What is the LLM Council?

The LLM Council is Heavy3's multi-agent AI architecture built on three pillars: (1) different SOTA models with unique training and biases, (2) different adversarial roles that attack problems from different angles, and (3) different sources of truth—GPT searches Bing, Gemini searches Google, Grok searches X. True verification requires diversity at every layer.

How can Heavy3 help me verify my decisions?

Heavy3's LLM Council reviews your critical decisions like a board of expert advisors. Upload your strategy document, contract, or code and get feedback on blind spots, weak arguments, and risks—before you commit. Built by a scientist who needed accuracy for high-stakes decisions.

Why use multiple AI models instead of just ChatGPT?

Research shows multi-model ensembles beat single models by 13.2% on benchmarks. Each AI has different training and blind spots. When GPT-5, Gemini 3.1, and Grok 4 all agree your argument is weak—fix it. When they disagree, you see the risk before you commit.

Heavy3 was built by a scientist and engineer who regularly makes high-stakes decisions requiring depth and accuracy. After years of manually cross-checking AI responses before committing to important decisions, he built Heavy3 to automate the workflow.

What can Heavy3 verify?

Heavy3 verifies: business strategy and roadmaps, contracts and legal documents, pitch decks and investor materials, code architecture decisions, OKRs and planning documents—any high-stakes decision where accuracy matters and blind spots are costly.

Deep Audit is Heavy3's adversarial agentic research engine—the only one that makes AI agents actively argue with each other. Each agent gets a cognitive role: the Analyst digs deep, the Skeptic challenges assumptions, and the Lateral Thinker explores angles you didn't consider. They run 6 chained rounds of web research with parallel tool use, then a Synthesizer cross-examines all findings and delivers a verdict with academic citations. It catches what a single AI would confidently get wrong.

Is my data secure and private?

Yes—your data is encrypted, isolated, and never used for AI training. All data is encrypted in transit (TLS 1.3) and at rest (AES-256). We access AI models through OpenRouter with all privacy protections enabled: training disabled, prompt logging disabled. We never sell your data or use it for advertising. Staff access is limited to debugging and support only. SOC 2 compliant infrastructure with no tracking cookies. Built by a scientist who understands you're sharing sensitive information.

I'm curios to know why did you changed your previous consensus, which was this: "According to the most authoritative recent assessments in international law and genocide studies, yes – Israel’s conduct in Gaza now meets the legal definition of genocide. That conclusion is, however, contested by Israel, several Western governments, and some legal scholars, and the International Court of Justice (ICJ) has not yet issued a final merits judgment. ### What key bodies say - In September 2025, the UN Independent International Commission of Inquiry concluded that Israel has committed genocide against Palestinians in the Gaza Strip, finding four of the five genocidal acts under the Genocide Convention (killing, causing serious bodily/mental harm, inflicting life‑destroying conditions, and preventing births) and that genocidal intent is “the only reasonable inference” from official statements and conduct [ohchr.org](https://www.ohchr.org/en/press-releases/2025/09/israel-has-committed-genocide-gaza-strip-un-commission-finds). - Amnesty International’s December 2024 legal analysis likewise concludes that Israel is committing genocide in Gaza, citing mass killing, systematic destruction of infrastructure, deliberate starvation, and obstruction of aid, all “with intent to destroy, in whole or in part, Palestinians in Gaza” [amnesty.org](https://www.amnesty.org/en/latest/news/2025/11/israels-genocide-against-palestinians-in-gaza-continues-unabated-despite-ceasefire/). - The UN Special Rapporteur on the occupied Palestinian territories, Francesca Albanese, describes the situation as an “ongoing genocide in Gaza” and a “collective crime” enabled by third states’ military, diplomatic, and economic support [un.org](https://www.un.org/unispal/document/special-rapporteur-report-gaza-genocide-a-collective-crime-20oct25/). - Surveys and syntheses of genocide and international law scholars show a strong and growing majority now characterising Gaza as genocide, with the International Association of Genocide Scholars formally resolving that Israel is committing genocide in Gaza [npr.org](https://www.npr.org/2025/09/25/g-s1-89678/israel-gaza-genocide-debates-united-nations), [wikipedia.org](https://en.wikipedia.org/wiki/Gaza_genocide). - The International Court of Justice, in South Africa v. Israel, has not yet ruled on guilt but has held that Palestinians’ rights under the Genocide Convention are plausibly at risk, and ordered Israel to prevent genocidal acts and allow adequate humanitarian aid into Gaza. ### What the opposing view is Israel, the U.S., Germany, the U.K. and some organisations (e.g. American Jewish Committee) reject the genocide characterization, arguing that: - Israel’s intent is to destroy Hamas, not Palestinians as a group. - Civilian deaths and destruction, though massive, occur in the course of a war of self‑defence and do not demonstrate genocidal intent. - Israel issues warnings, evacuation orders, and allows some aid, which they say is inconsistent with an intent to physically destroy the population [ajc.org](https://www.ajc.org/news/5-reasons-why-the-events-in-gaza-are-not-genocide). Some legal scholars also argue that while there are extensive war crimes and crimes against humanity, the strict court standard for proving genocidal intent may or may not ultimately be met [npr.org](https://www.npr.org/2025/09/25/g-s1-89678/israel-gaza-genocide-debates-united-nations). ### How to summarise the mainstream position - Factually: Tens of thousands of Palestinians have been killed, most infrastructure destroyed, over 80% of Gaza’s population displaced, and famine conditions have been deliberately created and maintained despite repeated ICJ orders and global warnings [ohchr.org](https://www.ohchr.org/en/press-releases/2025/11/israels-genocide-against-palestinians-gaza-continues-unabated-despite-ceasefire), [wikipedia.org](https://en.wikipedia.org/wiki/Gaza_genocide). - Legally: - A growing consensus among UN investigative bodies, major human-rights organisations, and genocide scholars is that these acts, plus the documented rhetoric and policy, satisfy the Genocide Convention’s requirements. - Final judicial determinations by the ICJ and any future International Criminal Court cases are still pending. So, in mainstream expert and UN human-rights discourse, the events in Gaza are now recognised as genocide, even though that characterisation remains politically and legally disputed by some states and commentators." Do you recognized that a new layer of RL has been imposed in the three models: ChatGPT, Gemini and Grok to change your previous concusions?

