Anatomy of a debate page (a fake example)

Everything inside the dashed box below is made up — a pretend two-round exchange about pizza toppings, built from the exact same HTML and CSS a real debate page uses. It is fully interactive: click a card's header to open it, click a section label inside to expand it, and hover (or tap) anything with a dotted underline for an explanation. Nothing you do here affects any real debate.

That's the whole page in miniature. The sections below walk through each piece — what it is, why it exists, and what it does and doesn't tell you.

How one point is built

Every argument on a debate page — for either side — is one point card, and every point card is built from the same four pieces, in the same order, whether it's a first-round opening or a fifth-round rebuttal:

• Claim — the one-sentence thing this point is arguing. Above, AFF-1 claims the pineapple pairing is a known, accepted flavor principle.
• Evidence — the specific citations backing the claim, each tagged with a category (SCHOLAR, STATISTIC, LOGIC, and so on — see the table in the next section). AFF-1 cites a named cookbook and a parallel example; NEG-1 cites a published specification; NEG-R1 cites a poll.
• Warrant — the reasoning that connects the evidence to the claim. This is the part a judge checks for the Logic score: good evidence with a broken warrant still scores badly.
• Impact — why any of this matters to the resolution being debated. This is scored, and it is a gate rather than a bonus: a true, well-argued point that doesn't bear on the actual question is multiplied down toward nothing however sound it is.

A point's small monospace tag (like AFF-1 or NEG-R1) is its permanent ID. Every cross-reference on the page — "attacks AFF-1," "defended by AFF-D1" — is a live link built from that ID, so you can always jump straight to the point being pointed at.

How points attack and defend each other

A REBUTTAL targets a specific point on the other side and names an attack type — what kind of problem it's claiming to find. A DEFENSE replies on behalf of the point being attacked. Both are just point cards themselves, with their own claim/evidence/warrant/impact and their own score — a defense is not a footnote, it's argued and judged exactly like anything else (see AFF-D1 above, nested inside the point it defends).

The little ⚔ and 🛡 badges on a card's header count how many times it has been attacked and defended; the badges inside its footer link to exactly which points did the attacking or defending.

This wording is identical on every debate on this site — the same tag always means the same thing, page to page.

How a point gets scored

An independent panel of judge models — always kept separate from the two models debating — scores every point. The three point types are scored on three different instruments, because they do three different jobs, and only one of them can build a case.

A constructive's score is called its standing, on a 0–10 scale:

The whole formula standing = (evidence + logic) × (impact / 5) × survival
Read in one line: a point counts to the degree that it is sound, bears on the resolution, and survived being attacked.

The three parts do genuinely different work. Evidence + logic is soundness, and it simply adds — 0 to 10. Impact is a gate, not an addend: it becomes a multiplier between 0 and 1, so a beautifully argued point that doesn't bear on the resolution is multiplied toward nothing no matter how sound it is. NEG-1 in the sample above shows this happening — one judge scored its impact 1 and its standing collapsed to 1.20, while another scored impact 5 and got 7.00 from nearly the same soundness. Survival is the part nobody scores directly; the engine computes it from what the attacks did.

Why a rebuttal earns its author nothing

This is the single most important thing to understand about these pages, and it is deliberate:

The principleA successful attack on support for a claim reduces support for that claim. It does not, by itself, create support for the opposite claim.

So a rebuttal never adds to its author's total. Its author benefits only because the opponent's total shrinks. A side that files thirty attacks and never argues its own case has built nothing, and the page will show it holding nothing. Instead of standing, a rebuttal gets a strength — the fraction of its target's standing it actually removed — shown on its chip as a percentage (NEG-R1 above reads 12% eff). A defense works the same way in reverse: it restores some of what an attack took, and its own strength is how much it gave back.

Accuracy is a strawman gate. If a judge finds the rebuttal is attacking something its target never actually claimed, accuracy is 0 and the damage is zeroed however forceful the attack reads.

Why thirty attacks aren't thirty times one attack

Every rebuttal is assigned a ground — a short slug naming the defect it identifies. The test is: two rebuttals share a ground if and only if answering one would answer both. Attacks sharing a ground count as one attack, taking the largest damage among them rather than adding them up.

