Ask the Record

Headline, under the oracle's own dynamic sizing — confidence-scaled leverage inside True Edge's declared 2-10x family band, on a 2% shared risk budget, fees inside every row, from the sealed weekly-cadence walk (True Edge's book sealed 2026-08-31), with the dollar figures re-stated 2026-10-05 under the settle-at-horizon-end rule: 615.2 R over 520 graded decisions across 13.3 years, 102.2% CAGR, −42.1% maximum drawdown, 59.7% win rate. The grade on that measurement is proven.

The book is a weekly-cadence walk of decision points; the card does not call weekly. It scanned 664 decision points and graded 520 of them in its most active horizon lane — 664 minus 520 leaves 144 points where that lane produced no trade — which works out to roughly one call every nine days. The 2600 graded rows are those 520 decisions carried across 5 horizon lanes.

The cumulative figure, which belongs in an expanded receipt view and not in a headline, is +1,164,983.9%: the same sealed book compounded — a 2 percent shared budget, the 2-10x confidence-scaled band, five horizon stacks off those 520 decisions — under one settle rule. Each trade is sized when it opens, from equity that holds only trades already settled; its result is booked when it settles, at its horizon end, the latest it could have; nothing compounds before it settles. At most 8 trades were open at once. The span runs from the first entry to the last settlement, which is why it reads 13.3 years rather than the cutoff span. Beside it, the flat 1x baseline on the same book: 947.1% with a −9.8% maximum drawdown. Both figures are replay of a frozen past. Neither is a forecast, and the difference between them is sizing, not a better prediction.

These dollar figures were re-stated on 2026-10-05. The earlier presentation compounded trades in the order they opened, so a trade could be sized off results that had not settled yet; the correction, with the old and new figures side by side, has its own answer on this page.

Four cards, the same ten fields each, dynamic and flat-1x beside each other. No card gets a summary where another gets detail.

True Edge — 102.2% CAGR dynamic (19.3% flat 1x), +1,164,983.9% total return dynamic (947.1% flat), −42.1% maximum drawdown dynamic (−9.8% flat), 615.2 R, 2600 graded rows, 13.3 years, 59.7% win rate. 9 proven / 1 mixed / 0 unproven.

LHCS / Curve — 8.6% CAGR dynamic (4.4% flat 1x), 198.4% total return dynamic (78% flat), −35.6% maximum drawdown dynamic (−18.7% flat), 124.8 R, 732 graded rows, 13.25 years, 49.6% win rate. 2 proven / 4 mixed / 0 unproven. Its span has matched its siblings' since the 2026-08-31 depth restoration returned the 2013-2017 history to its call source; before that repair it abstained across the early years rather than guess behind a data-depth limit. Short-side book on the horizons that have one: 0.8 R at 7d over 6 trades and 0.7 R at 30d over 35 trades, both graded proven — small totals that won often, and the 7d sample is small, which is why its count is printed beside it; the card trades no 1d or 3d short book, and both of those fields read absent (absent at 3d).

MODEL — 8.2% CAGR dynamic (6.2% flat 1x), 179.9% total return dynamic (117.9% flat), −10.3% maximum drawdown dynamic (−7.7% flat), 168.7 R, 647 graded rows, 13.05 years, 47.1% win rate. 0 proven / 5 mixed / 3 unproven. This is the statistical baseline; its positive result comes from the long side, and its short side sits close to zero on every horizon — 0 R at 1d, −0.1 R at 3d, −1 R at 7d and −1.2 R at 30d.

The Voice — 27.2% CAGR dynamic (14.7% flat 1x), +2,314.0% total return dynamic (510.1% flat), −55.8% maximum drawdown dynamic (−30.8% flat), 411.5 R, 2068 graded rows, 13.22 years, 41.5% win rate. 0 proven / 7 mixed / 1 unproven. It wins under half the time and still carries a large positive R, which is what a wide-tail card looks like.

Grades: True Edge proven; LHCS / Curve, MODEL and The Voice all mixed evidence at the card level. The dynamic and flat columns differ by sizing alone — same trades, same entries, same exits, same fees. Every dollar column compounds under the same settle rule: sized at open from settled equity only, booked at the horizon end; at most 8, 7, 7 and 7 trades were open at once on True Edge, LHCS / Curve, MODEL and The Voice respectively.

The honest comparison, on both of our bases, against a buy-and-hold figure measured on our own tape over True Edge's span of decision points, first to last.

Bitcoin bought and held on our tape from 2013-04-28 to 2026-07-17 compounds at 59.7% a year on Bitstamp (59.9% on Bitfinex). The sealed presentation carries an external public-price-history estimate of 55% beside that; the headline is the tape measurement. Against our cards' dynamic sizing that beats three of the four: The Voice 27.2%, LHCS / Curve 8.6% and MODEL 8.2% are all below it, and only True Edge at 102.2% is above. Against the flat 1x measurement basis it beats all four: True Edge 19.3%, The Voice 14.7%, LHCS / Curve 4.4%, MODEL 6.2%.

LHCS / Curve's first decision point is 2013-05-17, a few weeks after True Edge's, so its span-matched comparison is the tape over its own 13.17-year window of decision points, 61.7% a year. Against that benchmark LHCS / Curve at 8.6% is below it.

Our CAGR figures are compounded under the settle rule — each trade sized at open from settled equity only and booked at its horizon end — and the card spans run to the last settlement, a few days to a month past the last decision point.

