Applied Spark Multi-Voice Protocol and LLM Red Team Workflow

Introduction: Why single-model prompting and shared-thread agent chats fail

In technical or clinical LLM analysis, the primary risk is not gibberish, but polished multi-model consensus around outdated textbook dogma or fabricated citations.

The Spark (Kıvılcım) multi-voice protocol and its embedded LLM red team workflow are engineered to break both failure modes: independent voices answer the same anonymised brief in total isolation; no model enters synthesis by authority; every citation is audited against external registries; and every verdict must state the exact measurement that would falsify it.

1. The seven-step Spark protocol and the S0 author backbone

As documented live on spark.html, every Spark round executes seven sequential gates:

1. 01 · Prior Expectation: Before any model is called, the orchestration channel (Anu · Claude Fable) records a single-line prior expectation in the permanent scorecard, graded at completion as CONFIRMED (EVET), SHARPENED (KESKİNLEŞTİ), or REJECTED (HAYIR).

2. 02 · Baseline Measurement (DURUM): Existing repository metrics and experimental counts are compiled first.

3. 03 · Anonymised Brief: An identity-stripped brief is dispatched so models cannot infer what answer the author prefers.

4. 04 · Two-Tier Voice Collision (S0 Author + Blind External Voices): The domain author (S0 · Yusuf Aziz) builds the structural backbone and field verdicts—never reduced to "one of eleven voices"—while isolated cloud and local channels run blind to one another.

5. 05 · Reaper Citation Filter and Red Team Sieve: Every external citation is queried by the Reaper auditor agent against Crossref, doi.org, and PubMed E-utilities. Fabricated references and generic textbook summaries (tagged EZBER / ROTE) are mechanically stripped before synthesis.

6. 06 · Voice × Thesis Matrix and Proof Cards: Conflicts between voices are mapped into a traceable matrix with exact file:line citations and dilemma cards, culminating in the final Spark verdict.

7. 07 · Scorecard Grading and Reaper Stamp: The initial expectation is scored against the final verdict, and the page receives a machine-verifiable Reaper stamp recording verified and eliminated citations.

2. Field case study: 10 external voices × 11 questions and 420 audited citations

On October 18–19, 2026, during Digital Dentistry: What Is and What Should Be, 10 isolated voices (S1–S10) answered an 11-question brief against 24 S0 field verdicts:

Source: rapor/kivilcim/SEKTOR_MAKALESI_1018/reaper_kaynak.md:28-44 (Crossref + doi.org + PubMed E-utilities audit, October 19, 2026).

3. Inside the LLM red team workflow: fake DOIs, shared memorisation, and 72-hour quarantine

Three empirical findings from the 420-citation audit illustrate why naive multi-agent voting fails without an adversarial red team layer:

• Finding 1 — Correlated hallucinations break naive majority voting: Local voices S6 (Gemma 4 31B), S8 (Enki), and S9 (Enlil) ran in completely separate processes with zero communication, yet all three fabricated the exact same non-existent citations: Joda et al. 2017 "Digital technology in fixed prosthodontics", British Dental Journal and Sailer et al. 2009 "A systematic review…", IJP.
• Finding 2 — "Knows the paper, invents the identifier" (Light Fabrication): In headless voice S10, the cited papers and authors were real, but 7 out of 45 citations carried invented PMIDs or DOIs resolving to unrelated articles.
• Finding 3 — The EZBER (Rote) filter and 72-hour quarantine: Responses that cite real papers but merely repeat generic textbook summaries without addressing the specific field dilemma are tagged EZBER and excluded.

4. How 4 textbook assumptions shared by 9 models collapsed under field measurement

Red teaming also exposes where multi-model consensus diverges from physical reality.

1. Overturned Assumption 1 ("Clinical need precedes device adoption"): Field reality operates in reverse: technology imposes need (T11.5), and where the device exists, measurement exists.

2. Overturned Assumption 2 ("Lithium disilicate for anterior, zirconia for posterior"): Properly bonded single-unit lithium disilicate restorations achieve posterior survival comparable to anterior crowns.

3. Overturned Assumption 3 ("Pressed veneers beat chairside CAD/CAM"): In-house Clinical Lab workflows standardize same-day ceramic/composite laminates and injection moulding, while gradient zirconia discs dissolve old full-arch boundaries.

4. Overturned Assumption 4 ("Hospital quality paperwork raises the clinical floor"): Regulatory quality documentation (SKS) measures paperwork compliance rather than clinical survival.

5. Convergence metrics, scorecard distribution, and falsification gate

Across all 67 Spark rounds recorded in the autonomy scorecard (73 evaluation rows), the initial expectation was confirmed unchanged in only 13 rounds (CONFIRMED), modified and sharpened by counter-evidence in 44 rounds (SHARPENED), and outright falsified in 3 rounds (REJECTED). In three-iteration collision rounds, the inter-voice convergence index rises along a measured 0.30 → 0.48 → 0.68 trajectory.

Falsification gate: If on ≥20 blind synthesis tasks a single model yields fewer citation errors and higher expert accuracy than the 7-step Spark + Reaper pipeline, the protocol fails.

Related core pages: Spark Protocol · Reaper Auditor · Autonomy & Live Meters · Digital Dentistry: What Is and What Should Be · Machine-Readable Records (/bot/) · Event-Driven Task Bus (task_bus.jsonl) · Context-Blind Auditor (Reaper 360/394) · 3D Smile Design