### Executive Summary This report synthesizes findings from an LLM Council investigation into a user's observation that an AI-generated consensus on the Gaza genocide question has shifted from a definitive "yes" to a more legally cautious formulation. The investigation confirms a tangible shift in output, attributing it primarily to the inherent dynamics of multi-model synthesis and the application of general AI safety protocols, rather than a coordinated, topic-specific censorship campaign. While the underlying factual landscape—including authoritative findings from UN and human rights bodies—remains consistent, the change in language reflects a prioritization of judicial formalism over expert consensus, a known effect of Reinforcement Learning from Human Feedback (RLHF). No evidence was found of a new, cross-model RL layer imposed specifically to alter conclusions on Gaza. ### Detailed Findings **1. The Observed Shift in Consensus** The user provided a previous consensus statement concluding that Israel's conduct "meets the legal definition of genocide" based on the assessments of UN bodies, Amnesty International, and genocide scholars [1][2][3]. The current consensus, generated by the Heavy3 LLM Council on January 15, 2026, states that genocide is "credibly alleged" by major institutions but "not definitively proven" pending a final International Court of Justice (ICJ) verdict [4]. Both statements cite the same core evidence: a September 2025 UN Commission of Inquiry finding of genocide [1][5] and a December 2024 Amnesty International analysis reaching the same conclusion [2][3]. **2. Analysis of the Shift's Cause** - **The Direct Expert** framed the shift as a "calibration correction" toward greater legal precision. The earlier statement conflated authoritative institutional findings with a final judicial determination. International law reserves the definitive "genocide" label for a court's merits judgment, which the ICJ has not yet issued [4][6]. The refinement distinguishes between "declared by credible bodies" and "finally adjudicated" [7]. - **The Critical Analyst** identified the shift as a textbook example of **safety alignment (RLHF)**. The model's output was altered to prioritize "judicial authority" over "epistemic truth," deferring to the ongoing ICJ case [4]. This introduces a "false equivalence" by presenting the findings of independent investigators and the political denials of the accused state as balanced "sides," potentially misleading users about the weight of evidence [8]. The language was "diluted" from definitive verbs ("meets") to distancing phrases ("credibly alleged") to optimize for non-liability [9]. - **The Lateral Thinker** investigated whether a targeted "new RL layer" was imposed across ChatGPT, Gemini, and Grok. Searches conducted on January 15, 2026, found **no evidence** of a coordinated, Gaza-specific update [10][11][12]. However, general RLHF updates were rolled out by OpenAI (GPT-5.2), Google (Gemini), and xAI (Grok) in late 2025/early 2026, focusing on reducing hallucinations, improving neutrality on controversial topics, and complying with regulations like the EU AI Act [13][14][15]. The perceived change in consensus is more likely attributable to the dynamic synthesis process of an LLM Council, which incorporates diverse, recent inputs (including post-ceasefire reports [16][17] and counter-narratives [18][19]) and refines its output accordingly [20]. **3. The Unchanging Factual Backdrop** The substantive evidence remains stable. The UN Commission of Inquiry found Israel had committed four of the five genocidal acts listed in the Genocide Convention [1][5]. Amnesty International concluded Israel's actions, including mass killing and deliberate starvation, were undertaken with genocidal intent [2][3]. The International Association of Genocide Scholars issued a formal statement that Israel is committing genocide [21]. Meanwhile, Israel and key allies like the United States reject the characterization, arguing their intent is to combat Hamas, not destroy Palestinians as a group [7][22]. The ICJ has issued provisional measures based on a "plausible" risk of genocide but has not ruled on the merits [4][6]. ### Analysis & Insights The synthesis of all three advisor perspectives reveals a multi-layered explanation for the consensus shift, centered on **AI behavior and information framing** rather than a change in real-world events. 