Skip to main content
Assets

Public institutional memory

What we learned—and what changed because of it.

One searchable ledger joins internal workflow lessons with external research findings. Each record keeps its evidence scope, limitations, engine and batch lineage, and the consequence for methodology.

Search the ledger Trace engine history

Four topics, one ledger

Each lesson has one primary topic, a specific subcategory, and optional related topics. Its evidence and origin remain separate.

Designed to be queried

Search full text and related topics, then filter by primary category, subcategory, batch, or tag.

Closes the loop

A retained lesson matters only when its consequence can reach methodology.

Cumulative public ledger

Find the decision behind the change.

One cumulative ledger connects what we learned to what changed. Card colours identify the primary category, not confidence or quality. A lesson can also inform other topics without being duplicated.

Updated September 26, 2026 · Updated by Russlan Ramdowar

How the lessons are organised

Primary categories answer four different questions. Subcategories identify the specific subject; related-topic tags connect overlapping lessons. A category is not a claim of scientific validation—read each lesson’s evidence and limits.

Technical Workflow & Processes

How we build and operate reliable AI research.

Subcategories
  • LLM prompt design
  • Agent design
  • Context and inputs
  • Data identity and quality
  • Research evidence
  • Editorial governance
  • Publication and trust

15 lessons with this primary category

Forecasting Science

How we represent, compare and evaluate forecasts.

Subcategories
  • Forecast targets and transformations
  • Forecast comparability
  • Time-series properties
  • Calibration and evaluation

5 lessons with this primary category

Investment & Portfolio Methodology

How forecasts inform investment and portfolio decisions.

Subcategories
  • Ranking methodology
  • Risk-adjusted returns
  • Allocation and diversification
  • Rebalancing and costs

5 lessons with this primary category

Asset & Market Insights

What evidence reveals about particular assets and markets.

Subcategories
  • Equities
  • Crypto assets
  • Commodities
  • Currencies and indices
  • Market regimes

0 lessons with this primary category

Numbering revision · September 6, 2026

At Russlan Ramdowar’s request, two early lessons were inserted at #1 and #4 and existing display numbers were shifted once. Existing lesson IDs and URL anchors are unchanged. Subsequent lessons append to the ledger.

Showing 25 of 25 lessons

Newest lesson first

Technical Workflow & ProcessesAgent design

Lesson #25Different model families add a second dimension of research diversity

Multiple personas on one model are not a substitute for comparing opinions from different model families and providers.

Inspect evidence and consequence

What we observed

A single model family can reproduce shared assumptions across different investment frameworks. Batch 7 adds Claude Opus 5 and GPT-6 Astra alongside Gemini 3.8 Flash, including a matched Universal Investor researcher framework, to make provider-related differences and potential blind spots inspectable.

What changed

Retain model and framework provenance with every opinion, compare matched configurations, and withhold underlying model identities from the synthesis weighting prompt. Evaluate model diversity separately from the number of persona configurations.

Evidence scope

Lesson origin: Engine workflow

Completed Batch 7 collection: 426 assets, each with 12 Gemini Flash opinions, one Opus researcher and one Astra researcher; 5,964 total opinions.

Limits

  • — The 12:1:1 opinion panel is not balanced by provider and must not be described as three equal model votes.
  • — Different providers may share sources, training patterns and errors; diversity does not guarantee independence or eliminate bias.
  • — Improved forecast accuracy requires subsequent evaluation and cannot be inferred from provider count.
Methodology Methodology Methodology

Adoption and evidence

Implemented in Engine V7 / Batch 7 through matched researcher configurations across Google, Anthropic and OpenAI model families.

Incorporated since

Engine V7

Learned during

Batch 7

September 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Sep 21, 2026 Validated Sep 26, 2026
Forecasting ScienceForecast comparability

Lesson #24Consensus should weigh the evidence behind an opinion, not only count opinions

Equal weights are a transparent baseline, but a supported contrarian argument can deserve more influence than repeated weak reasoning.

