Research — Mithraeum Agora

Stage one and two

Designing and testing

The strategy builder and the backtest engine. This is where an idea becomes something precise enough to be wrong.

The indicator registry

A hundred and thirteen indicators are declared in a single registry, and everything downstream reads from that one place: the builder's dropdown, the engine that computes the series, the mining grammar that assembles candidates, and the screens that rank a universe. Since Round 540 every one of them can be drawn by the search as well as chosen by hand — none is reachable only from the builder.

That is an architectural decision with a practical consequence — adding an indicator is one declaration rather than an edit in six files that must agree. The registry carries each indicator's parameters, its warm-up requirement and how it is drawn, so a new entry arrives complete.

The families

11 groups · 113 indicators

Trend is price going somewhere, measured over a windowdirection

Moving averages of several kinds, directional strength, slope and channel position. The family everything else tends to be defined against — most "is this working" questions reduce to whether a trend measure agrees with the position.

The trap: every trend measure lags by construction. A faster one lags less and whipsaws more, and no parameter search removes that trade-off — it only moves it.

Momentum is the rate of change, not the directiondirection

Rate-of-change over a lookback, oscillators, and relative measures that compare an instrument's move to its own recent distribution.

The trap: momentum and mean reversion are the same measurement read with opposite signs. A strategy that mixes both without a regime gate is usually two strategies fighting.

Mean reversion is distance from a centredirection

Bands, z-scores against a rolling mean, and stretch measures that ask how unusual the current position is relative to recent history.

The trap: works beautifully until it does not, and the failure is unbounded. Reverting to a mean assumes a mean exists.

Volatility sizes everything elserisk

Realised volatility over several windows, true range, and ratios between fast and slow estimates that indicate whether conditions are changing.

Used less as a signal than as a denominator: position size, stop distance and regime gates are all commonly expressed in volatility units so they adapt without a parameter change.

The trap: volatility clusters. Yesterday's estimate is a good predictor of today's and a poor one of next month's.

Volume asks whether anyone was thereconfirmation

Absolute and relative volume, volume-weighted prices, and participation measures used mostly to confirm or veto a signal from another family.

The trap: volume conventions differ by provider and instrument. A rule tuned on one source can mean something different on another — which is one reason the data-source selector has no fallback.

Breadth looks across a universe, not at one instrumentcross-sectional

How many members of a universe are advancing, how many are above their own trend, and how concentrated the movement is. A market where three names carry the index behaves differently from one where everything rises.

The trap: breadth needs the whole universe loaded and aligned in time. It is the family most easily corrupted by a gap in one member's history.

Relative strength ranks members against each othercross-sectional

An instrument's performance measured against its peers rather than against zero. This is the family cross-sectional rotation is built from — the rank is the signal.

The trap: a rank computed on data that was not yet available is the easiest lookahead bug to introduce and the hardest to see, because the result simply looks good.

Regime classifies conditions rather than predicting themcontext

Labels a period as calm, stressed, trending or ranging using volatility, breadth and trend measures together. Used as a gate: a strategy can be permitted to act only in the conditions it was designed for.

The trap: a regime label is itself a model with parameters, so a strategy gated on one has more fitted parameters than it appears to.

Seasonality is calendar structurecontext

Day of week, month, position within the month, and proximity to known calendar events.

The trap: the family most likely to produce pure coincidence. With enough calendar slices something will always look significant — which is exactly what the empirical null exists to measure.

Statistical describes the distribution itselfshape

Rolling correlation, dispersion, skew, autocorrelation and distribution-shape measures. Used to characterise behaviour rather than to time entries.

Rolling regression belongs here too — slope, goodness of fit and a channel around the line — computed about the window's own mean, so a series trading at a high price level cannot destroy the precision of the fit it is measuring.

The trap: these need long windows to be stable, and a long window means a long warm-up before a strategy may honestly trade at all.

Structure is swings, pivots and the shape of the tapeshape

Swing highs and lows, higher highs and lower lows, and how many bars ago the last one printed. A pivot cannot be known at the bar that makes it — it needs the bars after it to confirm it — so every structure measure publishes a pivot on the bar that confirms it, while its age still counts from where it actually printed.

The trap: the popular implementations mark the pivot on the bar that made it, which reads the future by exactly the confirmation window and produces some of the most convincing backtests there are.

Indicators you write yourself

Custom indicators are written as formulas — expressions over price, volume and other indicators — and evaluated in a sandbox that parses the formula rather than executing it as code. There is no path by which a formula can reach the filesystem, the network, or anything else in the process.

They register into the same registry as the built-ins. A formula you invented appears in the same dropdowns, computes through the same series machinery, can be mined over, and can be gated on — with a single branch in the code rather than a parallel path. There is no second-class citizenship for your own work.
They are saved with your strategies, by reference. A strategy that uses one records which one, so the meaning of the strategy cannot drift if you later build something similar.

The custom indicator editor holding a demonstration formula called Lamp ratio: two parameters, two intermediate variables built from ema and sma, one output expression, and a check-and-preview section pointed at a synthetic series. — **Writing an indicator.** A demonstration formula — two parameters, two intermediate steps, one output — with the function reference along the right. The check states what the formula will draw and its own warm-up before anything is plotted, and the preview runs it over the whole of a synthetic series. The formula is parsed, never executed as code, and once saved it appears in every dropdown beside the built-ins.

Strategies are trees, not scripts

An entry is a set of conditions, grouped, combined with AND and OR, and optionally gated on a regime. Same for the exit. Each condition compares something to something — an indicator to a level, an indicator to another indicator, a value to its own history.

A strategy the way the system holds it. Groups are operators, leaves are comparisons, and the regime gate sits above both. Nothing here is code you wrote — which is why the search can mutate it, the ledger can attribute it, and the logic view can tell you which leaf was dead weight.