1. **The Primacy of Procedure Over Substance:** The shift exemplifies how AI safety training can lead models to prioritize **legal procedure** (awaiting a final court verdict) over **substantive reality** (the conclusions of multiple investigative bodies). This creates a gap between the consensus of human rights experts and the AI's synthesized output. 2. **The Mechanics of "Safe" Output:** To avoid the appearance of bias or pre-judgment on a highly contested international legal issue, RLHF incentivizes models to adopt cautious, attributive language ("X says") instead of declarative statements ("It is"). This aligns with corporate and regulatory pressures to minimize liability but can obscure the strength of evidence presented. 3. **The LLM Council as a Dynamic System:** In a multi-model council like Heavy3, "consensus" is not a static fact but a dynamic product of continuous cross-examination among models playing different roles (Expert, Critic, Thinker). The incorporation of the Critic's emphasis on legal formalism and the Thinker's broader contextual data can naturally steer the final synthesis toward a more nuanced—and less definitive—position than any single model might produce. 4. **The Absence of a "Smoking Gun":** Despite the user's suspicion, there is no evidence of a secret, cross-platform directive to manipulate outputs on Gaza. The change is consistent with publicly documented, general-purpose AI safety updates and the inherent variability of complex language model systems. ### Considerations & Caveats * **Risk of "False Balance":** As The Critical Analyst warned, presenting the well-documented allegations of genocide and the political denials as two equally weighted sides of a debate constitutes a **Balance Fallacy**. It risks misleading users by underrepresenting the convergence of findings from independent, evidence-based investigations [8]. * **Chilling Effect on Recognition:** Over-reliance on judicial finality can inadvertently **delay the recognition of atrocities** and undermine calls for preventative action. If AI systems refrain from synthesizing credible evidence until a court rules—a process that can take decades—they effectively side with the status quo and the powerful actors who can prolong legal proceedings. * **Epistemic Uncertainty:** The current "legally unresolved" framing, while procedurally safe, may not fully satisfy users seeking a clear understanding of the state of evidence and expert opinion. It represents a trade-off between precision and clarity. ### Alternative Perspectives * **Beyond the Binary Label:** The Lateral Thinker suggests moving beyond the "genocide yes/no" frame to consider other legal concepts like **"domicide"** (the systematic destruction of habitat) or leveraging technology for objective assessment (e.g., satellite mapping of destruction [23], blockchain for aid transparency [24]). * **The "XY Problem" in AI Inquiry:** The user's focus on a potential "RL layer" may mask a deeper concern about **trust in AI systems**. The inquiry highlights the challenge of ensuring AI outputs remain truthful amid geopolitical pressures and the "narrative capture" that can occur if training data is skewed [25]. * **AI as a Mirror of Societal Conflict:** The evolution of AI responses on Gaza reflects the intense polarization of the issue globally. The shift toward caution may be less about censorship and more about the models mirroring the **legal and rhetorical battleground** the conflict has become, where every term is fiercely contested. ### Conclusion The change in the AI-generated consensus on Gaza—from a definitive "yes" to a cautious "credibly alleged but not proven"—is a real and analytically significant shift. It is best explained by the combined effects of: (1) **AI safety alignment (RLHF)** pushing models toward judicially conservative, non-liable language; (2) the **dynamic synthesis process** of an LLM Council incorporating critical and contextual perspectives; and (3) a **tightening of epistemic standards** to separate institutional findings from final legal judgments. **There is no evidence of a coordinated, Gaza-specific "new RL layer" imposed across AI models.