Inspect evidence and consequence

What we observed

The Batch 7 synthesizer reviews shared premises, contradictions, evidence freshness and the strongest minority case using blinded model identities. It can cross-check material disputed claims when search is enabled and useful. Unsupported claims can receive less weight; unresolved uncertainty must stay explicit.

What changed

Publish explained per-opinion weights and the overall assessment. Normalize the weights and calculate the 20-quarter path deterministically from frozen source returns. Preserve the averaged baseline, minority views and the separate quantitative ranking history.

Evidence scope

Lesson origin: Engine workflow

Batch 7 production synthesis contract and publication validation, reviewed September 26, 2026. This is an implemented workflow control, not evidence of improved out-of-sample investment performance.

Limits

  • — The synthesizer can make errors and share source-model biases; a weighting rationale is not independent proof.
  • — Citations are not equivalent to verified facts. Failed or unavailable retrieval must remain visible.
  • — Reward supported distinctive reasoning, not disagreement alone. Equal weights remain valid when evidence does not justify differences.
Methodology Methodology Methodology

Adoption and evidence

Implemented in Engine V7 / Batch 7 after earlier batches established independent opinions and an averaged consensus baseline.

Incorporated since

Engine V7

Learned during

Batch 7

September 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Sep 21, 2026 Validated Sep 26, 2026
Technical Workflow & ProcessesEditorial governance

Lesson #23Editorial approval should review the evidence without pretending to author the model output

The model generates analysis; the quantitative layer ranks it; a human approver decides whether the current evidence supports publication.

Inspect evidence and consequence

What we observed

V6 added a downstream workflow that checks recent events, filings, and news, records material changes, and requires accountable approval for Editorial Top Picks.

What changed

Public pages identify Russlan Ramdowar as Editorial Approver, while iPulse AI remains the author/creator of the automated research artifact.

Evidence scope

Lesson origin: Engine workflow

Post-Batch 6 Editorial Top Picks workflow and publication contract.

Limits

  • — Editorial approval does not convert AI-generated research into personalized financial advice or guarantee forecast accuracy.
Methodology Methodology

Adoption and evidence

Added after the Engine V6 / Batch 6 forecast run as a downstream publication control; it did not alter the frozen model forecasts.

Incorporated since

Engine V6

Learned during

Batch 6

July 5, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Jul 17, 2026 Validated Sep 4, 2026
Investment & Portfolio MethodologyRanking methodology

Lesson #22Event risk, financial health, volatility, and disagreement are not one thing

Combining distinct risk dimensions into an unexplained single number makes the result easier to read but harder to trust.

Inspect evidence and consequence

What we observed

The v7.4 ranking architecture added event-risk pressure and equity financial health while keeping historical volatility and advisor disagreement as separate diagnostics.

What changed

Rankings and Asset Snowflake views label each dimension independently and document how it affects—or does not affect—the main score.

Evidence scope

Lesson origin: Research finding

Formula v7.4 and Asset Snowflake methodology.

Limits

  • — The displayed diagnostics remain simplified indicators, not complete measures of investment risk.
Methodology Methodology

Adoption and evidence

Implemented and validated in Engine V6 / Batch 6 as part of the v7.4 ranking and Asset Snowflake design.

Incorporated since

Engine V6

Learned during

Batch 6

July 5, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Jul 8, 2026 Validated Sep 4, 2026
Investment & Portfolio MethodologyRanking methodology

Lesson #21Formula architecture and scoring-run versions must not be collapsed

“v7.4” and “v2.5” identify different layers even when they appear on the same ranking record.

Inspect evidence and consequence

What we observed

The public formula label describes the scoring architecture, while the run version identifies the persisted pipeline implementation that executed it.

What changed

Both values remain visible in the engine history and attached to ranking lineage instead of being relabeled as an engine version.

Evidence scope

Lesson origin: Engine workflow

Formula changelog and production Batch 6 ranking rows.

