Skip to main content
Assets

Public engine ledger

One engine name. Every material version visible.

iPulse AI Engine V7 is the current end-to-end research operating system. This ledger separates its architecture from the batches that ran it and from the ranking formulas later applied to frozen forecasts.

Updated Updated by Russlan Ramdowar

Trace V1–V7 Open Lessons Learned

The identity model

Three ledgers answer three different questions.

A single version number cannot honestly describe architecture, a particular execution, and a later analytical formula. The relationship is explicit.

Engine release

What operating system existed?

iPulse AI Engine V7

Execution batch

What configuration actually ran?

Batch 6 · July 5, 2026

Formula snapshot

How were frozen forecasts ranked?

Formula v7.4 · run v2.5

Release ledger

From first baseline to V7.

Counts describe completed public batch cohorts. Capability labels describe what the generation run actually used; downstream additions are marked separately.

V7
Engine V7Batch 7Current

Model diversity and explained consensus weights

Batch 7 compares matched researcher configurations across three model families and introduces explained per-opinion synthesizer weights. Averaged and synthesized consensus remain separately inspectable.

Completed forecasts
5,964
Covered assets
426
Pipeline lineage
Governed individual opinions, weighted synthesis and separate quantitative rankings
Output contract
Individual schema v10 with richer research fields; seo_geo_synthesis_v7 with explained opinion weights

Generation capability record

Global contextUsed in run

Governed world-state context accompanies the individual forecasts.

Asset fundamentalsEligible assets

Currency-normalized fundamentals for eligible equities.

Consensus synthesisUsed in run

The synthesizer assigns explained per-opinion weights; Python calculates and compounds the weighted quarterly returns.

Cross-asset rankingsUsed in run

Portfolio simulations use the frozen original quantitative leaderboard, separate from the synthesized forecast path.

Editorial reviewUsed in run

Separate editorial publication approval remains downstream of forecasts and rankings.

Model diversity comparisonUsed in run

Matched Universal Investor researcher configuration across Flash, Opus and Astra; each asset has 12 Flash, one Opus and one Astra opinion.

Evidence-based opinion weightingUsed in run

Blinded opinion identities support lower or zero weights for weak claims and higher weights for supported insights. Weights are normalized and the complete source path is frozen.

Introduced before this run

  • Compare three model families using the same researcher framework.
  • Replace equal opinion contribution in the synthesized path with explained, normalized judgment weights.
  • Preserve both averaged and synthesized consensus; weighting remains experimental and does not establish better accuracy.

Learned after this run

  • Judgment weights can reduce the contribution of weak reasoning, but accuracy gains require subsequent evaluation.
  • Model diversity and framework diversity are separate comparisons.
  • Keep averaged consensus, synthesized paths and quantitative portfolio rankings distinct.

Model and persona configuration

Gemini 3.8 Flash, Claude Opus 5 and GPT-6 Astra

14 opinions per asset: 12 Gemini Flash configurations plus matched Opus and Astra researchers

The synthesizer sees blinded opinion identities and investment frameworks. It assigns weights and rationales; Python calculates quarterly weighted returns and compounds them. The portfolio track record uses the frozen original quantitative leaderboard, not these synthesized prices.

V6
Engine V6Batch 6

Open research operating system

V6 added currency-normalized fundamentals, a larger asset universe, compact outputs, mature rankings, public forecast history, and an editorial approval stage around published Top Picks.

Completed forecasts
4,523
Covered assets
377
Pipeline lineage
Prediction pipeline v6; request metadata v2
Output contract
Structured output schema v9

Generation capability record

Global contextUsed in run

Governed global context was supplied.

Asset fundamentalsEligible assets

Currency-normalized fundamentals were supplied for eligible equities.

Consensus synthesisUsed in run

Independent forecasts were synthesized into asset-level consensus records.

Cross-asset rankingsUsed in run

Current v7.4 architecture adds event-risk pressure and equity financial health; historical snapshots retain their formula identifiers.

Editorial reviewAdded downstream

The formal editorial-review workflow was added after Batch 6 generation and now governs publication of Editorial Top Picks.

Introduced before this run

  • Aligned reporting and ticker currencies and corrected P/E and related fundamental calculations.
  • Expanded the analyzed asset universe and shortened selected narrative fields.

Learned after this run

  • Forecast generation, quantitative ranking, and editorial approval must remain distinct, traceable stages.
  • Frozen forecasts can be re-scored as ranking methodology improves, but the original forecast content must never be rewritten.
  • Public forecast receipts, batch history, and clear methodology provenance create a stronger trust model than an opaque latest-only result.

Model and persona configuration

Gemini 3.1 preview configurations

Governed multi-persona configurations with Thinker and Researcher modes

Editorial review is part of the public V6 operating system, but it did not influence the raw Batch 6 forecasts: it was added downstream after the deep-analysis execution.

V5
Engine V5Batch 5

Consensus and fundamental-context expansion

V5 added the second-layer asset consensus workflow, equity fundamentals, consistent subject-matter-expert language, glossary support, and a shorter schema for more usable reports.

