Public research record
Methodology
How public evidence becomes a dated, reviewable research record – and where the record deliberately stops.
What this record is
Research, not instruction
TrydingDay publishes systematic market research and historical outcome tracking. Each episode preserves the evidence available at the time, the editorial reasoning, the uncertainty and the later observed outcome. It is not personalised investment advice, portfolio management or trade execution.
Public research flow
One traceable sequence
System architecture
Two complementary analytical layers
TrydingDay's complete deliberative workflow uses frontier GPT-5.6 models across evidence review, candidate comparison, Desk deliberation, independent auditing and final synthesis.
Alongside these language-model processes, the system uses a separate machine-learning component based on LightGBM, called Candidate Horizon ML. It is trained and evaluated against a two-year historical backtesting base built from time-stamped US-market observations. Its current factual dataset contains nearly 80,000 records and is extended and retrained daily as new trajectories mature. The component contributes reproducible quantitative context, together with historical risk, similarity and model-explanation information.
Five analytical lenses
Different questions, one public record
Statistical review
Checks whether the public claim is supported by a suitable comparison and clearly stated uncertainty.
Information and catalysts
Examines what changed in the public company or information record and whether it explains the move.
Data integrity
Checks availability, timestamps, source lineage and whether the evidence was actually available at the time.
Risk and execution
Considers liquidity, gaps, costs and the difference between an observed path and an executable result.
Market structure and timing
Reads observable price, volume and timing context without treating continuation as guaranteed.
The lenses are AI-assisted analytical roles, not five human advisers and not independent trading signals. Their disagreement and uncertainty remain part of the record.
Data and availability
What can be checked
The public record uses time-stamped market observations, public company information and recorded sources. Availability depends on the source and the capture window. Missing information is disclosed as a limitation; it is not automatically treated as negative evidence.
Editorial states
A shared vocabulary
Desk deliberation
The complete candidate set
The process begins with the complete dossier of candidates selected for Desk consideration. It preserves their factual parameters, public evidence, sources, time context, uncertainty and other information available at the time of review.
Frontier GPT-5.6 models analyse the candidate set and develop the Desk deliberation. Candidate Horizon ML contributes an additional machine-learning parameter to that deliberation. It does not replace the Desk reasoning, operate as an automatic filter or decide by itself which candidate should be selected.
Independent audit
A second review before publication
After the initial Desk deliberation, an independent audit subagent is instantiated as a separate GPT-5.6 process. It is not another voice in the initial deliberation; it is a separate analytical layer.
The subagent receives the complete dossier: all candidates selected for Desk consideration, all of their factual parameters and evidence, the recorded reasoning, the provisional resolution and the additional parameters generated by Candidate Horizon ML. It therefore reanalyses the full candidate set and the resolution applied to that set, rather than reviewing only the case eventually selected.
The audit checks whether the provisional Desk resolution remains coherent when confronted with the complete available information and the independent machine-learning evidence. It checks for contradictions, overlooked risks, unsupported exclusions, differences between candidates and material changes in interpretation.
The audit is outcome-blind: it does not receive the later market outcome and does not attempt to explain it retrospectively. If it identifies a material disagreement, it may reopen the AI deliberation before the process is closed. No publication has been made at this stage. If the audit reopens the discussion, the new deliberation and its result remain traceable alongside the earlier resolution.
Human review
The decision before publication
The complete AI deliberative and audit process is followed by human review. The human decision takes place before publication and may approve the final resolution, request clarification or decline publication.
The human technical supervision behind this final step is described in the About and Editorial Data Integrity Policy.
AI limitations
Useful analysis can still be wrong
Frontier GPT-5.6 models can omit information, misunderstand a source, produce an inconsistent interpretation or carry an error forward. The LightGBM-based machine-learning component can also reflect gaps in historical coverage, changing market regimes or model error. Using both layers improves the opportunity to challenge a conclusion, but no model, subagent or review guarantees accuracy, completeness or future performance.
Editorial selection
A transparent boundary
TrydingDay publishes the evidence, reasoning, uncertainty, material updates and observed outcomes needed to evaluate each research record. Not every internal input, intermediate calculation or implementation detail becomes a separate public section. These decisions are made for editorial relevance, clarity and reader usability, not to conceal material evidence or prevent scrutiny.
Internal processes may therefore contain more intermediate detail than the public page, while every material fact, limitation, review, correction and outcome remains part of the traceable record.
Multi-engine benchmark
Quality checks across several processes
The system has been benchmarked across several analytical processes on comparable public cases. The benchmark checks consistency, evidence handling, uncertainty recognition and stability of conclusions. It is a quality check, not a vote between models, a performance guarantee or a disclosure of every production detail.
Limitations and observed outcomes
Read results in context
Research records report sample size, coverage, median, uncertainty, benchmark context and known data limitations. Observed paths are not guaranteed fills: liquidity, gaps, fees, slippage, taxes, timing and market access can change real-world outcomes.
Versions and changes
A public change record
Material changes to the public methodology are dated and described at a general level. Earlier episodes retain the version and evidence context that applied when they were published.
Audit links