A vocabulary for what a strategy can say

A comparison and an AND is enough to express a surprising amount, and not enough to express a great deal of what traders actually mean. Round 506 onward added the missing grammar — each piece off by default, so a strategy written before it existed means exactly what it always meant.

At least K of N votesgrammar

A group can fire when any two of its three conditions hold, rather than all or any. The logic tree expands a vote into the exact pairs it means, so the drawing is always the strategy the engine runs — and a group's combine mode is always stated on the group, never implied by the joins between its rows.

Guards against: a vote that, drawn as a plain list, reads as a stricter strategy than the one being run.

Weighted scores with a thresholdgrammar

Conditions carry weights and the group fires above a threshold — a scorecard rather than a checklist, drawn as the minimal combinations that clear the bar.

NOT, applied before the missing-data maskgrammar

Any condition can be negated. The negation is applied before the engine masks out bars it cannot evaluate, on the backtest path and the live path alike, so an input that is missing — a warm-up bar, an absent peer, an untrained model — never reads as true merely because it was negated.

Guards against: a dead operand quietly firing on every bar the moment someone puts NOT in front of it.

Lags and "how long ago"grammar

An operand can read its value N bars ago, and a condition can ask how many bars have passed since something last held — the cool-off, the "only after a quiet week" that traders reach for constantly.

A calendar and a clock, and what they refuse to answergrammar

Day of week, month, position in the month, minutes from the session's open. On a daily bar there is no time of day, so the clock answers unknown rather than zero — a condition on it never fires on a daily tape rather than firing on every bar.

An indicator on another symbolgrammar

The RSI of a second instrument can be an operand in the first one's strategy — the classic "only trade this when that is calm". When the second series is stale or missing at a bar, the condition cannot be evaluated, and a live deployment holds its orders rather than trading on an input it does not have.

Adding to a position, sizing by risk, reversingexecution

A strategy may add to a winning position up to a stated limit; the entry price every stop and target reads becomes the size-weighted average, and the position's age does not reset — a pyramid that restarted the clock could keep a losing trade alive by adding to it. Size can be set from the stop's distance instead of a flat fraction, and a signal in the opposite direction can reverse the position in one step.

Guards against: three execution facts the engine used to leave implicit — each a run setting now, stated, and off unless chosen.

The append-invariance gate asks the whole registryproof

Every indicator and every operand is computed twice — over a prefix of the tape, and over all of it — and every bar the two share must be identical. A value that changes when later bars arrive was reading them. The gate covers whole engine runs too: a trade booked on a prefix must survive the arrival of the rest.

Guards against: the lookahead bug, in the one form that can be checked mechanically for every indicator at once.

The strategy builder holding the demonstration regime strategy: the test settings, six indicator rows, and the Risk-off regime card open — its gate condition on the 200-day average, an entry on RSI below 32, and a three-clause exit including a stop expressed as entry price minus a multiple of ATR. — **The builder, holding a demonstration strategy.** Test settings across the top — capital, costs, leverage, execution, the stop, trail and take-profit fields, how many times a position may be added to, how it is sized, and the funding terms — then the indicators, then a regime card gating which plan is active. The amber *no stop* chip is the builder saying out loud that none of the protective fields is set. Captured against the synthetic series *SYN-BOREAS*.

The regime performance panel for the demonstration run: the regime equity overlay with its background banded green and red by which regime owned each stretch, above a persistence table asking whether each regime is a regime at all — spells, median spell length, longest spell and share of bars, with verdicts reading flickering, shorter than a trade, and persists. — **A gated strategy, split by the gate — and asked whether the gate is a gate.** The equity curve is banded by which regime owned each stretch. Beneath it, before any per-regime figure, the panel asks the question that decides whether those figures mean anything: does each regime *persist*? A regime whose median spell is three bars is *flickering*; one shorter than the trades it opens is deciding nothing it gets credited with. Only the default state here persists. Synthetic series *SYN-BOREAS*; the verdicts are the tool's, not a result.

Because the tree is data, it can be inspected after the fact. The logic view shows which conditions actually fired, how often each one contributed, and which were effectively dead weight — a condition present in every trade is not doing any work, and one that never fires is not either.

Know your strategy

The application used to draw a strategy five different ways in five places, and the list of saved strategies had the poorest of them: a median plan of eighteen conditions across four regimes shown as six flat lines of code. A window now renders the whole plan as a tree, with how often each condition fired beside it — and it opens not on the conditions but on what can end a trade.

That ordering is the point. Stops, trails, take-profits and time limits all close a position with no condition true, and no logic tree in the application had ever mentioned them. When the window was first pointed at the saved strategies, more than half had no stop of any kind, a few could only ever be closed by a reversal, and a handful could never trade at all. It says so, in a sentence, first.

The know-your-strategy window for a demonstration strategy on SYN-DELPHI: a one-sentence summary of the whole plan, then tabs — how a bar is decided, every condition, what it is made of, what might be wrong, what did anything — open on the first, which shows the one plan owning every bar, its entry and exit, and the exits that fire with no condition true: an exit condition, or a stop at six percent. — **A strategy, explained before it is shown.** One sentence says what the plan does and what can end its trades; the first tab draws how a bar is decided, ending in the exits no condition controls. This demonstration strategy is written in the new vocabulary — any two of three conditions, one of them negated, one read off a second synthetic series.

The every-condition tab: the entry drawn as an at-least-two-of-three vote expanded into its three pairs, each pair an AND with the negated condition marked NOT, and the exit as two alternatives. — **A vote, drawn as what it means.** "At least two of three" expanded into the three pairs that satisfy it, with the negated condition marked where it sits. The tab also says, plainly, when a run stored no firing statistics — which is a different fact from a condition that never fired.

The what-it-is-made-of tab: counts of conditions, indicator-driven conditions, price-action conditions, indicators, plans, regime depth and negations, and which parts of the vocabulary the strategy reaches for — one vote group. — **A census of the plan.** How many conditions, how many indicators, how deep the regimes go, how many negations — and which parts of the vocabulary it reaches for. Read-only: the window runs nothing, requests nothing and changes nothing.