Completed forecasts
3,742
Covered assets
312
Pipeline lineage
Prediction pipeline v6; request metadata v2
Output contract
Structured output schema v9

Generation capability record

Global contextUsed in run

Governed world-state context was included.

Asset fundamentalsEligible assets

Asset-specific fundamentals were introduced for equities, not for asset classes without company financial statements.

Consensus synthesisUsed in run

A second batch layer combined independent advisor forecasts into an asset consensus.

Cross-asset rankingsUsed in run

The v7.2/v7.3 ranking generation incorporated dividend, dispersion, volatility, and cash-hurdle logic.

Editorial reviewNot used

The formal editorial stage was introduced after the next deep-analysis execution.

Introduced before this run

  • Added asset-level consensus synthesis after independent advisor forecasts.
  • Added asset-specific fundamentals after global context for equity assets.
  • Standardized report language at subject-matter-expert level and added glossary assistance.
  • Increased Thinker-mode use and reduced overly long output fields.

Learned after this run

  • Fundamental reporting currency can differ from ticker currency; mixing them can corrupt valuation ratios and confuse downstream reasoning.
  • Consensus summaries need to preserve disagreement instead of merely averaging voices.

Model and persona configuration

Gemini 3.1 preview configurations

Governed persona set with more Thinker-mode executions

V4
Engine V4Batch 4

Instruction-retention hardening

The fourth release hardened long-context execution with repeated critical instructions and a prompt-sandwich structure, while preserving the verified global-context layer.

Completed forecasts
3,437
Covered assets
312
Pipeline lineage
Prediction pipeline v6; request metadata v2
Output contract
Primarily schema v5, with a small schema v6 transition cohort

Generation capability record

Global contextUsed in run

Governed global context was included in the run.

Asset fundamentalsNot used

Asset-specific fundamentals were not yet part of the run.

Consensus synthesisNot used

The dedicated asset-consensus layer followed this batch.

Cross-asset rankingsUsed in run

Rankings used the then-current post-generation formula and were later eligible for reproducible re-scoring.

Editorial reviewNot used

The editorial publication workflow did not yet exist.

Introduced before this run

  • Repeated the task goal and critical asset instructions at both ends of long prompt assemblies.
  • Kept global context while reducing the risk that essential instructions were lost inside 50,000-plus-token prompts.

Learned after this run

  • Long-context workflows need deliberate instruction placement, not merely more context.
  • Independent forecasts need an asset-level synthesis layer so readers can inspect one governed consensus overview.

Model and persona configuration

Gemini 3 Pro preview configurations

Eight governed personas with broader asset coverage

V3
Engine V3Batch 3

Verified global-context start

Batch 3 is the verified start of supplied global market context. It also expanded asset coverage, simplified the persona set, and opened more configuration lineage publicly.

Completed forecasts
2,719
Covered assets
247
Pipeline lineage
Prediction pipeline v6; request metadata v2
Output contract
Structured output schema v5

Generation capability record

Global contextUsed in run

Verified 2025–2026 world-state context was supplied. This is the first batch for which that claim is valid.

Asset fundamentalsNot used

Asset-specific fundamentals were not yet part of the run.

Consensus synthesisNot used

The dedicated asset-consensus layer had not yet been introduced.

Cross-asset rankingsUsed in run

Post-generation ranking work existed, with formulas continuing to evolve after the run.

Editorial reviewNot used

The editorial publication workflow did not yet exist.

Introduced before this run

  • Added global events and quantitative world-state context, arranged to preserve temporal proximity.
  • Expanded the asset universe and further revised the structured output schema.
  • Removed the ambiguous Nostradamus persona configuration.

Learned after this run

  • Model tool use is not reliable enough to substitute for a governed current-context package.
  • Context should represent the world state without repeatedly over-emphasizing one event.
  • Dividend history and past volatility belong in comparable ranking logic.
  • Public configuration lineage materially improves inspectability and methodological accountability.

Model and persona configuration

Gemini 3 Pro preview configurations

Eight governed personas

V2
Engine V2Batch 2

Persona-governed analysis

The workflow moved from generic prompt configurations to named investment personas, stricter time-series rules, and a much more explicit output contract.

Completed forecasts
1,413
Covered assets
212
Pipeline lineage
Prediction pipeline v4; request metadata v2
Output contract
Structured output schema v4

Generation capability record

Global contextNot used

Batch 2 did not receive the later world-state context files.

Asset fundamentalsNot used

No governed fundamentals input was supplied.

Consensus synthesisNot used

The asset-consensus layer had not yet been introduced.

Cross-asset rankingsAdded downstream

The first ranking implementation was created after this run, beginning March 27, 2026.

Editorial reviewNot used

The editorial publication workflow did not yet exist.

Introduced before this run

  • Added persona, investment-framework, behavior, and communication-style configuration.
  • Tightened forecast-anchor and horizon-step requirements in output schema v4.
  • Standardized the main forecast path around a five-year structure.

