Introducing Batch 7: our largest research release yet

Batch 7 expands coverage from 377 assets in Batch 6 to 426: 49 additional instruments, an increase of approximately 13%. The completed collection contains 5,964 individual opinions, up from 4,523 in the previous batch. That is approximately 32% more opinions available for comparison. These are counts of completed research records, not measures of investment performance.

Each asset has 14 opinions: 12 Gemini 3.8 Flash configurations, one Claude Opus 5 researcher and one GPT-6 Astra researcher. The coverage spans the platform’s five research categories: equities, cryptoassets, indices and funds, commodities, and foreign exchange. Readers can examine both the asset-level conclusion and the different paths that contributed to it.

Scale matters because it gives the research process a broader set of cases to inspect. A framework that appears persuasive for one company may behave differently for an index, a currency pair or a digital asset. A larger universe makes those differences easier to investigate, while preserving a common structure for comparing assumptions, risks and forecast horizons.

An equal-weight average is a useful baseline. It is simple to calculate and easy to explain. But it gives a well-supported argument the same influence as a report built on a stale premise, a questionable comparison or an internal contradiction. Agreement alone cannot tell us whether the underlying reasoning is sound.

V7 adds a synthesizer that reviews the supplied opinions, identifies common causal patterns and examines the strongest opposing cases. It assigns a weight and a written rationale to every opinion, including any opinion assigned zero weight. The assessment considers the quality and freshness of evidence, the plausibility of assumptions and the extent to which apparently separate views repeat the same reasoning.

The synthesizer receives blinded opinion identities and the associated investment frameworks. Underlying model, provider and execution-mode labels are withheld from this assessment. A familiar provider name should not earn extra influence, and a minority view should not lose influence simply because it disagrees with the majority.

Weights apply to the complete supplied forecast paths. After validation and normalization, code calculates the weighted return for each of the 20 quarterly steps and compounds those returns from the recorded anchor price. The language model explains its allocation; it does not invent a separate numerical consensus path. Readers can inspect the weights, their rationales and the frozen opinions that support the calculation.

The new layer is designed to reduce the influence of hallucinations and unsupported claims before they shape the synthesized conclusion. Its instructions explicitly call for checking premises, dates, units and evidence. When search is enabled and useful, the synthesizer can investigate material disputed or suspicious claims against primary sources within the recorded evidence cutoff.

A citation supplied by an individual opinion is not automatically fresh verification. If retrieval is unavailable or a fact cannot be checked, the uncertainty must remain visible. When verification changes an opinion’s weight, the assessment should explain the finding and cite the source actually retrieved. An unverified assertion must not be relabeled as a proven hallucination.

These controls add a deliberate review step between individual reasoning and the combined result. They do not certify that every factual statement is correct. Models can share the same mistaken premise, sources can be incomplete, and the synthesizer can itself make errors. The advance is a clearer mechanism for challenging weak claims and documenting the effect of that challenge.

Some of the most useful research begins with an observation the majority has missed. One opinion may identify a balance-sheet constraint, a change in unit economics or a catalyst that challenges the dominant narrative. A simple average can dilute that insight even when its reasoning is stronger than the surrounding consensus.

V7 asks the synthesizer to assess the strongest contrarian case diligently. A distinctive argument can receive more influence when it is supported by evidence and a credible causal explanation. Confidence, eloquence and novelty are not sufficient reasons to increase a weight. Equal weights remain acceptable when the evidence does not justify a difference.

This changes what we ask of consensus. The goal is to explain where the evidence converges, where important uncertainty remains, and which disagreement deserves further attention. The minority argument stays inspectable alongside the majority view, allowing the reader to judge the synthesis rather than accept a single unexplained answer.

Investment-framework diversity and model-family diversity address different questions. Different analytical personas can ask different questions of the same evidence. Different providers can also expose differences in reasoning, retrieval behavior and assumptions that a single model family may reproduce across many personas.

Batch 7 adds Anthropic’s Claude Opus 5 and OpenAI’s GPT-6 Astra alongside Google’s Gemini 3.8 Flash. The three families share a matched Universal Investor researcher framework for comparison. That makes it possible to investigate model differences while holding an important part of the analytical assignment constant.

Broader provider coverage is a practical control against relying exclusively on one family’s blind spots. It does not guarantee independent errors or remove bias. The panel remains asymmetric: 12 of the 14 opinions use Gemini, with one each from Opus and Astra. Framework, mode and model contributions must therefore be interpreted separately rather than presented as three equally represented teams.

V7’s research contracts make room for a more complete explanation of the investment case. Alongside the headline and thesis, consensus records organize the prevailing market narrative, the gap between that narrative and the research, potential catalysts, macroeconomic alignment, drivers, frictions, tail opportunities, tail risks and material disagreement.

Thesis watchpoints identify developments that could strengthen or weaken the case. Quarterly rationales explain the mechanisms behind the forecast path. Equity-specific assessments separate earnings and financial trends, capital expenditure, balance-sheet strength and profitability, helping readers distinguish business quality from the price the market may be willing to pay.

The weighted synthesis contract adds explicit opinion allocations, individual weighting rationales and an overall weighting assessment. Expert and accessible thesis fields serve different reading needs without asking the reader to reconstruct the investment case from a long unstructured answer. Deliberate field budgets keep the research detailed enough to inspect and consistent enough to compare.

We have also strengthened the capture of source references, actual retrieval evidence, input context and configuration lineage. The records distinguish a model’s own citations from provider-reported search activity and from the frozen material supplied to the run. More references improve traceability; their presence alone does not establish source quality or prove that a claim was independently verified.

A new engine should make its changes easier to examine. Engine V7 is a public system version; Batch 7 identifies the execution cohort. Output-schema versions and ranking-formula versions describe separate components. Keeping these identifiers distinct helps readers understand exactly what changed without rewriting the history of older forecasts.

The averaged consensus remains available alongside the synthesized view. Individual opinions stay frozen. Quantitative rankings and portfolio simulations remain separate from the synthesizer-weighted path; the portfolio track record uses its recorded original quantitative leaderboard. A change in synthesis must not be presented as an improvement in historical portfolio results.

This release also adds two lessons to our public Lessons Learned ledger: why consensus needs evidence-sensitive judgment beyond equal opinion weights, and why multiple model families matter alongside multiple investment frameworks. Both record the practical change and its limitations. The next question is empirical: how stable are the allocations, and do they improve outcomes as observations accumulate?

Start with an asset you know. Read its consensus, compare the individual AI opinions, and inspect the explanation behind the synthesized weights. Follow the sources where available. Look for a material disagreement and decide whether the synthesis represents it fairly. The value of the release is in that ability to examine the reasoning for yourself.

For Future Edge Group, open research means making the process inspectable: methodology, configurations, evidence, past forecasts, evaluation, limitations and lessons. It does not mean that every source dataset, full prompt or production component is publicly redistributable. The boundary should be clear, and the research record should remain useful on either side of it.

Batch 7 is our largest completed research release and a substantial step forward in how we organize and challenge AI-generated investment opinions. Broader coverage gives us more to study. Model diversity gives us more perspectives to compare. Evidence-sensitive synthesis gives us a more explicit way to decide what deserves influence. Together, they define iPulse AI Engine V7.