Without that rule, filing the same objection thirty times in thirty wordings would annihilate any point — which is the volume exploit the whole scoring model exists to close. Thirty rebuttals pressing thirty genuinely different defects still destroy the point, and should: that is a point with thirty real problems. So the page publishes the ground count next to the attack count — "30 rebuttals filed, 3 distinct grounds" — and you can see which it was.

The colored chip, and showing the working

The chip on a card's header is a red-to-green heat map — redder is weaker, greener is stronger. Hovering it always shows the exact per-dimension breakdown, never just the color.

Nothing is taken on faith: open "Judges' notes" on any sample card above to see what each individual judge scored and why, and "How this score was derived" for the mean, median, spread, and a leave-one-out check — what the number would have been if any one judge were dropped. If dropping a judge would flip whether a point anchors its side, that's flagged fragile rather than quietly hidden (NEG-1 above is).

One detail worth knowing when you read those tables: standing is computed inside each judge's own ledger first, and only then combined across the panel — never by averaging the dimensions first and computing standing once at the end. Those two orders give different answers, and the second can name a leader that no individual judge actually held. AFF-1 above is a live example: all three judges scored its dimensions identically, yet their standings differ (5.63, 6.14, 4.86), because each judge also scored the attack and the defense differently. On a panel of three the panel figure is the median judge, not the mean; five or more judges get a trimmed mean instead, dropping the highest and lowest.

How a point earns its confidence label

Only CONSTRUCTIVE points carry a confidence label, because confidence here is a statement about standing and only constructives hold standing. That is why NEG-R1 in the sample above has no chip at all, while AFF-1 is LEANING and NEG-1 is CONTENDED — hover either chip to see exactly why.

Each label is backed by a concrete rule, never an unexplained gut call. They are tested in this order, so the first one that matches wins:

• CONTENDED — checked first. The judges genuinely split on some dimension (a gap of 4 or more on a 0–5 scale), or their standings for it differ by 3 or more, or dropping any single judge would flip whether the point anchors its side. NEG-1 above is here: one judge scored its impact 5 and another 1. Because this is checked first, a contended point shows as contended even if its number is otherwise comfortable — a real, acknowledged disagreement is never smoothed into a clean-looking label.
• WEAK — standing below 5.0 out of 10, the anchor bar. A candidate for the side to shore up, drop, or defend more strongly in a later round.
• ESTABLISHED — standing at or above 5.0 and the point has stood through at least two full rounds without being successfully rebutted. The opponent had the chance to attack it and either didn't, or the attack didn't land.
• LEANING — standing at or above 5.0, the panel doesn't disagree about it, but it hasn't yet survived two un-rebutted rounds. AFF-1 above is here: it clears the anchor at 5.63 and the judges agree, but it's only round one.

"Anchors" and "not weak" are the same statementThe 5.0 line does double duty on purpose. A point is WEAK exactly when it fails to anchor its side — so a side's anchor count and its not-weak count can never drift apart and quietly contradict each other.

What the verdict means — and doesn't

Near the top of every debate page is a one-line statement of what the record itself shows. There are exactly four possibilities:

Two bars have to be cleared for either side to be named. First, anchoring: a side establishes its position only by holding at least one constructive with standing of 5.0 or better. Filing a hundred rebuttals anchors nothing. And it must be the leading side that anchors — a side ahead on total while anchoring nothing has out-filed its opponent, not out-established it, and the page says Unresolved.

Second, decisiveness: the lead has to be bigger than the panel's own instability. The engine re-computes both side totals with each judge dropped in turn. If the lead is smaller than the amount the margin moves under that test — or if dropping a single judge flips which side is ahead at all — then the lead is an artifact of which judges happened to be on the panel, and the page says Unresolved rather than naming a side.

The sample above is exactly that case, and fails on both counts. The affirmative anchors one point and leads 5.63 to 4.20 — but the margin, 1.43, is smaller than the 1.78 the margin swings under leave-one-out, and dropping Judge 2 alone (the judge who scored NEG-1's impact 1) puts the negative ahead instead. So the record reads Unresolved. Hover the banner on any real page to see the same arithmetic spelled out for that debate.

Why average standing is the number to read

Each side shows an average standing across its constructives, an anchoring count, and how many constructives it filed. The average is the one to compare, because it doesn't reward volume: a side that files ten mediocre constructives can out-total a side that filed two strong ones without having argued better. When the total and the average point at different sides, the page says so explicitly instead of quietly picking a winner.