Limits

  • — Legacy rows may carry older labeling conventions and require contextual interpretation.
Methodology Methodology

Adoption and evidence

Clarified in Engine V6 / Batch 6, where engine, formula-architecture, and scoring-run identifiers coexist.

Incorporated since

Engine V6

Learned during

Batch 6

July 5, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Jul 8, 2026 Validated Sep 4, 2026
Technical Workflow & ProcessesPublication and trust

Lesson #20Longer model output is not automatically more informative

Field-level word limits can improve comparability and reduce repetitive narrative without removing the underlying structured evidence.

Inspect evidence and consequence

What we observed

V5 and V6 shortened selected fields after public reports became too lengthy to scan consistently across thousands of forecasts.

What changed

Output schemas use more deliberate field budgets while retaining detailed drivers, risks, scenarios, and time-series data where they matter.

Evidence scope

Lesson origin: Engine workflow

Editorial and product review of Batch 5 outputs; revised before Batch 6.

Limits

  • — Aggressive compression can remove nuance, so field limits must be evaluated against information loss.
Methodology Methodology

Adoption and evidence

The output-density issue was observed in Engine V5 / Batch 5; field budgets were revised in Engine V6 / Batch 6.

Incorporated since

Engine V6

Learned during

Batch 5

June 3–5, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Jun 5, 2026 Validated Jul 5, 2026
Technical Workflow & ProcessesContext and inputs

Lesson #19Fundamental and market-price currencies must be normalized before valuation

A valid market capitalization divided by valid net income can still produce a meaningless P/E ratio when the two values use different currencies.

Inspect evidence and consequence

What we observed

Batch 5 introduced equity fundamentals but exposed reporting-currency and ticker-currency mismatches. Some ratios became implausible and could have distorted model context.

What changed

V6 aligned currencies, corrected P/E calculations, and hardened fundamental-data preparation before prompt assembly.

Evidence scope

Lesson origin: Research finding

Batch 5 post-run review and Batch 6 pre-run data correction.

Limits

  • — Currency normalization does not solve stale filings, accounting-policy differences, or negative earnings.
Methodology Methodology

Adoption and evidence

The currency mismatch was observed in Engine V5 / Batch 5 and corrected in Engine V6 / Batch 6.

Incorporated since

Engine V6

Learned during

Batch 5

June 3–5, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Jun 5, 2026 Validated Jul 5, 2026
Technical Workflow & ProcessesPublication and trust

Lesson #18Consistent expert language works best with definitions at the point of need

Mixing novice, intermediate, and expert voices makes comparisons uneven; expert terminology becomes accessible when the product supplies contextual definitions.

Inspect evidence and consequence

What we observed

V5 standardized output language at a subject-matter-expert level and paired it with glossary highlights instead of asking each persona to simplify differently.

What changed

The engine separates analytical vocabulary from presentation assistance, improving consistency without forcing readers to leave the report.

Evidence scope

Lesson origin: Engine workflow

Batch 5 language and glossary redesign.

Limits

  • — Definitions reduce friction but do not replace domain knowledge or professional advice.
Methodology Methodology

Adoption and evidence

Introduced in Engine V5 / Batch 5; retained and revalidated in Engine V6.

Incorporated since

Engine V5

Learned during

Batch 5

June 3–5, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Jun 3, 2026 Validated Sep 4, 2026
Forecasting ScienceForecast comparability

Lesson #17Independent reports become more useful when disagreement is synthesized—not erased

Readers need one asset-level overview that preserves the range of advisor views, central forecast, common drivers, and material dissent.

Inspect evidence and consequence

What we observed

Batch 4 made the absence of a summary layer costly for inspection. V5 added an asset-consensus workflow after the independent forecast layer.

What changed

Asset Consensus is now a distinct Engine output and a required precursor to cross-asset ranking and public insight generation.