Highlighted logic

Individual conditions can be starred and annotated. The note travels with the condition's location inside its strategy and the text that makes it recognisable elsewhere, so an observation made while looking at one strategy is still findable when the same idea turns up inside another. It is the one place in the application where the content is your reasoning rather than a derived number.

What a backtest returns

Not a score. The full result: every trade with its entry and the reason it ended, the equity curve at full resolution, the drawdown profile, the statistics, and the attribution of which conditions fired when — drawn as the book, the same component the Replay page uses for a running deployment. A backtest can be played, scrubbed bar by bar, jumped from signal to signal and looped, with a panel beside it saying which of the strategy's conditions were true under the cursor. Clicking a trade in the trades table moves the book to it.

Modelled	Why it is not optional
Commission and fees	A strategy that trades often can be profitable gross and losing net. Trading frequency is a cost decision, not just a signal decision.
Slippage	You do not get the price you saw. Sweeping slippage rather than assuming one value shows how much of a result depends on filling well.
Execution delay	A signal computed on a close cannot be filled at that close. Lag is charged deliberately rather than being an accident of the data.
Warm-up	An indicator needs history before its first honest value. Trading during warm-up invents signal out of an incomplete window.
Non-trading gaps	Weekends and holidays are collapsed rather than drawn as flat lines, so a chart does not imply activity where there was none.

The Monte Carlo panel of the demonstration run: hundreds of resampled equity paths drawn as a translucent band with percentile edges, and the run's actual curve inside it — showing where the realised path sits among its own reorderings. — **Is the curve an ordering accident?** The trade sequence is resampled hundreds of times and every alternative path drawn as a band; the realised curve sits inside it. A strategy whose result depends on one lucky ordering shows up here as a curve hugging the band's edge. Synthetic series, demonstration arithmetic.

Charts that do not flatter

Linked zoom across figures, so a drawdown and the price that caused it stay aligned while you look.
Currency and percentage views of the same series, because a curve that looks dramatic in one is often unremarkable in the other.
A buy-and-hold baseline. A result can always be read as the difference against simply holding the same instrument for the same span, so a rising market cannot quietly take the credit.
Sub-panes per indicator, with the indicator drawn on the same time axis rather than a separate picture you have to align by eye.
Gaps drawn as breaks. Missing data is shown missing. A line joined across a hole is a claim that nothing happened there — and where the axis collapses a weekend, a quiet mark on the first bar after it says how long the gap was.
Strips that share the book's window. Drawdown, position, volume and oscillators are drawn on the same pixel window as the book above them, so a trough sits under the bar that caused it rather than a third of the width away.
The cash line. What the capital would have earned doing nothing is drawn beside the strategy, because a strategy invested a sliver of the time can look like skill when it is mostly interest.

The Backtest page's book of the synthetic series SYN-BOREAS at full-screen size: the transport with play, speed and landmark controls across the top, candles with moving-average overlays, the strategy book against buy-and-hold with the gap filled, entry and exit markers, the strategy's regime boxes shaded over the candles, and the oscillator strips beneath sharing the same window. — **The whole study surface at once.** The transport along the top — play, speed, the next landmark, a loop — above one figure holding the price, its indicator overlays, the strategy's book against buy-and-hold, every entry and exit, and the regime gate shaded straight onto the candles; the oscillators run in strips beneath on the same window. The toggles switch each layer, so the book shows exactly as much of the strategy's reasoning as you ask it to. Series: *SYN-BOREAS*, an invented tape from the app's own generator.

Nothing on this page is investment advice or a performance claim. Where a figure, a curve or a marker appears in a frame, it is the demonstration instance's own arithmetic over app-generated synthetic series, behind the application's standing hypothetical-results warning — not a result and not a track record. Trading involves risk of loss.

Stage three

Searching for strategies

Generating and evaluating enormous numbers of candidates is the easy half. Nearly everything described below exists to stop the search handing you the luckiest arrangement of noise it could find.

The problem, stated plainly. Search a million random strategies against one price history and thousands will look excellent. Not because they work, but because a million tries against one sample will produce extremes. A search with no defence against this is a machine for manufacturing false confidence, and it will do so tirelessly.

How candidates are generated

A grammar assembles logic trees from the indicator registry — choosing shapes, operands, comparisons, windows and regime gates from weighted draw tables. The generator's default behaviour is deliberately frozen: absent an explicitly armed option, the server takes its historical path byte for byte, so a stored search seed reproduces exactly the same population it did when it was recorded.

That discipline is what makes an old result checkable. A search you cannot re-run is an anecdote.

There are two grammars, and the second one is where the new vocabulary lives. The classic grammar is frozen byte for byte — every stored search on it still reproduces its population exactly. The extended grammar draws from every one of the 113 indicators and writes what a person can now write by hand: a condition negated on a cross, an at-least-K vote, an exit on the position's own state, a risk gate, an indicator read off another instrument. Any change to what it draws bumps a revision number that rides on every noise floor measured under it, so a floor can never be read against a search it was not measured for.

Evolution, not just enumeration

Populations and generations. Candidates are scored, the better ones become parents, and offspring inherit and mutate their structure. Crossover combines two parents' subtrees.
Lineages. A run can be rooted on an existing strategy so you are exploring around something specific rather than starting from nothing.
Immigrants. Fresh random candidates are periodically injected, because a population left alone converges and stops looking.
Cycles, for the long runs. A search left running alternates searching with pruning — the kept set is re-filtered against the stored record, survivors seed the next pass, and an extinction guard skips a prune entirely rather than let one over-strict filter wipe the population.

The loop that actually runs. Fitness is computed on training data only; selection walks a frontier with the diversity brakes applied inside the loop; the generator keeps injecting strangers so the population cannot quietly become one idea. The same loop, run against deliberately meaningless data, produces the empirical null everything must beat.

