00 / ABSTRACT
Confidence is an output,
not an input.
This note documents an engineering approach to decision systems: separate what was observable from what was inferred, price the action under explicit execution assumptions, and keep strategy selection apart from evaluation. The examples come from MT5 validation work and memecoin launch-risk research. They illustrate correctness issues; they do not establish a profitable strategy.
The central result is methodological: a favorable metric cannot repair a broken observation boundary. Statistical correction, protocol decoding, and execution modeling answer different questions. Each needs its own evidence.
01 / THE EXECUTION BOUNDARY
A stop price is
not a fill guarantee.
A candle can open beyond a standing stop. Filling the simulated exit at the stop hides the intervening loss. The corrected evaluator uses the opening executable quote when a stop has been crossed; it then considers the unordered high and low. A target already reached at the open takes precedence over a later reversal, with favorable price improvement capped.
The open is below the stop. The additional loss reaches the return and drawdown calculations.
Illustrative zero-spread example. When the opening price is above the stop, this fixture assumes the bar subsequently touches the stop. R is the result divided by the original 5-unit risk.The research evaluator remains an OHLC approximation. A BID-versus-midpoint source mismatch, additional slippage, commissions, and swaps require separate treatment. MetaQuotes documents that bar zero is still current, which is why completed-bar selection is another independent requirement. [1]
02 / THE SEARCH BOUNDARY
Search harder.
Demand more evidence.
Ranking hundreds of rules on one data window creates a selection problem. A confidence interval from the winning row does not describe an independently selected strategy. The Deflated Sharpe Ratio literature addresses selection bias and non-normal returns; the Probability of Backtest Overfitting framework examines the reliability of strategy selection. Neither establishes that a feature was available at decision time. [2] [3]
≈ 0.103 draws per tail
A two-sided Bonferroni percentile interval at this resolution is estimated from effectively fewer than one expected draw in each extreme tail. Interpolating a minimum does not create missing resolution.
The revised research screen withholds that interval. Its local numerical rule requires at least 20 expected tail draws and 10 launch-time clusters; those thresholds are diagnostics, not a theorem guaranteeing coverage. Same-window rankings remain exploratory. Repeated nightly searches and earlier discarded trials also matter.
A 180-second per-launch recorder cannot evaluate a 170-second policy entered 30 seconds after creation. Requiring the complete policy horizon avoids keeping only the early winners. Censoring does not prove the recorder had uninterrupted coverage.
03 / THE OBSERVABILITY BOUNDARY
Unknown is a state.
It is never a clean bill.
A launch-risk review should distinguish missing data, observed controls, and modeled exit constraints. Mint and freeze authorities represent capabilities; Solana also defines extension-specific authority types. An active authority or concentrated wallet cluster does not prove fraudulent intent. [4]
BLOCKED · insufficient real reserves
Virtual reserves: 30 SOL / 1,000,000 tokens. Exit: 50,000 tokens. Assumed sell fee: 1.25%; fixed fee: 0.0005 SOL. This simplified illustration covers reserves and observation status only. The full offline assessor also checks authorities, extension coverage, concentration, creator history, quote age, and impact. It never authorizes execution.The gross curve expression is x × q / (y + q), where x and y are virtual reserves and q is the exit quantity. Available real reserves cap the displayed bound. A real transaction may fail instead of filling partially. Priority fees, router charges, transaction ordering, and competing trades must be supplied or modeled separately.
The venue's help page currently lists a 1.25% pre-graduation trade fee. Historical replays need their historical fee regime. The laboratory now records explicit venue, router, network, and early-entry tip assumptions rather than presenting a universal cost. [5]
04 / PROTOCOL REVERSE ENGINEERING
Decode the structure.
Preserve the uncertainty.
Start from the program's published interface definition, then map binary fields into a normalized observation. Verify the event discriminator, field widths, byte order, boolean encoding, string length, and truncation behavior. A parser that returns plausible numbers for malformed bytes contaminates every downstream result.
- IdentifyPin the public interface revision and event discriminator.
- DecodeValidate each consumed field and reject malformed prefixes.
- BoundDescribe unparsed tails, quote denomination, commitment, and observation gaps.
- ReproduceKeep synthetic byte fixtures, truncation cases, and a regression trace.
This is compatibility and observability work on a public protocol. It does not recover identities behind wallets or establish undisclosed information. Evolving quote assets and appended event fields require explicit schema support; a legacy SOL scale must not silently stand in for another asset. Public Pump interface definitions ↗
05 / THE RESEARCH SEQUENCE
A repeatable path
from idea to assessment.
Record the hypothesis, parameter family, cost regime, rejection criteria, and evaluation dates before looking at the next window.
Preserve receipt time and source provenance. Distinguish unknown coverage from an observed absence of activity.
Create a counterexample that fails under the old rule. Test chronology, units, boundaries, gaps, and invalid inputs.
Evaluate adverse fees, liquidity constraints, delayed fills, and ambiguous ordering. Keep the losses in the record.
Select on one window. Freeze the selected configuration. Evaluate on independent later data without retuning.
Compare prospective decisions with completed outcomes and observed fills. Count exclusions and unknowns before reporting performance.
06 / PRIMARY SOURCES
Read the originals.
- MetaQuotes. copy_rates_from_pos.
Bar indexing and returned fields. Accessed September 10, 2026.
- Bailey & López de Prado (2014). The Deflated Sharpe Ratio.
Selection bias, backtest overfitting, and non-normality.
- Bailey, Borwein, López de Prado & Zhu. The Probability of Backtest Overfitting.
Reliability of strategy selection under repeated search.
- Solana. Set Authority.
Mint, freeze, and extension-specific authority roles.
- Pump.fun. Transaction Fees.
Current help-page schedule; historical and routing costs require separate evidence.
The linked papers are third-party research. This page is an original engineering synthesis and interactive explanation, not a reproduction of those papers or a peer-reviewed publication.