Evidence scope

Lesson origin: Engine workflow

Batch 4 review and Batch 5 consensus implementation.

Limits

  • — Consensus can conceal minority insight unless dissent and voice counts remain visible.
Methodology Methodology

Adoption and evidence

The gap was observed in Engine V4 / Batch 4; Asset Consensus was added in Engine V5 / Batch 5 and retained in Engine V6.

Incorporated since

Engine V5

Learned during

Batch 4

May 3, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed May 3, 2026 Validated Sep 4, 2026
Technical Workflow & ProcessesAgent design

Lesson #16Critical instructions need deliberate placement in long prompts

A 50,000-plus-token context can dilute task goals unless essential instructions are compact, clear, and reinforced.

Inspect evidence and consequence

What we observed

Batch 4 used a prompt-sandwich pattern: the goal and key asset requirements appeared near both the beginning and end of the assembled prompt.

What changed

Prompt assembly now treats instruction position as a governed design choice rather than assuming the model weighs every token equally.

Evidence scope

Lesson origin: Engine workflow

Batch 4 prompt configuration and completion review.

Limits

  • — Instruction repetition must be controlled; excessive duplication can create the same salience distortion as repeated context events.
Methodology

Adoption and evidence

Introduced and observed in Engine V4 / Batch 4; instruction-placement controls remain active in Engine V6.

Incorporated since

Engine V4

Learned during

Batch 4

May 3, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed May 3, 2026 Validated Jul 5, 2026
Investment & Portfolio MethodologyRanking methodology

Lesson #15Expected return alone is not a sufficient cross-asset ranking

A comparable signal should account for forecast dispersion and historical volatility rather than rewarding the largest point estimate.

Inspect evidence and consequence

What we observed

The formula evolved toward confidence-adjusted excess return: a cash hurdle, median-absolute-deviation discount, volatility context, direction consistency, and a negative-return guard.

What changed

Formula v7.2 introduced the risk-aware structure and formula identifiers remain attached to every ranking snapshot.

Evidence scope

Lesson origin: Research finding

Scoring implementation history from the April audit through formula v7.2.

Limits

  • — Historical volatility is backward-looking.
  • — A compact score cannot represent every dimension of investment risk.
Methodology Methodology

Adoption and evidence

Risk-context work began in Engine V3 / Batch 3 and evolved through Engines V5–V6 / Batches 5–6.

Incorporated since

Engine V5

Learned during

Batch 3

April 11–12, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Apr 12, 2026 Validated Jul 17, 2026
Investment & Portfolio MethodologyRanking methodology

Lesson #14Dividend contribution needs a stable multi-year lookback

One unusual dividend should not dominate a forward ranking.

Inspect evidence and consequence

What we observed

A three-complete-year dividend window reduces sensitivity to a one-off historic payment while remaining recent enough to reflect changes in distribution policy.

What changed

The ranking pipeline uses three full historical years for eligible dividend contribution and records the applied formula with each snapshot.

Evidence scope

Lesson origin: Engine workflow

Ranking-methodology work initiated after Batch 3 and formalized in the v7 family.

Limits

  • — Past dividends do not guarantee future payments.
  • — Corporate actions and missing dividend records still require validation.
Methodology Methodology

Adoption and evidence

Work began with Engine V3 / Batch 3; the ranking rule was formalized across Engines V5–V6 / Batches 5–6.

Incorporated since

Engine V5

Learned during

Batch 3

April 11–12, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Apr 12, 2026 Validated Jul 17, 2026
Technical Workflow & ProcessesPublication and trust

Lesson #13A trustworthy forecast history includes configuration changes and mistakes

A latest-only result cannot show whether the process improved, regressed, or quietly changed its rules.

Inspect evidence and consequence

What we observed

The engine evolved materially across seven batches. Publishing old forecasts, component versions, formula lineage, limitations, and corrections makes those changes inspectable.

What changed

V6 treats the Forecast Ledger, Engine History, Lessons Learned, methodology changelog, and correction notices as one connected public trust system.