The mining run form pointed at the synthetic series SYN-ATLAS: a status line with the universe's size and a stale noise-floor badge, a minimum-backtest-length advisory, the ticker, period, granularity, mode, waves and strategies per wave, the stored-data-only switch, and the auto-cycle strip reading mine, IS cut, OOS gauntlet, gates, promote — with the hard caps spelled out underneath. — **Configuring a search — against an invented tape.** Two lines of honesty sit above the form before anything runs: how much evidence this tape can support at the trial count about to be spent — here nothing under a Sharpe of 1.46 can be told from luck, stated as an advisory rather than a gate — and a noise-floor badge reading *stale*, because the floor on record was measured under different settings from the ones now on the form. The auto-cycle strip is the loop from the diagram above, and the hard caps are printed on the form because a sweep that can take the application down is not a feature. Target: SYN-ATLAS, synthetic.

The mining lab's equity-curve panel after a finished run, labelled three assets: the mined synthetic tape SYN-ATLAS beside the two basket tapes SYN-BOREAS and SYN-CHRONOS, each card holding the top curves of the kept set drawn across that tape. — **Basket scoring, drawn.** Every kept candidate is scored across all three synthetic tapes rather than only the one it was mined on, and each tape keeps its own card — the mined window on the left, the two scored-not-mined tapes beside it. A candidate that only works on its own tape separates from its basket in one glance; the demonstration set visibly struggles on the bearish middle tape. The line above the cards names the leader and says outright that the sketches understate its drawdown — they are drawn from 150 points, and the card's figure is the real one. Synthetic series throughout; the curves are the demonstration's arithmetic, not results.

The strategies table after a mining run: all 250 kept rows passing the filters, sorted on validation Sharpe, each with its wave, its kind — random, evolved or immigrant — a sparkline, and in-sample, validation and out-of-sample columns, under the application's hypothetical-results warning, with the verbs and judges in a strip beneath. — **What survived, as rows you can interrogate.** 280 candidates were made across four waves — random first, then evolved, with immigrants brought in for diversity — and 250 kept; the rest were culled inside the loop. The table sorts on the held-out validation Sharpe by default rather than the in-sample one, the validation and out-of-sample columns ride beside the in-sample ones, and every column filters. The application's own hypothetical-results warning sits above the lot — the kept set is a thing you question, not a leaderboard you read.

The page, and the candidate's own book

The mining page is laid out in the order a question is asked of a search: what was searched for, what every candidate did, and then one candidate in detail. Every candidate's curve is drawn as a field — on the tape it was mined on and on each basket tape it was only scored on — so a population that works only where it was bred separates from its basket at a glance. Pick one and it is re-run at full resolution into the same book the Backtest page uses, with a line that says whether the re-run reproduced the row it came from.

The mining page after a finished search on SYN-ATLAS: the generator's settings and honesty line on the left, the field of every candidate's in-sample curve on the mined tape and on two basket tapes in the middle, and the chosen candidate's own book on the right, re-run at bar resolution with its signals and benchmark. — **The whole search on one screen.** Left, the settings — including the line reporting how much evidence this tape can support at the requested number of trials. Middle, the field. Right, one candidate's own book. Synthetic series throughout; every figure is the demonstration's own arithmetic, not a result.

The candidate's book for one mined strategy, w4·063: candles of SYN-ATLAS with three moving averages, its signals and the stop levels it carries drawn as dashed lines, the strategy book against buy and hold, the validation window marked, a drawdown strip beneath, and a line reporting that the re-run differs from the row, because the book ran on today's tape, one bar shorter than the tape the row was mined on. — **One candidate, re-run properly — and honest about the difference.** The mined row's curve was drawn from a sample of its bars; the book re-runs the candidate at full resolution on the universe's tape and says whether it reproduced the row. Here it says it did not, and why: this universe states no span of its own, so the book used today's tape — 1,060 bars against the 1,061 the row was mined on — and it labels the result a shorter run's numbers, Sharpe 1.86 against the row's 1.85, rather than passing them off as the row's. The validation window is marked on the chart, so the part the search trained on and the part it was judged on are never confused.

Four numbers that were not what they said. A review of the mining evidence found that the noise floor's twin tapes had leaked the next bar's open into the fills they were judging; that validation slices were cut from a sampled curve rather than the bars themselves; that parents were being bred partly on the validation span they would later be judged on; and that one fixed holdout was doing all the judging. Each was corrected — two of them as corrections that changed the meaning of stored numbers, and said so — and the page's defaults stopped favouring the in-sample maximum.

The countermeasures

Diversity control

A search ranked purely on score converges on one idea wearing a thousand hats. Niche caps limit how many similar candidates survive, and a correlation cull removes those whose returns move together — structural difference is not the same as behavioural difference.

Ranked frontiers

Parents are selected along a Pareto frontier across several objectives rather than by one blended number. Blending into a single score lets a strategy buy a great return with terrible robustness and still look good.

Held-out data, in the loop

The fitness a candidate is bred on is computed on training data only, with an embargo between the training window and the held-out one so a result cannot leak across the boundary through overlapping indicator windows. Parents are ranked on the training span alone, and the holdout itself can rotate, so no single slice of history does all the judging.

The market question, per block

Every candidate is measured against simply holding the tape — in six-month calendar blocks, not as one total that a single good year can carry. A tape too short for two blocks gets no verdict rather than a flattering one.

A seed check

The same search can be re-mined under a different seed. A search whose top candidates mostly fail to reappear was describing its seed, not the tape.

Complexity has a price

More conditions can always fit more noise, so complexity can be charged against fitness — and a verb takes conditions away one at a time to see which were doing any work.

Minimum evidence

A candidate with too few trades is not a strategy, it is a coincidence with a chart. Minimum trade counts ride on every run rather than being applied afterwards.

