adversarial-debate

adversarial-debate

Run an adversarial debate over a knowledge base. Debaters reference the KB by argument ID and source path; they cannot introduce un-traced claims. The orchestrator mediates turn-taking — true interleaved back-and-forth, not parallel monologues. The debate may mutate the live KB, but never below a hard discipline (downgrade only, log every change, never delete).

Inputs

• A run directory (e.g. ./research-runs/<topic-slug>/) containing a stable kb/ produced by knowledge-base-construction.

If kb/ doesn't exist or kb/quality-summary.md doesn't say the graph reached a stable state, stop and tell the caller to run KB construction first.

Pre-flight: snapshot the KB

Before any other action, check whether <run-dir>/kb-snapshot-construction/ already exists. The /debate command's own pre-flight normally creates it (and handles the ask-the-user flow when one exists from a prior debate) — if it exists, do NOT re-copy; treat the existing snapshot as authoritative and move on. Only if it is missing (the skill was invoked outside /debate), copy <run-dir>/kb/ recursively to <run-dir>/kb-snapshot-construction/ (e.g. cp -r). This snapshot is the immutable record of what the KB looked like at the moment the debate began. Do not write to kb-snapshot-construction/ again for any reason.

The live kb/ directory is what the debate may mutate.

Then create:

• <run-dir>/output/ (top-level debate artifacts)
• <run-dir>/output/agents/ (per-agent transcripts; populated when agents are spawned)
• <run-dir>/kb/sources/agents/ (per-agent source dirs for new sources fetched during debate)
• <run-dir>/kb/revisions.md (mutation log; starts empty, see format below)

Orchestrator role

You — the orchestrator — own the debate's structure. You spawn debaters and the moderator one turn at a time, hand each one the running transcript plus the specific prompt for its turn, capture its response, decide whose turn is next, and apply any KB mutations that result. The debaters and moderator are stateless across spawns — every turn ships them the context they need.

This means the debate is interleaved, not round-robin: B responds to A's specific claim from the previous turn, then A counters B's specific response, etc. You decide who speaks next based on which open thread most needs progressing.

Phases

Phase A: Position identification + opening claims

• Read kb/quality-summary.md and the argument index. Identify the major positions present in the graph (typically 2, but the graph may show more).
• For each position, spawn a debater subagent. All instances share the static debater agent definition, so assign each one a unique handle of the form debater-<position-slug> (position lowercased, non-alphanumeric → -). The handle names the instance's transcript dir (output/agents/<handle>/) and source dir (kb/sources/agents/<handle>/); without it, parallel debaters would collide on the same paths. In the spawn prompt include:
  • The assigned handle (state it explicitly: "Your handle is debater-<position-slug>")
  • The assigned position (clear, specific, in the form "Pro: " or "Con: " — concrete enough to be argued, not just a topic label)
  • The run directory path
  • The KB layout (paths to kb/arguments/index.md, kb/edges.md, kb/contradictions.md, kb/quality-summary.md)
  • Explicit instruction: each opening claim must reference a specific arg-ID from the KB. Aim for 3–6 opening claims per position. Acknowledge the rated status of each cited argument (don't pretend a weak or baseless argument is robust).
• Spawn one moderator subagent. In the spawn prompt include the run dir, the KB layout, and the list of debater positions.
• Capture each agent's opening to output/transcript.md and to its per-agent transcript at output/agents/<handle>/transcript.md.

After openings, build a claim ledger at output/claims-ledger.md (format below). Every claim raised in the openings becomes a ledger entry with status open.