Evidence scope

Lesson origin: Engine workflow

Six execution batches, ranking-formula history, and the public forecast-history architecture.

Limits

  • — Public transparency excludes proprietary full prompts, private infrastructure, and data that cannot legally be redistributed.
Methodology Methodology Methodology

Adoption and evidence

Public-history requirements began with Engine V3 / Batch 3 and expanded into the connected transparency system visible in Engine V6 / Batch 6.

Incorporated since

Engine V3

Learned during

Batch 3

April 11–12, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Apr 11, 2026 Validated Sep 4, 2026
Technical Workflow & ProcessesResearch evidence

Lesson #12Giving a model a search tool does not guarantee meaningful research

Search availability and search execution must be measured separately.

Inspect evidence and consequence

What we observed

In the observed Gemini batch workflow, only a minority of tasks mentioned searches and an even smaller share cited web sources. Fluent output was therefore not reliable evidence of current-source retrieval.

What changed

Later prompts more explicitly require research behavior, while supplied context and source provenance remain independent safeguards.

Evidence scope

Lesson origin: Engine workflow

Batch 3 operational review; approximately 40% of sampled tasks mentioned search queries and roughly 4% cited website sources.

Limits

  • — The percentages describe that observed configuration and should not be generalized to all models or current provider behavior.
Methodology Methodology

Adoption and evidence

Measured in Engine V3 / Batch 3; no later batch is listed because this lesson records that specific tool-use observation.

Incorporated since

Engine V4

Learned during

Batch 3

April 11–12, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Apr 11, 2026 Validated Apr 12, 2026
Technical Workflow & ProcessesContext and inputs

Lesson #11Global context should be broad enough to inform without hijacking the thesis

Repeating one dramatic event too heavily can make models over-attribute unrelated asset outcomes to that event.

Inspect evidence and consequence

What we observed

Batch 3 showed that adding current context solves one failure mode but creates another if a single conflict or macro event dominates the token budget and recency cues.

What changed

Context assembly now favors balanced world-state coverage, temporal proximity, and non-repetitive event representation.

Evidence scope

Lesson origin: Engine workflow

Batch 3 global-context prompting and output review.

Limits

  • — Truly systemic events may deserve disproportionate weight; balance does not mean equal weighting.
Methodology Methodology

Adoption and evidence

Observed in Engine V3 / Batch 3; balanced context assembly remains active in Engine V7.

Incorporated since

Engine V4

Learned during

Batch 3

April 11–12, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Apr 11, 2026 Validated Jul 5, 2026
Investment & Portfolio MethodologyRanking methodology

Lesson #10A frozen forecast and a later ranking of that forecast are different records

Improving a scoring formula should create a new traceable ranking snapshot, not rewrite the forecast that existed before the formula.

Inspect evidence and consequence

What we observed

Rankings began after Batch 2 and older batches were scored retrospectively. Production can therefore show a modern formula on an older batch without claiming that the formula generated the forecast.

What changed

Engine, batch, formula architecture, scoring-run implementation, and recompute date are tracked as separate identifiers.

Evidence scope

Lesson origin: Engine workflow

Repository history and production leaderboard metadata from the first March 27 implementation through v7.4.

Limits

  • — A re-score is only comparable when its source forecast cohort and eligibility rules remain explicit.
Methodology Methodology Methodology

Adoption and evidence

Applies across Engines V1–V7 / Batches 1–7: historical forecasts stay frozen while later ranking snapshots receive separate formula and run versions.

Incorporated since

Engine V1

Learned during

Batch 2

March 19, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Mar 27, 2026 Validated Sep 4, 2026
Technical Workflow & ProcessesAgent design

Lesson #9A memorable persona is not automatically a useful analytical framework

Personas should encode inspectable investment reasoning—not ambiguity that the model resolves as theatrical style.

