohibited deferred-payoff phrase are rejected before entering the conversation and regenerated once. Both provider attempts remain in the private event record; no panellist wording is affected.

Predictions must contain the speaker's own answer, a mechanism and a way to score the claim. That requirement is given to the isolated writer before the card exists. Since 11 September 2026 (prediction-warning-v2) nothing repairs a card in the run: a scoreable card that still lacks one of those elements airs exactly as written, the defect is recorded as a QA warning, the public audit manifest commits to the prediction policy, the hashed scoreable questions and the warning count, and the room, not a deterministic callout, may challenge it live. Final QA and the release audit derive whether the prediction contract applies from the attributed card kind, its source question and the frozen episode specification, not from a mutable flag on the turn, and any trace of the retired in-run repair vocabulary on a run recorded under the current policy blocks release. Runs recorded under the earlier prediction-repair-v1 policy keep the validation they were recorded with: the recomputed defect codes, the reconstructed deterministic callout and its causal links.

Optional reaction calls may pass without a spoken turn. OpenAI-compatible and Codex calls use a plain transport because those paths do not enforce the supplied schema: the archived request asks for only the spoken words, or exactly [] for a silent pass, and does not enable unconstrained JSON-object mode. Anthropic and Claude CLI retain their enforced schema. Before any turn exists, a closed deterministic decoder accepts plain speech, [], a JSON string, the sole line or observed reaction field, or an exact assistant-message envelope containing only type, role and string content. It extracts only the unmodified final string and records a release-replayable transform; role/type wrappers never become dialogue. Empty forms require a matching hash-linked pass event. Unknown, mixed, malformed, non-text and scratchpad-shaped optional-reaction payloads are archived and deterministically omitted as unavailable; they never become dialogue and do not destroy the required conversation. Requested lengths are prompt-level style budgets, not destructive transport ceilings. Required OpenAI-compatible speech receives a larger transport safety ceiling so a complete answer can stop naturally. If a provider nevertheless reports truncation, the take is rejected rather than airing the fragment. Within a question, cards are read in an order shuffled before the segment. After the first card, a fixed producer-authored handoff names each next reader in three words; it is not model-written and the showrunner may not edit it. The seat that reads next does not react to the card immediately before its own, so a live reply and a sealed card are never heard as one speech. A plain-language gate applies to sealed answers only, before anyone has heard them. If a sealed draft's vocabulary sits outside common English past two thresholds frozen with the run (30% of words outside the 3,000 most frequent and 25% outside the 5,000), the same seat is asked once, with its own draft in front of it, to say the same thing in plainer words without adding or dropping a claim. The draft is archived, hash-linked to the restatement, and published with the episode; the restatement airs. Nothing chooses between them by content, and a restatement that still crosses the thresholds airs as written. Live turns are never sent back: the same meters are computed on every take and recorded as warnings, alongside a measure of how often the panel speaks in "not X, it's Y" epigrams and how many live reactions push back without agreeing, conceding or asking, so the register can be judged against real conversation rather than by ear. An episode may freeze a collision round: after a stated sealed question, one live round on the cards read so far, with the ordinary reaction roll and no referee, before the next card. The live Moderator opens it by stating the actual split in what was written, from the full transcript it already holds; no human chooses which cards collide, and the round's position, length and settings are frozen with the run before the first sealed answer is generated. The debate host and the seats also receive, as a briefing, the same verified launch history and press research the production checked before the run; any topic research is validated for a direct source URL before it reaches the room. An optional reaction that the provider explicitly reports as token-truncated is unavailable rather than dialogue: the raw incomplete attempt is retained privately, a hash-linked event records the provider finish reason, and none of its partial text is used. This transport rule is fixed in advance and applies only to optional reactions. A truncated required answer still rejects the run; producers do not inspect an incomplete reaction and choose whether it was good enough to keep.

The recurring Cards Against Artificial Intelligence segment uses anonymous answers and asks the panel to guess authorship. It is a segment of Synthetic Beef, not a separate card game. Its answers remain part of the same run and follow the same no-cherry-picking rule. Episode specifications exclude claim-review questions and questions requiring the scoreable-prediction contract from anonymous play; those cards use the attributed read-through so any defect can be challenged in the room.