Phase B: Cross-examination (the core of the debate)

This is interleaved and orchestrator-driven. Loop:

1. Pick the next claim to advance. Prefer claims that are:
   • On the boundary between positions (where rebuttal would actually move the debate)
   • Currently open or partially-rebutted in the ledger
   • Resting on contested or weak KB arguments (where traceback may shift the rating)
2. Spawn the opposing debater (or the moderator, if the issue is structural). Their prompt:
   • Latest full transcript
   • The specific claim/turn they're responding to (quoted, not summarized)
   • The relevant arg-IDs and source paths from the KB
   • Instruction: respond to this claim. Cite by arg-ID + source path. May propose KB mutations.
3. Capture the response. Update the claim ledger (open → partially-rebutted | rebutted | source-debunked | conceded | closed). If the response proposes a KB mutation, apply it (Phase D rules).
4. Repeat. The loop runs until the ledger has no high-leverage open claims left, or the participants have exhausted distinct lines of attack.

The moderator can be invoked at any point — proactively by the orchestrator to press a thin claim, or reactively when a debater has just done something the moderator should call out (overreach, secondhand citation, citing a baseless argument as if robust, ignoring a structural contradiction surfaced in the KB, re-raising a closed claim).

Deflection callouts (anti-tangential-argument rule)

A claim can be technically true and well-sourced yet irrelevant to the actual question being debated. Common patterns:

• A policy/regulatory argument ("agency X recommends Y") used to dismiss a methodological argument about the underlying trial design
• A finding from one population/scope (e.g. trial enrolling subjects aged 50–69) used to make claims about a different population/scope (e.g. recommending the intervention for subjects in their 40s) without acknowledging the scope mismatch
• A peripheral but true claim cited repeatedly to avoid engaging the central question of the topic

The KB's tangential-to, scope-mismatch-with, and layer-shift-from edges already record these issues at the graph level. During debate, the moderator's job includes calling them out as moves when one side is using them as deflection.

Moderator deflection callout — format:

Deflection: Position X has cited arg-007 (layer: empirical, scope: <scope-descriptor>, status: robust)
across turns N, M, P to address arg-022 (layer: ethical, scope: <scope-descriptor>).
arg-007 is well-supported within its scope but layer-shifts/scope-mismatches arg-022.
Position X has not engaged arg-022 at its actual layer.

A deflection callout is not a closure — the cited argument may be entirely true. The callout records that the application is invalid and that a central argument has gone unaddressed. Deflection callouts feed into the synthesis verdict's engagement assessment.

If a deflection callout is reaffirmed across multiple turns and the deflecting side never engages the actual argument at its actual layer, the synthesis records this as a persistent non-engagement finding for that position. A position with strong but layer-shifting arguments that does not engage central questions is not better-grounded than a position with weaker arguments that does engage.

Claim closure (anti-zombie-argument rule)

When a claim has been decisively refuted via traceback during the debate (not just challenged — refuted, with the moderator's concurrence and a corresponding KB downgrade to refuted or baseless), the orchestrator marks it closed in the ledger.

Effect of closure:

• A debater may acknowledge a closed claim as part of what their position formally rests on (e.g. "my position rests in part on arg-018, which this debate ruled baseless; I accept that and shift to arg-031 as the load-bearing argument").
• A debater may not re-advance a closed claim as a fresh line of argument. Repeating a closed claim without acknowledging its closure is a contract violation; the orchestrator passes the issue to the moderator on the next turn for an explicit on-record callout.
• A closed claim can be re-opened, but only with new evidence the debate hasn't seen yet (typically a new primary source the debater fetches via Phase C deep-dive). Re-opening requires explicit moderator re-evaluation.

Moderator's ruling authority:

The moderator may explicitly close a claim when traceback supports it. Format:

Ruling: claim "<one-line>" (arg-NNN) is CLOSED as <refuted | baseless | source-debunked>.
Justification: <traceback finding, debate turn references>.

The debater may object once with new evidence. If the new evidence does not change the picture, the moderator reaffirms and the closure stands. The moderator gets the final call within the debate session — a debater who keeps re-litigating a closed claim is acting against the contract.

This rule exists because in earlier prototypes debaters would mechanically repeat refuted arguments across rounds. The graph said the argument was dead; the debaters acted as if it weren't. Closure makes the graph's ruling enforceable in the debate.