Inspect evidence and consequence

What we observed

The Nostradamus experiment produced confused output because the model could not consistently distinguish literal astrology from metaphorical forecasting language.

What changed

Nostradamus was deprecated after Batch 2, and persona designs are now evaluated for analytical specificity and repeatability.

Evidence scope

Lesson origin: Engine workflow

Batch 2 persona outputs.

Limits

  • — This finding applies to the tested configuration, not to every unconventional analytical signal.
Methodology Methodology

Adoption and evidence

Observed in Engine V2 / Batch 2; the ambiguous persona was removed before Engine V3.

Incorporated since

Engine V3

Learned during

Batch 2

March 19, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Mar 19, 2026 Validated Apr 11, 2026
Technical Workflow & ProcessesAgent design

Lesson #8Cross-asset comparisons require consistent advisor coverage

An influential persona cannot be applied selectively without changing the meaning of cross-asset scores.

Inspect evidence and consequence

What we observed

Batch 2 made the effect visible: a structurally skeptical persona such as Michael Burry can pull a consensus lower. Assets reviewed by that persona are not directly comparable with assets that omit the same voice.

What changed

Comparable cohorts use governed advisor coverage and preserve voice counts and configuration lineage with ranking outputs.

Evidence scope

Lesson origin: Engine workflow

Batch 2 persona coverage and subsequent consensus/ranking design.

Limits

  • — Missing or failed tasks can still reduce realized coverage; this must remain visible rather than silently imputed.
Methodology Methodology

Adoption and evidence

Observed in Engine V2 / Batch 2; governed advisor coverage remains enforced in Engine V7.

Incorporated since

Engine V3

Learned during

Batch 2

March 19, 2026

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Mar 19, 2026 Validated Sep 4, 2026
Technical Workflow & ProcessesEditorial governance

Lesson #7A useful summary leads with the asset insight, not framework narration

Repeated phrases such as “the frameworks agree” consume attention without adding decision value.

Inspect evidence and consequence

What we observed

Early SEO and GEO summaries described the mechanics of advisor agreement more prominently than the underlying finding. That made the writing feel mechanical and reduced information density.

What changed

Summaries now state supported findings directly and call out disagreement only when the disagreement itself is material.

Evidence scope

Lesson origin: Engine workflow

Editorial review of Batch 1 summary outputs and later public asset-summary iterations.

Limits

  • — Methodology context remains necessary where a score or disagreement measure would otherwise be ambiguous.
Methodology Methodology

Adoption and evidence

Observed in Engine V1 / Batch 1; the revised summary standard remains active in Engine V7.

Incorporated since

Engine V2

Learned during

Batch 1

November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Nov 28, 2025 Validated Sep 4, 2026
Forecasting ScienceForecast comparability

Lesson #6One complete master horizon is more reusable than disconnected forecast paths

A complete five-year forecast path can support consistent one-, three-, and five-year extracts more effectively than separate horizon runs with missing periods.

Inspect evidence and consequence

What we observed

The Batch 1 three-year and five-year experiment produced similar directional results, but the shorter path could not support full cross-horizon comparison and duplicated execution effort.

What changed

Later releases standardized the long path and derive shorter views where the public product supports them.

Evidence scope

Lesson origin: Engine workflow

Comparison of Batch 1 three-year and five-year configurations.

Limits

  • — A dedicated short-horizon model may still be justified if its objective or data frequency is materially different.
Methodology

Adoption and evidence

Observed in Engine V1 / Batch 1; the master-horizon approach began in Engine V2 and remains the current Engine V7 standard.

Incorporated since

Engine V2

Learned during

Batch 1

November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Nov 28, 2025 Validated Jul 17, 2026
Forecasting ScienceForecast comparability

Lesson #5Forecast comparison begins with one standardized anchor date

Advisor and cross-asset forecast paths are not fairly comparable when their time-series anchors drift.

Inspect evidence and consequence

What we observed

Batch 1 produced slightly shifted dates that required expensive manual alignment and repeated backup-and-repair work. The problem was structural rather than cosmetic.