Learned after this run

  • The same influential personas must cover the comparison universe or cross-asset scoring becomes structurally unfair.
  • Abstract personas can create stylistic ambiguity instead of useful financial reasoning; Nostradamus was retired.
  • A ranking layer was needed to compare the growing set of frozen forecasts.

Model and persona configuration

Gemini 3 Pro preview

Nine persona configurations, including the later-retired Nostradamus experiment

V1
Engine V1Batch 1

Structured market-wide baseline

The first large-scale execution established a comparable structured-output baseline, but still relied on a single prompt configuration and legacy horizon handling.

Completed forecasts
2,400
Covered assets
204
Pipeline lineage
Legacy request metadata v1; pre-versioned generation workflow
Output contract
Mixed early structured schemas (1, 3, and 4)

Generation capability record

Global contextNot used

No global-context component was supplied to Batch 1.

Asset fundamentalsNot used

No governed fundamentals input was supplied.

Consensus synthesisNot used

Forecasts were produced independently without the later asset-consensus layer.

Cross-asset rankingsAdded downstream

Ranking did not exist at generation time; frozen forecasts were eligible for later retrospective scoring.

Editorial reviewNot used

The editorial publication workflow did not yet exist.

Introduced before this run

  • Introduced machine-readable structured JSON outputs for comparison and later inspection.
  • Ran separate three-year and five-year forecast configurations as an early horizon experiment.

Learned after this run

  • Explicit bullish or bearish persona instructions bias the forecast instead of testing the most reasonable scenario.
  • API models do not inherit the browsing behavior of consumer chat applications; current context must be deliberately supplied or retrieved.
  • Forecast dates require one standardized anchor to remain comparable across advisors and assets.
  • A single five-year path can support shorter-horizon extracts more consistently than separate, incomplete paths.

Model and persona configuration

Gemini 2.5 GA and Gemini 3 preview configurations

No governed persona layer

Separate downstream ledger

How the ranking formula evolved.

Rankings are derived from frozen forecasts. Older batches can be re-scored under a later formula, so each snapshot carries its own formula and run identifiers.

Initial ranking system

First cross-asset consensus ranking

  • — Introduced a comparable post-generation scoring and leaderboard workflow after Batch 2.
  • — Allowed already-frozen Batch 1 and Batch 2 forecasts to be evaluated retrospectively.

This formula did not generate Batch 1 or Batch 2 forecasts. It was a later analytical layer applied to their frozen outputs.

Formula audit v6

Validity and deterministic-ranking audit

  • — Hardened numeric validity checks and dividend caps.
  • — Made tie-breaking deterministic and expanded formula diagnostics.

The public architecture label and the stored scoring-run version remain separate identifiers.

Formula v7.2

Confidence-adjusted excess return

  • — Added a cash-return hurdle, median-absolute-deviation confidence discount, and historical-volatility context.
  • — Combined risk-adjusted return with direction consistency and added a negative-return guard.

Dividend contribution uses three complete historical years; volatility uses history anchored to the original batch analysis date.

Formula v7.3

Cash-hurdle calibration

  • — Lowered the annual cash hurdle from 4.0% to 3.6%.
  • — Preserved the v7 confidence-adjusted architecture while recording the new formula label on generated leaderboard rows.

Production rows for older batches can carry this later formula because the forecasts were re-scored without altering their original content.

Current

Formula v7.4 · run v2.5

Risk-aware current ranking architecture

  • — Added event-risk pressure, risk-resilience presentation, and equity financial-health diagnostics.
  • — Retired the public three-year leaderboard path and kept one-year and five-year comparison views.
  • — Preserved formula and scoring-run identifiers with every ranking snapshot.

“v7.4” names the public scoring architecture. “v2.5” identifies the persisted scoring run implementation; they are not interchangeable.

Versioning contract

What makes a version change.

Future releases follow a stricter semantic contract. Component versions stay independent and remain frozen with each prediction.

Major engine version

V7

Changes when a material workflow stage, governed input class, output layer, or publication-control boundary is added, removed, or fundamentally redesigned.

Minor engine version

V7.1

Reserved for compatible configuration or methodology improvements that materially affect future runs without changing the engine architecture.

Patch version

V7.1.1

Documentation, provenance, or metadata corrections that do not change forecasts, scoring results, or workflow behavior.

Execution batch

Batch 7

An immutable run of a particular engine and component configuration. A batch is evidence of execution, not a synonym for an engine release.

Component version

schema v10 / synthesis v7

Models, personas, task configs, pipelines, input packages, and output schemas keep independent versions recorded with each prediction.

Ranking formula

v7.4 / run v2.5

A downstream analytical version stored on each ranking snapshot. It can evolve and re-score frozen forecasts without changing their engine or batch identity.

Prediction-level source of truth

The engine label never replaces component lineage.

Every deep-analysis record can carry its exact model release, task configuration, persona, mode, input package, pipeline, output schema, and later ranking identifiers. V7 is the current public system version; those component versions explain precisely what executed.

Inspect data provenance Read Lessons Learned