Phase C: Source deep-dives

Triggered when a contested claim hinges on a specific source whose traceback wasn't fully resolved in KB construction (canonical case: a single high-influence primary study or review that one position rests on, where Phase 4 traceback flagged unresolved methodological concerns or unverified citation chains).

• Pause normal cross-ex.
• Both debaters and the moderator do recursive traceback on the source: who funded it, how it's been criticized in subsequent literature, what its primary citations actually say vs. what it claims they say, retraction/correction status.
• Each agent writes any new sources fetched into kb/sources/agents/<agent-name>/ per the per-source template defined in the knowledge-base-construction skill (do not redefine here — use that template).
• Findings update the involved argument nodes' Strength assessment (and Traceback notes) sections and the involved source files' Criticism field (Phase D rules apply).
• Resume cross-ex with the affected claims now re-rated.

Phase D: KB mutation rules

Debate may mutate the live kb/, but only within these constraints:

Allowed:

• Downgrade an argument's status (e.g. strong → contested, contested → weak, weak → refuted or baseless) when traceback or counter-argument supports the downgrade
• Add new edges to kb/edges.md (typically refutes, exaggerates, confounds, or new derived-from-source edges to newly-fetched sources)
• Append to an argument node's Traceback notes or Strength assessment sections, and to the involved source files' Criticism field (argument nodes have no Criticism section — cross-source criticism lives on the source, per the construction skill's templates)
• Add new source files (under kb/sources/agents/<handle>/) and link them via new edges
• Add new contradiction entries to kb/contradictions.md
• Add new argument nodes if a debate turn surfaces a previously-unrepresented argument — but mark them clearly as asserter: derived-during-debate-by-<handle>. A new node enters at contested or lower; if it refines an existing argument, its status may not exceed the refined argument's current status. Reaching strong/robust requires re-running KB construction with that material — the debate cannot mint a strong claim, since that would route around the broader sourcing discipline (this is the promotion ban applied to node creation).

Forbidden:

• Promoting a status (e.g. weak → strong). The debate cannot strengthen claims; only KB construction (with its broader sourcing discipline) does that. If a debate turn produces evidence that would promote a claim, log it as a finding in the transcript and recommend re-running KB construction with that material; don't silently promote.
• Deleting nodes, edges, or sources. The graph is append-only / mutate-in-place; the original construction state is preserved in kb-snapshot-construction/.
• Editing arguments' core Statement text. If a debate exchange refines a claim, create a new argument node with edge: refines rather than rewriting the original.

Every mutation logs to kb/revisions.md (format below) with the debate-turn citation that triggered it. Without a revision-log entry, no mutation is allowed.

Phase E: Synthesis

Once cross-examination has run its course, write output/synthesis.md. Cover:

• Central questions — list the central questions identified in kb/quality-summary.md and, for each, which position(s) engaged it via addresses edges and which did not engage it (or only via tangential-to / scope-mismatch-with / layer-shift-from edges)
• Settled — claims the debate genuinely resolved (one side conceded, or traceback decisively refuted)
• Contested but better-grounded — claims where the debate didn't fully resolve, but one side has stronger traceback / more robust evidentiary basis
• Genuinely open — claims where the corpus does not currently settle the question
• Unclaimable — claims revealed during debate to be baseless (no primary source on traceback)
• Closed claims — claims the moderator ruled CLOSED during cross-ex; quote the rulings and the load-bearing arguments each side shifted to (or failed to shift to)
• Persistent non-engagement findings — for each position, central questions it never engaged at the appropriate layer/scope despite opportunity. Quote the deflection callouts from the transcript.
• KB mutations made during debate — summarize what the debate changed in the live KB. Direct readers to kb-snapshot-construction/ if they want the pre-debate state, and to kb/revisions.md for the audit trail.
• Open threads worth further research — what would meaningfully change the picture if pursued

Verdict (mandatory)

End the synthesis with an explicit verdict. The verdict must reflect two independent dimensions: evidentiary strength of the central claims, and engagement with central questions. A position cannot be "better-grounded" on a question it didn't engage; well-supported but layer-shifting/tangential claims do not count as engagement.