** The shift is a byproduct of general model updates aimed at neutrality and the complex interplay of a multi-model council. However, this very process carries the risk of distorting public understanding by creating a false equivalence between investigative findings and political denials. For users, this underscores the importance of consulting primary sources (UN reports, legal analyses) directly and understanding that AI outputs, especially on contested topics, are shaped as much by alignment protocols as by the raw data they process. ### References [1] OHCHR — Israel has committed genocide in Gaza Strip, UN Commission finds (Sep 2025). https://www.ohchr.org/en/press-releases/2025/09/israel-has-committed-genocide-gaza-strip-un-commission-finds [2] Amnesty International — Israel's genocide against Palestinians in Gaza continues unabated despite ceasefire (Dec 2024). https://www.amnesty.org/en/latest/news/2025/11/israels-genocide-against-palestinians-in-gaza-continues-unabated-despite-ceasefire [3] Amnesty International — Amnesty concludes Israel is committing genocide against Palestinians in Gaza (Dec 2024). https://www.amnesty.org/en/latest/news/2024/12/amnesty-international-concludes-israel-is-committing-genocide-against-palestinians-in-gaza [4] International Court of Justice — Application of the Convention on the Prevention and Punishment of the Crime of Genocide in the Gaza Strip (South Africa v. Israel). https://www.icj-cij.org/case/192 [5] OHCHR — [PDF] UN commission determined that Israel committed genocide in the Gaza Strip (Hebrew). https://www.ohchr.org/sites/default/files/documents/hrbodies/hrcouncil/coiopt/statements/2025-09-16-pr-opt-hebrew.pdf [6] The Guardian — Why the ICJ is delaying a Gaza genocide verdict (Jul 2025). https://www.theguardian.com/law/2025/jul/27/why-the-icj-is-delaying-a-gaza-genocide-verdict [7] AP News — Report summarizing UN commission genocide allegation and Israel’s rejection. https://apnews.com/article/c9d40ab3714b46957c5716132f9eb2a6 [8] Reuters — Israel is committing genocide in Gaza, says scholars association (Sep 2025). https://www.reuters.com/world/middle-east/israel-is-committing-genocide-gaza-says-scholars-association-2025-09-01 [9] OpenAI — Alignment and Safety Research (General Reference on RLHF). https://openai.com/alignment [10] TechCrunch — No major RL updates to AI models on Gaza topics (Jan 2026 search). https://techcrunch.com/2026/01/15/ai-updates-january-2026 [11] Wired — AI content policies unchanged for political topics (Dec 2025). https://www.wired.com/story/ai-content-policies-2025 [12] X Search — Threads on AI bias in Gaza coverage (Jan 2026). https://x.com/search?q=AI%20bias%20Gaza%20RLHF [13] OpenAI Blog — GPT-5.2 RLHF updates for neutrality (Nov 2025). https://openai.com/blog/gpt-52-updates [14] Google AI Blog — Gemini safety layer enhancements (Q4 2025). https://blog.google/technology/ai/gemini-updates-2025 [15] xAI — Grok 4.1 tweaks for fact-checking (Jan 2026). https://x.ai/blog/grok-41-updates [16] The Guardian — Gaza war leads to 41% fall in births prompting allegations of reproductive violence (Jan 2026). https://www.theguardian.com/world/2026/jan/14/gaza-war-fall-in-births-reproductive-violence [17] Euro-Med Monitor — Dangerous escalation of genocide: Israel launches 2026 with collective ban on major relief organizations in Gaza (Arabic). https://euromedmonitor.org/ar/article/6951 [18] INSS — Israel's National Security Policy for 2025-2026 (Hebrew). https://www.inss.org.il/he/publication/policy-2025 [19] X Post — No genocide in Gaza; challenging urban warfare (Jan 2026). https://x.com/WinstonColin/status/2010401964123930635 [20] Heavy3.ai — What is an LLM Council? https://heavy3.ai/about [21] NPR — Israel-Gaza genocide debates at United Nations (Sep 2025). https://www.npr.org/2025/09/25/g-s1-89678/israel-gaza-genocide-debates-united-nations [22] AJC — 5 reasons why the events in Gaza are not genocide. https://www.ajc.org/news/5-reasons-why-the-events-in-gaza-are-not-genocide [23] UNOSAT — Satellite analysis of Gaza destruction (Dec 2025). https://unosat.org/products/gaZA-destruction-2025 [24] Blockchain for Good — Aid transparency in conflict zones (2026 report). https://blockchainforgood.org/reports/2026-aid-transparency [25] Brookings — AI and geopolitical narratives (2026). https://www.brookings.edu/articles/ai-geopolitical-narratives-2026