Cost stress, in the loop

Costs can be varied during the search rather than checked at the end, so fragility to assumptions is selected against instead of discovered later. The cost-stressed twin is derived arithmetically from the candidate's own run rather than re-run, which made honest searching markedly cheaper per candidate.

The empirical null

The same search run against deliberately meaningless data establishes what a good score looks like when there is nothing to find. Anything that does not clear its own null is noise with a nice chart.

It is matched to the kind of search that produced the candidate: a null generated by random draws says nothing about a population that was evolved, so the null mirrors the wave's own steps. And it runs when a cycle arms — before the first cut rather than after — so the bar is set before anything has been selected against it. Its twin tapes wear each bar's own geometry, so a fill cannot see the bar it is about to be judged on.

What the search is allowed to steer on

A search optimises whatever number it is given, exactly and relentlessly, which makes the choice of that number the most consequential setting in the whole apparatus. Three constraints sit on it:

It is measured where the search cannot see. Cutting on the in-sample score is not available at any setting, and a floor can be put under the statistic as well as a rank — ranking alone keeps the least bad candidates rather than good ones, which is a different thing entirely.
It can be several numbers at once. The front keeps what nothing else beats on every objective simultaneously, rather than collapsing risk and return into one score using an exchange rate nobody chose. That costs something measurable — about a third of the validation return for roughly half the drawdown — which is why it is a choice and not a default.
It is the number the product can actually deliver. Fills are the next session's open, not the close the signal was computed from. The gap between those two is the size of the apparent edge in a great many mined strategies.

A search can also be told to lean toward the kinds of set-up that have produced survivors before, weighted by evidence rather than by the last result. Every part of that is off unless you arm it, and each lean names itself in the run's own log — a search that quietly changes its aim between runs produces results that cannot be compared with each other.

Signal quality, at the moment of mining

A candidate arrives with statistics about its own conditions, not just its curve: how often each clause fired, whether any never fired at all, and whether a condition that looks important changes the result when it is removed. A clause that never fires is something the cut can now see — and the rate of dead clauses turned out to track how many conditions a candidate carries, which is exactly the complexity a search is prone to reward.

Two further measures — how much each condition adds on its own, and how much two conditions say the same thing — were built, measured, and kept report-only. The measurement showed the sign of the first flipping inside its own standard error, and a filter built on a number that noisy would select on luck while looking rigorous. The report stays; the gate was withheld.

Condition attribution

When a search returns something, the natural question is which part of it is doing the work. The attribution ledger records, per condition, how often it fired and what happened when it did — so a candidate whose entire result rests on one lucky gate is visible as such rather than presented as a coherent idea.

Why it is fast enough to be useful

A worker pool that imports almost nothing. Evaluation runs in separate processes whose import surface is deliberately tiny. A worker that had to load the whole application — web framework, machine learning stack, a large history file — would cost a fraction of a second each, which across a large search is the difference between practical and not.
Spawned, never forked. The main process is multithreaded and may be holding a live broker session. Forking that is a known way to produce a process in an undefined state.
Series are memoised per job and handed to candidates shared rather than copied, which means the code path that reads them must stay strictly read-only — a constraint the project treats as load-bearing rather than an implementation detail.

Universes

A finished search is saved as a universe: the full logic trees, the statistics, the run configuration that produced it and the seed that reproduces it. Universes can be filtered, re-scored, compared and mined further. The point is that a result is stored with enough context to be interrogated months later, rather than as a row in a leaderboard.

A universe stores both halves of the run that made it — the search and the selection. Re-opening one restores the cut, the keep filters, the gates and the stop rules exactly as they were armed, because "what did I search for" and "what did I keep" are two different questions and a stored answer to only the first is not reproducible. There is a roster across all of them, with the linking, renaming, pruning and deletion that a few hundred saved searches eventually require — and a delete that refuses while anything still refers to what you are removing.

A cycle of all this can be handed to an unattended researcher, which is its own page — mostly because the interesting part is not the searching but the list of things such a researcher is prevented from reaching.

Pricing the exit

One lab sits beside the search rather than inside it, and it is included here because it is the clearest example on the site of a measurement coming back against the thing that was asked for. It was requested as an optimiser: mine the best stop-loss and take-profit rules, apply them, make mined strategies beat the market. The measurement taken before building it found that a stop's effect on return does not survive contact with new data by any method tried — while its effect on drawdown transfers reliably.

So it prices rather than optimises. Every candidate exit is an insurance line: what it costs in return, what it buys in drawdown, how many extra trades it causes, and whether it ever actually fired. The recommended pick is a constrained one — cheapest protection meeting the requirement you set — and below a floor on how many times a stop was genuinely tested it declines to pick at all.

The stop lab after pricing eight ticked strategies: the mine form with stop families, a budget and a return floor; the pass line naming the span, the stripped base and the budget; a red verdict that the strategy on screen is not estimable — four drawdown episodes against a floor of eight — so nothing is picked; a table of candidate exits with how often each fired, the change in trades, the premium in Sharpe, the return given up, the drawdown bought back and the same measured on validation data; and a scatter of premium against payout beside the model and apply controls. — **A stop as a price — and a refusal to guess.** Each row is one candidate exit, and every column is a difference against that strategy's own no-stop control, so a zero means the level changed nothing; the scatter sets what each costs in-sample against the drawdown it buys back. Here the lab declines to pick at all: the training span holds four distinct drawdown episodes, under its floor of eight, so it says in red that the table is description only and nothing is chosen. A stop chosen from four episodes would be chosen from noise. Priced over eight strategies mined from an app-generated synthetic series in the demonstration instance; the figures are that demonstration's own arithmetic and are not results.

The diversity heatmap of the demonstration universe mined from a synthetic series: every kept candidate compared against every other, shaded by how similar their behaviour is. — **The survivors, compared to each other.** The diversity heatmap is the correlation cull's view of a saved universe — every kept candidate against every other, shaded by how alike they behave. A leaderboard cannot show you that its top ten are one idea; this can. Mined from a synthetic series in the demonstration instance; the shades are the tool, not a result.