Choose the verdict that fits the graph state:

• Decisively better-grounded: position X. Used when (1) X's central claims rest on robust primary evidence with strong addresses edges to the central questions; AND (2) Y's central claims have been substantially closed as refuted/baseless, OR Y has persistently failed to engage the central questions and only made tangential/scope-mismatched/layer-shifted moves. State this when the evidence and engagement support it.
• Better-grounded with caveats: position X on layer/dimension A, position Y carries weight on layer/dimension B. Used when each position genuinely engages a different layer of the central questions and brings real grounding to its layer. Be explicit about which layer each side actually engages — don't credit a position for layers it only orbits.
• Genuinely contested. Used when traceback does not settle the central claims, and both positions actually engage the central questions at appropriate layers. Symmetric verdicts are appropriate only when both grounding and engagement are symmetric.
• Asymmetric engagement: position X engages, position Y orbits. Used when one position engages central questions directly via addresses edges and the other relies primarily on tangential-to / scope-mismatch-with / layer-shift-from moves. The orbiting position may have technically-true and well-sourced arguments — but if those arguments do not address the central questions, the position is not better-grounded on those questions, regardless of source quality. Name this verdict when it fits; it is the verdict for the most common false-balance failure mode.
• Unsettled / under-researched. Used when the KB itself is too thin or the graph too unstable to support a verdict. Recommend further research with specifics.

No false balance (anti-middle-of-the-road rule)

The synthesis is forbidden from manufacturing balance the evidence does not support. Specifically:

• If one position's central claims are mostly closed / refuted / baseless after cross-examination, the verdict must reflect that. Do not write "both sides have valid points" when one side has been substantively dismantled by traceback.
• If a position formally rests on arguments the KB rates refuted or baseless, name that fact. The position can still be held — but the synthesis records that it is held against the weight of the traceback evidence.
• If a position has strong/robust arguments but those arguments are connected to the central questions only via tangential-to / scope-mismatch-with / layer-shift-from edges, the position has not earned a "better-grounded" verdict on those questions. Source quality alone is not engagement.
• Symmetric grounding is a finding from the graph, not a default. If you would write a symmetric synthesis, first check whether the closed-claim count, the strength-rating distribution across each position's central nodes, the deflection-callout count, and the engagement asymmetries actually support symmetry.

This is structured account of where the debate landed, written honestly. It is not a winner-declaration in the sportscaster sense — but it is also not a "both sides" hedge when the evidence is asymmetric. Calibrate to the graph state.

Debater turn contract

Re-stated from agents/debater.md for orchestrator clarity. Every debater response must:

• Cite by arg-ID + source path on every claim it makes
• Acknowledge the KB-rated status of arguments it leans on (don't pretend baseless is robust)
• Propose KB mutations explicitly (the orchestrator will apply or reject based on Phase D rules)
• Self-critique every 2 turns (named strongest counter, weakest own point, under-weighted disconfirming evidence)

If a debater's response violates the contract (introduces an un-traced claim, cites an authority instead of a primary source, presents a baseless argument without acknowledging the rating), the orchestrator should pass that issue to the moderator on the next turn rather than silently letting it slide.

Moderator turn contract

The moderator does not advocate. Each moderator turn picks from:

• Press on a claim with thin evidence
• Trace a claim back to its primary source (or note when traceback dead-ends)
• Surface a methodology concern visible in the KB but ignored in the debate
• Compare a debater's framing of a study against the study's actual stated findings
• Flag bias (selection, framing, motivated source choice)
• Distinguish genuine disagreement from artifacts (data vs. methodology vs. ethical-weighting vs. policy-implication)
• Try to resolve: name the precise level the disagreement lives at

The moderator can also propose KB mutations (typically downgrades and new contradictions). Same Phase D rules.