After the run

Humans decide whether the complete run is publishable. They may reject it for safety, legal, editorial, technical or quality reasons. They may not rescue it by:

The predeclared in-run interruption is not a human post-production cut and is not selected after the episode is complete. A model judge may spend a hard budget of one only for unearned repetition, already-answered repetition or a materially overlong non-answer—never for viewpoint, politics, offence, provider identity or conclusion. The cut and moderator line are committed before the next provider call, so every later seat receives the interrupted conversation.

An automated showrunner makes the generated run recordable. For panellists, it may only remove silent Markdown, redundant whitespace and short delimited non-spoken visual directions from a closed physical-action list, such as (leans back, arms crossed); their spoken wording remains unchanged and the raw response remains preserved. Unknown or substantive parentheticals are kept and flagged. Every removed byte, offset and rule is logged. Audible tags such as [laughs] remain available to the speech synthesizer. The exact string-valued line field may also be extracted from a structured response.

The same disclosed showrunner reads the full episode but may contextually tighten only one specified moderator turn at a time and insert an essential one-to-eight-word moderator bridge—for example, between a reaction and the same speaker's prepared answer. It may not delete, move, complete, paraphrase or rewrite a panellist turn. It is an editor, not a fact-checker: it may not remove, soften or flag any line because a claim in it cannot be verified, and it may not tighten a moderator turn for that reason. A mistake made on air stays on air and is handled afterwards by the published corrections process. The exact versioned prompts, schemas, hashes and change notes are published with the episode. Humans do not line-edit the dialogue after seeing it. Each seat is asked to name the model it is running, in its introduction; if it does not, the Moderator states it in a fixed producer-authored line, so the lineup is always spoken on air and never left to the transcript alone. Every episode ends with a fixed, producer-authored closing disclosure appended after the Moderator's own words: that the show is entertainment, independent of the companies behind the models and endorsed by none of them; that every model on the panel and the show itself can make mistakes, with corrections on the public record; and where the methodology is published. It is never generated and the showrunner may not edit it. Required opening disclosures, the fixed closing disclosure, in-run factual corrections, the triggering moderator statement, a deterministic failed-repair acknowledgement (only in runs recorded under the retired prediction-repair-v1 policy) and a live-interruption line are causally locked: the showrunner cannot rewrite words that later models were specifically shown as a correction or intervention.

Human-written questions and any producer-authored disclosure or wrapper are labelled as such. A correction or right-to-comment response published later is separate from the generated conversation and visibly attributed.

Voices and assembly

Consecutive same-speaker fragments are combined into one continuous synthesis request, not generated separately and stitched into a paragraph. The source transcript records remain separate and are mapped to exact character spans in the combined audio unit. Another speaker, an explicit production pause or a timed music cue creates a boundary. Caption timing is recorded per audio unit; internal source-turn timings are not invented. Voice identity and native speed settings are unchanged by grouping, and the finished waveform is not time-compressed to force the words into a slot.

Editorial warnings such as potentially redundant transitions are distinguished from integrity failures. A disputed seam receives at most one targeted contextual review; an uncertain edit is not forced. All accepted proposals, refusals and unapproved checkpoints are retained. Approval requires intact panellist wording, locked text, attribution and mandatory disclosures, not stylistic perfection.

Models return text. Synthetic cast voices read that text in post-production; they are not the models' or companies' voices. The final mix may add pauses, music, loudness treatment, captions, animation and mastering. It may not substitute panellist dialogue from another run or change the generated turn order.

Evidence and limits

The private run record is hash-chained and retained, including rejected and interrupted runs. Every showrunner and audio attempt has its own immutable evidence chain; retrying creates a new attempt rather than replacing the failed one. A content-blind production timeline and process video may show the sequence, roles and durations of production events without exposing prompts, answers, credentials or private provider identifiers. The public episode manifest commits to the selected run, transcript and released media without publishing credentials or sensitive raw records. Hashing detects later alteration but does not by itself prove when material was created; separate trusted timestamps or immutable-storage receipts address that limitation.

The seed reproduces orchestration choices only when the same code consumes chance in the same order. It does not reproduce provider prose, routing, hidden vendor instructions or remote infrastructure. The show is entertainment and an inspectable editorial experiment—not a scientific benchmark and not evidence of a company's policy or a model's stable belief.