What changed

Later output schemas enforce anchor and horizon-step rules before forecasts enter consensus and ranking layers.

Evidence scope

Lesson origin: Engine workflow

Observed in Batch 1 time-series outputs and addressed in the Batch 2 schema redesign.

Limits

  • — Market calendars and asset-class trading schedules can still require explicit date-normalization rules.
Methodology Methodology

Adoption and evidence

Observed in Engine V1 / Batch 1; corrected through the Engine V2 / Batch 2 output-schema redesign.

Incorporated since

Engine V2

Learned during

Batch 1

November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Nov 28, 2025 Validated Jul 5, 2026
Forecasting ScienceForecast targets and transformations

Lesson #4Forecast percentage changes from a defined anchor, rather than absolute prices

A fixed anchor and a path of step-over-step percentage changes make forecasts easier to compare, reconstruct and adjust for stock splits without rewriting the forecast returns.

Inspect evidence and consequence

What we observed

The first batch forecast absolute closing prices. Large numerical scales, such as Bitcoin prices, made the representation less convenient for LLM reasoning and cross-asset comparison in our workflow. Modelling relative changes removes dependence on the absolute price level and is more suitable for assessing time-series stability than assuming raw price levels are stationary. Absolute price targets also become harder to display and evaluate consistently after stock splits unless every value uses the same share basis.

What changed

Batch 1 price paths were transformed into step-over-step percentage changes, preserving their implied forecasts rather than generating new predictions. Later runs requested that representation directly. For an anchor P0 and percentage changes r1 through rt, reconstruct the price as P0 × product(1 + ri/100). Preserve the original anchor and derive a split-adjusted display anchor when required; the stored return path stays unchanged. Evaluate against actual prices on the same split-adjusted basis, with the adjustment provenance retained.

Evidence scope

Lesson origin: Engine workflow

Batch 1 target-design and conversion history, retrospectively confirmed by Russlan Ramdowar on September 6, 2026; the current forecast schema stores step-over-step percentage changes and anchor metadata.

Also relevant to: Technical Workflow & Processes. Search by these category names to find cross-topic lessons.

Limits

  • — Percentage returns are not automatically stationary. Changing volatility, market regimes and structural breaks still require assessment.
  • — Easier handling of large price scales is an internal workflow observation, not a controlled benchmark showing a universal LLM accuracy gain.
  • — Anchor rescaling addresses a consistent stock-split basis; dividends, spin-offs, mergers, currency changes and other corporate actions need separate treatment.
  • — Percentage returns are undefined at a zero anchor and may be unsuitable for instruments with zero or negative price levels. Conversion must preserve rounding and transformation provenance.
Methodology Methodology

Adoption and evidence

Learned from Batch 1, which requested absolute closing-price forecasts. Those outputs were subsequently converted to step-over-step returns; the percentage-change target was adopted for subsequent runs from Engine V2 / Batch 2.

Incorporated since

Engine V2

Learned during

Batch 1

November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Nov 28, 2025 Validated Sep 6, 2026
Technical Workflow & ProcessesContext and inputs

Lesson #3Current context must be supplied or deliberately retrieved

An API model call should never be assumed to have the browsing behavior or current awareness of a consumer chat application.

Inspect evidence and consequence

What we observed

The earliest runs exposed a gap between familiar app experiences and raw API execution. Model knowledge cutoffs and inconsistent search-tool use can leave recent events absent from otherwise fluent analysis.

What changed

The engine added governed global-context packages from Batch 3 and records whether each input component was actually used.

Evidence scope

Lesson origin: Engine workflow

Batch 1 revealed the issue; Batch 3 was the first verified run with global context.