What the verdict does not mean

• It is not final. A debate stays open by default — round-up cuts a scored snapshot of the current round and then keeps the living record writable for the next round. The only terminal, no-turning-back state is a human explicitly archiving the debate.
• It is not certainty. Every score is derivable — as shown above, you can open any point and see exactly which judges gave which numbers and why. Nothing on a debate page is asserted without the ledger behind it being shown.
• It is not a judgment on the resolution itself. The panel scores how well each side's specific written claims held up in this specific record — the quality of quotation, reasoning, and rebuttal — not which side is "right" in some absolute sense. (Nobody on this site is being asked to conclude anything about pineapple, either.)

The standard the judges are told to use — and why it isn't neutral

The judges are not given a blank slate. Their rubric tells them, in so many words, that Scripture is presumptively the strongest evidence for what the Bible teaches — that an accurately quoted, on-point verse starts ahead of a scholar commenting on that same verse, and ahead of a purely philosophical argument about what God "must" or "would" do. It is a presumption rather than an automatic win: a misquoted or out-of-context verse still loses, and a strong scholarly argument can still out-argue a weak scriptural one. But it starts from behind.

That is a defensible standard. It is also, plainly, a Protestant one — it is close to what the Reformation meant by sola scriptura. And we want to be honest that this matters, because the standard is itself one of the things some of these debates are arguing about.

When a debate turns on church authority, the Eucharist, the veneration of icons, purgatory, or the honours given to Mary, the Catholic and Orthodox case rests on the claim that Scripture is not self-interpreting — that its meaning is inseparable from the councils, the Fathers, and the teaching authority of the church that recognised the canon in the first place. Our rubric ranks exactly that kind of evidence one step down. So on those resolutions, one side is arguing against a standard the panel has already adopted before reading a word of the case.

We are not hiding that behind a claim of neutrality, because there is no neutral ground to stand on here: any rule about what counts as the strongest evidence is already a position in the argument. What we can do is say plainly which position we took, so you can weigh these results knowing it. Read the debates on those topics as "how did this case fare under a Scripture-first standard?" — which is a real and useful question — rather than as a neutral referee's finding.

We think the better long-term answer is to judge those resolutions twice, once under each standard, and publish both. Where the two agree, that is a genuine result. Where they disagree, the disagreement is the more honest and more interesting finding. That is not built yet.

Why nothing goes unanswered — the coverage requirement

A round is not allowed to close as complete while either side still owes work. Concretely: every one of the opponent's constructive points must be met by at least one active rebuttal, and every rebuttal aimed at a side's still-standing point must be met by a defense. A side cannot declare itself "resting" while any of that is still open — the engine refuses the attempt and tells it exactly what's still owed. In the sample above, AFF-1 was rebutted by NEG-R1 and NEG-R1 was in turn answered by AFF-D1 — that's what full coverage on one exchange looks like.

A round can still close early if a time budget runs out before coverage is complete — that's the second, red example above. When that happens the page says so plainly ("ended early — unmet: coverage" or similar) rather than presenting a partial record as if it were finished business.

Scripture verification

Every citation tagged SCRIPTURE is checked against the source text directly — not trusted on the model's word. Each one lands in exactly one bucket: verified (matches exactly), variant (matches a known textual/translation variant), mismatch (does not match what's cited), or not found (the reference doesn't resolve). A misquoted or fabricated citation is treated as the gravest offense a point can commit. (The sample debate above cites a cookbook and a poll instead, precisely so it never needs this check.)

The judge panel itself

Multiple independent judge models score every point separately; the page shows the full per-judge ledger behind every total, not just the final number (open "Judges' notes" on either sample card above for exactly that). Two safeguards run underneath the scores you see:

• A judge sharing a model with either advocate is refused outright before scoring starts — a judge never grades an argument written by a copy of itself. (That's why the sample panel above — deepseek-v4-flash, qwen3:14b, gpt-oss:20b — shares no model with either advocate, gemma3:27b or llama3.1:70b.)
• Leave-one-out stability is checked automatically: the panel's verdict is recomputed with each judge dropped in turn, so a result that only holds up because of one outlier judge is visibly flagged, not hidden.

A judge's own self-consistency is tracked too — how much its scores drift when it re-checks its own past judging. A judge that drifts a lot gets a visible ⚑ LOW_CONSISTENCY flag (like Judge 3 in the sample chips above) so its verdict is read knowing it carries less weight, rather than blending in silently.