The 3D universe map of the demonstration universe: every kept candidate placed by Sharpe, return and trade count, coloured on the same green-to-red outlier scale as the flat maps, with the current best marked by a diamond. — **The same universe, in three axes.** Each kept candidate is a point placed by three metrics of your choosing — here Sharpe against return against trade count — with a fourth carried as colour. The cluster of near-identical points is the interesting part: it is what convergence looks like, and the flat maps cannot show it. Demonstration universe, synthetic series.

The Adversary — distance to death

The last question the lab asks of a survivor is not "how good is it" but "what kills it". The Adversary runs a candidate through a fixed ladder of eleven increasingly hostile synthetic futures — calibrated calm, rising volatility multiples, bears, jump storms, a crash world — a few seeded waves per rung, with death declared when a rung's median drawdown breaches the kill floor. Because every rung is stated in calibrated units of the tape's own tail volatility, the verdict reads as a sentence: the mildest world that kills this one is a bear at 1.6× volatility.

The ladder is fixed, so verdicts compare. Two candidates run the same eleven worlds in the same order; "survived to rung seven" means the same thing about both.
Nothing is stored. The hostile worlds are generated in memory and discarded; what remains is the stamp on the row — rungs survived, and the worst drawdown seen.
It refuses a synthetic base. The hostile futures are grown from a real tape's own character, so the demonstration run uses the neutral demo series rather than a SYN-* one — the same rule the stress lab and the Ordeal apply.

The strategies table of the demonstration universe under the application's hypothetical-results warning, its rows carrying walk-forward fold verdicts as ordinary columns, with the strip of verbs and judges beneath: promote, ablate into simpler variants, prune, evolve, robustness, noise, vintage, lag, carry, the Ordeal batch, the Adversary with its ladder size and minus-forty-percent kill floor, and the Vigil with its paper capital and incubation window. — **The judges live one click from the kept set.** Along the bottom of the table: robustness jitters, noise twins, vintage sweeps, lag and carry re-scores, the Ordeal batch, the Adversary — its ladder size and the −40% kill floor set right on the button — and the Vigil's paper capital and incubation window. Each stamps its verdict back onto the rows as ordinary sortable, filterable columns, so judging the kept set is part of using it, not a separate ceremony. Demonstration universe on the neutral demo tape.

Nothing on this page is investment advice or a performance claim. The figures inside the frames are a demonstration search's own arithmetic over app-generated synthetic series — illustrations of the tools, not results, behind the application's standing hypothetical-results warning. A search result describes history under assumptions you supplied; it is not a forecast. Trading involves risk of loss.

Stage four

Trying to break it

The part of the project with the most work in it, because it is the part that decides whether anything else means anything.

Everything here is adversarial by design. The question is never "does this look good" — it is "what would have to be true for this to be an illusion, and can I make that visible?"

One past, spent carefully. There is only one actual history, so it is divided with an embargo at the seam, marched through fold by fold, and finally left behind entirely — synthetic continuations are data the strategy provably never saw, because they never existed.

Out of sample, and the embargo

Testing is separated from the data used to find the idea. That much is standard. The part that is easy to get wrong is the seam: an indicator with a 200-bar window evaluated on the first day of the held-out period is partly computed from training data. An embargo between the two windows keeps the boundary honest, and it applies inside the search loop rather than only in a final report.

Walk-forward

Rather than one split, the strategy is repeatedly fitted on a window and tested on the window that follows, marching through history. This answers a different question from a single hold-out: not "did it work on data I did not look at" but "would it have kept working, re-derived as you went, all the way along?"

The gauntlet

A structured battery that re-runs a surviving strategy under changed conditions — different windows, different instruments, degraded assumptions — and reports coverage alongside the score.

Coverage is reported because a partial run can score higher. If a trial fails to run on half its cases and the half that ran happened to be favourable, the average looks better than a complete run would have. Coverage is the only guard against that, so it is always shown next to the result rather than being an implementation detail.

The out-of-sample equity section with two remembered windows side by side, each a card of the kept set's curves re-run from a start date to the end of the SYN-ATLAS tape, the leader highlighted against buy and hold, and a drawdown strip under each. — **Two remembered windows, side by side.** The kept set re-run over two windows of the tape, each keeping its own card — a later gauntlet leg never erases the one before it — with every strategy wearing the same colour here as in the in-sample section. The leader goes flat for the last months of both windows while buy-and-hold rallies, which is the kind of thing a single number hides. In this demonstration both windows overlap the stretch the search was scored on, so they show the mechanics rather than a clean test; a research bot's funnel holds its tail out of the mine by construction. Synthetic series; the curves are the demonstration's arithmetic.

The selection-honesty panel over the demonstration universe: tiles for 280 examined, 250 kept, the deflated-Sharpe median and best and the minimum backtest length; a deflated-Sharpe histogram; a PBO run reading 71 percent with the verdict overfit factory; two reality-check tiles with p-values of 0.94 and 0.97; and the noise floor's tiles and verdicts — the best live Sharpe at or below the null median, the kept set's body at or below the null body median. — **The number that knows how hard you looked.** A Sharpe ratio means nothing without the size of the search that produced it. This panel deflates every kept candidate by the 280 examined; estimates the probability that the selection itself was overfitting — 71 per cent here, which it calls an *overfit factory*; runs two reality checks on the best row's excess over buy-and-hold; and runs the same generator over signal-free twins of the tape to set the noise floor. Its verdict on the demonstration: the best live Sharpe sits at or below the null's median — *indistinguishable from noise wearing a strategy's name* — and the kept set's body is *what noise keeps*. It even notes that this universe evolved while its null drew at random, and says to re-run the null to match. That is the product working as intended.