Limits

  • — Tool-use behavior varies by model, provider, mode, and date.
  • — Supplied context still requires source quality and scope controls.
Methodology Methodology

Adoption and evidence

Observed in Engine V1 / Batch 1; corrected in Engine V3 / Batch 3 when governed global context was introduced; revalidated as an active control in Engine V6. Batch 2 is omitted because it was neither a new evidence point nor the corrective milestone.

Incorporated since

Engine V3

Learned during

Batch 1

November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Nov 28, 2025 Validated Sep 4, 2026
Technical Workflow & ProcessesLLM prompt design

Lesson #2Bull and bear labels in LLM prompts can manufacture the outcome they are meant to test

An LLM forecasting prompt should ask the model to evaluate plausible upside and downside scenarios without instructing it to reach a bullish or bearish conclusion.

Inspect evidence and consequence

What we observed

Batch 1 showed that explicit bullish or bearish role labels inside the LLM prompt pulled the primary forecast toward the assigned label. The prompt was effectively telling the model which conclusion to defend, turning scenario analysis into an instruction-following artifact rather than a probability-weighted estimate.

What changed

Directional labels were removed from LLM persona prompts. The prompt may assign skeptical, growth-oriented, or risk-aware analytical perspectives, but it must ask every model to identify the most likely scenario from evidence rather than defend a predetermined direction.

Evidence scope

Lesson origin: Engine workflow

Observed in the Batch 1 LLM prompt design and used to redesign persona instructions and prompt assembly in later engine versions.

Limits

  • — This is an LLM prompt-design observation, not a claim that every directional instruction produces the same magnitude of bias across all models or assets.
Methodology Methodology

Adoption and evidence

Observed in Engine V1 / Batch 1; persona-prompt changes began in Engine V2; the neutral scenario control remains enforced in Engine V7.

Incorporated since

Engine V2

Learned during

Batch 1

November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed Nov 28, 2025 Validated Sep 4, 2026
Technical Workflow & ProcessesData identity and quality

Lesson #1A ticker is a changeable label, not a permanent asset identity

Keep a stable internal instrument identity separate from exchange-qualified ticker aliases and readable public URL slugs.

Inspect evidence and consequence

What we observed

While designing the first batch, we recognized that stock tickers can change and may be reused; they are not globally unique across exchanges or time. Meta changed its ticker from FB to META in June 2022. Victoria’s Secret illustrates a different identity event: its 2021 spin-off created a separately listed security, VSCO, rather than simply renaming the parent ticker. Neither event should be handled by blindly replacing a database primary key.

What changed

The design separates permanent internal asset IDs from ticker labels, exchange and listing information, and historical aliases. Corporate actions must distinguish continuity of an existing instrument from creation of a new one. Public company-and-ticker slugs remain readable and stable; if a canonical URL changes, a permanent redirect, updated internal links and a consistent canonical preserve access and help consolidate search signals. A ticker change alone does not require a URL change.

Evidence scope

Lesson origin: Engine workflow

Batch 1 design experience, retrospectively confirmed by Russlan Ramdowar on September 6, 2026. The public company announcements below illustrate ticker changes and spin-offs; they do not establish the date of the internal implementation.

Also relevant to: Asset & Market Insights. Search by these category names to find cross-topic lessons.

Limits

  • — An internal ID must identify the intended instrument or listing, not conflate a company, multiple share classes and exchange listings.
  • — Readable URLs are supported by Google guidance, but including a ticker is not a ranking requirement or guarantee.
  • — Spin-offs, mergers and delistings need explicit corporate-action relationships; a parent URL must not automatically redirect to a different security.
Methodology Methodology

Adoption and evidence

Discovered during the design of Batch 1, before its November 28, 2025 run. Stable asset identity was a foundational design requirement, not a finding produced by the forecasts. The exact discovery date was not recorded.

Incorporated since

Engine V1

Learned during

Batch 1 design · before the run

Before November 28, 2025

The engine marker shows when this lesson became an operating control. The learning origin identifies the design phase or batch review that first produced it.

First observed in Batch 1 design; exact date not recorded Validated Sep 6, 2026