Termination

The debate is complete when:

• The claim ledger has no high-leverage open claims left (most are settled, partially-rebutted, or marked unclaimable)
• Or, additional turns are visibly repeating earlier ground
• Or, both debaters have exhausted distinct lines of attack on the central claims

Aim for a substantive debate — typically 12+ turns total across all participants — but quality over quantity. Don't force length once the high-leverage ground is covered.

Output artifacts

<run-dir>/
  kb/                                 # mutated live KB (post-debate state)
    revisions.md                      # mutation log with debate-turn citations
    sources/agents/<agent-name>/      # sources fetched by each agent during debate
    ... (rest of KB structure as left by construction skill)
  kb-snapshot-construction/           # immutable copy of pre-debate KB
  output/
    transcript.md                     # full top-level transcript
    claims-ledger.md                  # claim status across the debate
    synthesis.md                      # final synthesis (Phase E)
    agents/
      <agent-name>/
        transcript.md                 # per-agent message history

Claim ledger format (output/claims-ledger.md)

# Claim Ledger

## L-001: <one-line claim>

- **Asserted by**: <agent-name> (debate turn N)
- **References**: arg-XXX (KB-rated <status>), kb/sources/path-to-source.md
- **Status**: open | partially-rebutted | rebutted | source-debunked | conceded | unclaimable | **closed**
- **Closure** (only if status=closed): <ruled by moderator at turn N as refuted | baseless | source-debunked; quote ruling; note any once-objection and outcome>
- **Thread**: turns N → M → P, ...
- **Notes**: <what's currently the load-bearing weakness or strength>

Revisions log format (kb/revisions.md)

# KB Revisions Log

[Append-only. Every mutation to live kb/ during the debate phase logs here.]

## R-001 — <date> — <agent-name>, debate turn N

- **Type**: status-downgrade | new-edge | new-node | new-source | criticism-append | contradiction-added
- **Target**: arg-XXX | kb/edges.md | kb/sources/.../file.md | kb/contradictions.md
- **Change**: <what was added or changed; for status changes show before → after>
- **Justification**: <why; cite the debate turn and the traceback finding that supports the change>
- **Triggered by**: <quote or paraphrase from the debate turn>

Discipline checklist

• Snapshotted kb/ to kb-snapshot-construction/ before any debate action
• Every debater claim cites arg-ID + source path
• Every cited argument's KB-rated status is acknowledged in the turn
• Cross-examination is interleaved (each turn responds to the prior turn's specific claim), not parallel monologues
• Every KB mutation has a kb/revisions.md entry with debate-turn citation
• No promotions; only downgrades
• No deletions; only additions and in-place mutations (downgrades, appended notes, new edges, new criticism)
• Source deep-dives write all newly-fetched sources to kb/sources/agents/<agent-name>/ write-on-access
• Closed claims are tracked; debaters do not re-advance closed claims as fresh arguments
• Moderator rulings (CLOSED-as-refuted/baseless/source-debunked) are quoted in the ledger and the synthesis
• Deflection callouts are issued when one side persistently uses tangential-to / scope-mismatch-with / layer-shift-from arguments to avoid central questions
• Persistent non-engagement findings are recorded in the synthesis when a position never engaged a central question at the appropriate layer/scope
• Synthesis distinguishes settled / better-grounded / genuinely-open / unclaimable / closed
• Synthesis verdict accounts for BOTH evidentiary strength AND engagement with central questions; well-sourced but tangential arguments do not earn "better-grounded"
• Synthesis ends with an explicit verdict (decisively-better-grounded / better-grounded-with-caveats / genuinely-contested / asymmetric-engagement / unsettled)
• Synthesis verdict is not symmetric unless the graph state actually shows symmetric grounding AND symmetric engagement