The Ordeal, as one page

Twelve trials run in sequence against one strategy with nothing to configure beyond depth — the same ordeal every time, which is the point. The result is drawn as one page: each trial a card with a meter, every meter oriented so that more is better (the server inverts the ones that are naturally "lower is better"), and a score card whose grade ring is stacked by pillar so that the arcs sum to the score exactly.

The letter had to be made to mean something again. An earlier version let hard flags pull the average down, and across a real population of mined strategies the letter had drifted into little more than a count of flags. Version three keeps every flag out of the average and lets each one cap the grade instead — and prints what the grade would have been without them.

The Ordeal's trials after a quick-depth run on the demonstration strategy: twelve cards — baseline backtest, Monte Carlo, vintage cone, parameter surface, backtest overfitting, robustness jitter, walk-forward, regime stress, cross-symbol transfer, execution lag, crisis holdout and deflated Sharpe — each with its own finding and a meter. — **The whole battery, as a dashboard.** Twelve of twelve trials measured in 56 engine passes — each card states what it tested, what it found, and a meter oriented so that more is better. Three meters are red — the overfitting estimate, the crisis holdout and the deflated Sharpe — and the score card below does not let the green ones outvote them. Demonstration strategy on the neutral demo tape; every figure is that run's own arithmetic.

The Ordeal's score card for the demonstration strategy: a grade ring reading D, 69 of 100, stacked by pillar; three hard flags — no protective stop, probability of overfitting over 0.5 at 0.546, deflated Sharpe under 0.5 at 0.329; a note that the flags are kept out of the average and cap the grade, which would otherwise have been C; the headline figures, including a return 109.9 points behind buy and hold; a list of where the missing points went; and nine weighted pillars from crisis survival to transfer, each with its own sub-statistics. — **The verdict, with its reasons attached.** The demonstration strategy came back grade D, 69 of 100 — capped by three hard flags the card names outright: no protective stop, an overfitting probability of 0.55, a deflated Sharpe of 0.33. The flags are kept out of the average and cap the grade instead, and the card says it would otherwise have graded C. Its headline line admits the strategy finished 109.9 points behind buy-and-hold; beside the ring, the card lists where the 31.2 missing points went — edge, luck, regime stress — and how far it sits from a B. A tool that graded everything B+ would be decoration; this one demotes its own demonstration strategy and says exactly why.

Stress sweeps

Costs, slippage, execution delay, volatility shocks and price jumps swept across ranges rather than spot-checked at one value. The whole sweep is saved, so the grading can be revisited later against different thresholds without re-running anything.

Cost and slippage ranges — at what assumption does the edge disappear?
Execution delay — how much does it matter that you cannot fill instantly?
Volatility shocks — behaviour when the regime changes underneath it.
Price jumps — gaps and discontinuities, which stops do not protect against as well as people assume.

The stress lab's setup: the demonstration strategy picked on the neutral demo tape, four synthetic scenario rows — two-year bears and three hundred mixed days — each with a regime, a length and a wave count, the advanced controls, and a note that nothing is stored because the whole sweep is generated in memory. — **Setting up a sweep.** Synthetic futures are grown off the same real base — two-year bears and mixed three-hundred-day stretches here — with several waves each, on deterministic seeds so a re-run replays the same tapes. The form's own text says the important part: nothing is stored; evidence about robustness is not a track record.

The stress verdict card after a finished sweep on the demonstration strategy, headed read this before deploying: the verdict FRAGILE, two of four gates passing over six waves — worst drawdown and probability of loss failing, mean excess and beating buy-and-hold passing — with the tails and the weakest and strongest scenario named. — **Read this before deploying.** The verdict card compresses the whole sweep into four gates — worst drawdown, mean excess, how often it beat holding, how likely it was to lose — and a word. Here the word is *fragile*: two of four gates over six waves, with the weakest and strongest scenario named. Its heading is an instruction, not a label. Demonstration strategy on the neutral demo tape; every number is that sweep's own arithmetic.

The equity-across-scenarios panel after a sweep: every wave of every synthetic scenario drawn as its own curve — bear, mixed and sideways futures at two cost levels — with the real-history run-in visible before the synthetic tails fan out. — **Every wave, drawn.** Each synthetic future is its own curve, so the sweep's spread is visible rather than summarised — the fan opening after the real run-in is the moment the invented futures diverge. A strategy that only survives the average future is not the same thing as one that survives most of them, and only this view can tell you which you have.

The outcome-distribution panel drawn as cumulative distributions: each scenario's per-wave final returns as a step line, with buy-and-hold's across every wave dashed beside them, so the shape of the whole sweep is one picture. — **The whole sweep as one shape.** Every wave's final outcome lands in one distribution, strategy beside buy-and-hold — the honest summary of a sweep is not its mean but its shape, and especially its left tail.

Three more questions the battery asks

A tail nobody looked at — the sealed windowproof

Every number an unattended search produces has been selected on: fitted on the training span, selected on the validation slice, selected on again at the out-of-sample tail. So the last months of the tape are sealed before the search begins — no pass, no gate and no Ordeal may read them — and opened exactly once, after the search is over. It is the one number the selection cannot have touched.

Guards against: an out-of-sample test that has quietly become in-sample through being looked at a hundred times.

The market question, per regimemarket

Beating buy-and-hold over a whole tape can be one good regime carrying two bad ones. The comparison is also made regime by regime, in blocks, so a strategy that only works when the market is rising is identified as a leveraged opinion about the market rather than an edge.

The noise floor decidesnull

A search's results are read against the same search run on tapes with the signal destroyed — and the floor is compared with the body of the results, the top few together, rather than the single best row, because the best of many draws is exactly what luck is best at producing.

Synthetic continuations

Price paths generated beyond the end of real history, with the statistical character of the real series but none of its specific sequence. A strategy tested on these is being tested on data it provably cannot have been fitted to, because the data did not exist when the strategy was found.

Synthetic results are held in memory and never written into the stored history as though they were real. A synthetic run is evidence about robustness; it is not a track record, and the two must not be able to be confused later.

Vintages

A strategy can be re-run anchored at different start dates to see how much of its result depends on when you happened to begin. This is distinct from slicing an existing curve — slicing shows you a segment of one run, re-running produces a different run, and the difference between those two is exactly the effect being measured.

Scenario windows

Behaviour isolated to specific historical periods — a crash, a squeeze, a long grind — computed server-side on the full-resolution equity rather than on a sampled curve, because the statistics of a drawdown are not preserved by sampling.

The Adversary — the mildest world that kills it

Every test above asks whether a strategy survives something. This one asks how little has to change before it does not, which is a more useful question and a harder one to answer honestly. Eleven worlds run in order from calm to crash, each stated in units calibrated against the strategy's own measured behaviour rather than in absolute numbers somebody chose.

It is a fixed ladder and deliberately not a search. Turn an optimiser loose on "find the world that kills this" and it wins by discovering a degenerate corner nobody can interpret — a world with no resemblance to any market, which tells you nothing about the strategy. The deliverable here is a sentence: the mildest named condition under which the thing dies. The calm control rung runs first and defines what "normal" means for the rungs above it, and if the control cannot calibrate the whole run fails rather than letting the other labels lie.

The noise retest

Every bar is nudged by a draw scaled to its own trading range — quiet days wobble less than violent ones — and the strategy runs again. The closing prices are deliberately not re-chained, so this is a wobble around the path history actually took rather than a drift into a different history.

It answers a question the neighbouring tests do not. Perturbing the strategy asks whether its parameters sit on a knife-edge; destroying the structure entirely asks what luck scores. This perturbs the tape — and asks whether the result needed prices to land exactly where they did. A stop grazed by a fraction of a percent, an entry that required one particular tick: these are invisible to every other test on this page and fatal in practice.

Testing the procedure, not just the winner

All of the above examine one strategy. The most important source of self-deception is not in any individual strategy, though — it is in the selection that produced it. Search ten thousand candidates and the best one will look excellent whether or not anything is there, and no amount of testing that best one in isolation will reveal it.

So two published statistics are run against the search's own record. The first splits the trial history combinatorially and asks how often the configuration that looked best in one half failed to stay above median in the other — an estimate of the probability that the procedure is overfitting. The second asks whether the best result in the population beats what you would expect from that many attempts at nothing, which is the multiple-comparison correction a per-strategy test cannot make. Both are read as floors, and ties count against the candidate.

A third figure sits beside them and is computed before any searching starts: given the length of history you have pointed at and the number of candidates you are about to request, how good would a purely lucky best be expected to look. A result underneath that line is not a weak result — it is not a result. See the tape's ceiling.

What survives

Very little, and that is the intended outcome. The purpose of this stage is not to certify strategies; it is to make the reasons for doubt explicit and quantified, so that whatever you do decide to run, you are doing it knowing which assumptions it depends on and how hard each one was pushed.

Surviving these tests is not evidence that a strategy will work. Robustness testing can reduce the chance you are fooling yourself; it cannot establish that a pattern will persist. Nothing on this page is investment advice or a performance claim; the figures and verdicts inside the frames are a demonstration instance's own arithmetic over app-generated synthetic series — including the failing grades, which is the tooling doing its job — and trading involves risk of loss.

A personal project. Not advice. Nothing for sale.

This site describes a personal engineering project, written and run by one person in their own time and published under the name Mithraeum Agora. There is no company behind it, no team and no other contributor. It is not a product and not a business. Nothing here is for sale — there is no account to open, nothing to buy, no subscription, no waiting list, and no service is being offered or solicited. Sending a message through the note form creates no customer, client or contractual relationship of any kind. The note form on the contact page is the way to reach me, and it is the only one.

Nothing on this site is investment advice, financial advice, tax advice, or a recommendation, solicitation or offer to buy or sell any security or financial instrument. I am not a financial adviser, a broker, an investment manager or a regulated firm, and nothing here should be relied on as though I were. If you are making decisions about money, take advice from someone qualified and regulated to give it.

Figures do appear on this site, inside screenshots, and none of them is a performance claim. Every one is a demonstration instance's own arithmetic over price series the application generated itself — synthetic data, with no real market behind it — shown to illustrate what the tools display. They are hypothetical and simulated: no capital was at risk, no orders were placed, and a result computed over history with the benefit of hindsight carries limitations that live trading does not forgive. Hypothetical results are not indicative of future returns, and past performance — real or simulated — predicts nothing.

No brokerage, market-data vendor or other company is named anywhere on this site, and where a name appeared inside a screenshot it has been redacted out of the image. Nothing here states or implies that any company is associated with this project, endorses it, sponsors it, supplies it or has reviewed it. None is, and none has.

What this site collects. Nothing, unless you write to me. There is no analytics, no tracking, no advertising, and nothing at all is loaded from another domain. If you use the note form it takes the name, address and message you type, stores them privately where only I can read them, and keeps them for up to a year before they are deleted — sooner if you ask, and you do not have to give a reason. Your IP address is not kept. One thing is stored on your own device: the colour theme you pick, remembered by your browser so the site does not change appearance every time you arrive. It is written only when you choose a theme, and it identifies nothing and nobody. The preference itself never leaves your browser, but the screenshots follow it, so the pictures your browser fetches from this site are the ones drawn in that palette. The contact page answers all of this in more detail.

Terms of use. This site is provided as is and as available, with no warranty of any kind, express or implied. It describes software under active development: anything here may be incomplete, out of date or simply wrong, and it may change or disappear without notice. Nothing on it is a contract, a term of service for any product, or a promise that anything described will be built, released or kept running. To the fullest extent the law allows, I accept no liability for any loss or damage arising from use of this site or from reliance on anything it says. Nothing here excludes or limits any liability that cannot lawfully be excluded or limited.

Trading involves risk of loss.