Two things have to ride with that comparison or it misleads. First, the benchmark figure is era-dependent and the early years dominate it. Second, it is a compound-growth number with no drawdown beside it, and buying and holding this asset historically meant sitting through falls far deeper than anything on this page. We do not have a pinned drawdown figure for the benchmark, so we do not state one; the comparison above is return against return only, and that is a friendlier framing for the benchmark than a full risk comparison would be.

For scale against managed money, the same file carries: Renaissance Medallion net 39%, a crypto quant composite 30% (survivorship-biased), Citadel Wellington 15%, Millennium 14%, and the SG CTA index 7%. Those are public reported figures carried as reference points; they are not measurements this system made, and none of them is a like-for-like on asset, horizon, or fee structure. Grade on every benchmark line: not a measurement.

Both conventions, every card, because a drawdown quoted without its sizing convention is not an answer.

At the oracle's designed dynamic sizing: True Edge −42.1%, The Voice −55.8%, MODEL −10.3%, LHCS / Curve −35.6%. At the flat 1x measurement basis underneath, same trades and same exits: True Edge −9.8%, The Voice −30.8%, MODEL −7.7%, LHCS / Curve −18.7%.

Read that gap honestly: leverage multiplies the drawdown as faithfully as it multiplies the return, and the dynamic column is the one to plan around. The worst single number on this page is The Voice at −55.8% — a run that would have taken the account down by more than half before recovering — with True Edge next at −42.1%. Individual horizons inside the cards go deeper still. Every figure here is measured under the settle rule: a trade is sized at open from settled equity only and its result is booked at its horizon end, so a losing trade cannot hide behind a winner that had not settled yet.

What these numbers are not: a worst case. They are the worst drawdown that occurred on a specific frozen tape, and a tape that has not happened yet can be worse. Grade: proven as a measurement of that tape, and nothing more.

Three layers, and one of them is deliberately switched off.

Per-trade geometry, always on. Every graded row in every sealed book carries a protective stop and tiered take-profits, with fees inside the row and ambiguous bars resolved against us. That is not a control bolted on afterwards; it is how the rows were produced in the first place.

A hand kill switch, proven live. During the live graduation the founder's own hand flipped the kill flag, and the machine cancelled the resting stop, flattened the position, re-read the book bare and left foreign orders untouched. The record of that drill reads KILL DRILL PROVEN.

A market-defense entry gate, built and disarmed. There is a gate that would refuse new entries for selected cards while a live composite risk tier sits at or above a chosen rung. Its arming flag currently reads false, which means the gate is inert and the executor behaves exactly as it did before the gate existed. No code path in this system can set that flag; it is edited by hand.

Two disclosures ride with that gate. Its intended membership is narrow — voice and one sibling — because the defense study measured this class of overlay costing True Edge across the full tape, so True Edge is deliberately isolated from it and LHCS / Curve is held pending a hearing rather than excluded. And the overlay's protective value is unproven: zero of sixty-four cells reached the pre-registered bar. We built the machinery and we are not claiming it works.

The short version: it measures how human market behaviour has actually behaved inside recurring, precisely-dated calendar windows — and it is graded by exactly the same frozen-tape court as every other card here.

The longer version, because this is the question we get asked sideways. Markets are made of people, and people are periodic. Everyone in finance already accepts a version of this claim: the day-of-week effect, the turn-of-the-month effect, the January effect, options-expiry week, the end-of-quarter rebalance. Each is a time marker — a date computable years in advance, with no forecasting involved — and each is used as a filter. Instead of asking what the market will do, you ask: conditional on being inside this recurring window, how have humans historically behaved?

This card uses a wider set of time markers than the usual calendar ones. They are astronomical in origin: deterministic, computable centuries ahead, and — this is the part that matters — completely independent of anything in the market, which is what makes them a clean conditioning variable rather than something fitted to the outcome.

What is not being claimed, stated plainly because the honest version is less interesting than the caricature. No celestial causation. No providence. Nothing in this system claims the sky moves the price. The founder's own words on that idea: it is nonsense. The mechanism is irrelevant to the measurement and no claim is made about one. What is measured is human behaviour inside recurring windows, using the same conditional-pattern mathematics you would use for a day-of-week effect. Rename every marker to macro_window_07 and the statistics are identical and nobody asks this question.

Why it is worth measuring. Peer-reviewed work exists on periodic and environmental effects in market returns — lunar-cycle return studies, seasonal-affective and sunshine effects on equity returns, and the Atlanta Fed's own published work on geomagnetic storms and stock returns. There is also a century of practitioner lineage that never got a rigorous public court. What is rare is not the idea; it is this feature class combined with this measurement discipline — frozen tapes, rules registered before the answers were read, pessimistic accounting on ambiguous bars, fees inside every row, and a per-cell grade published whether it flatters the card or not. The stigma attached to the vocabulary is exactly why a pattern that survives that court would be unlikely to be crowded.

And here is what the court says today, which is the part a deck would leave out. The card's current sealed measure is 198.4% over 13.25 years at its own declared 1-3x dynamic sizing, average realized multiplier 2.09x (78% at the flat baseline), 124.8 R, −35.6% maximum drawdown (−18.7% flat), 732 graded rows, a 49.6% win rate over those 732 graded rows (520 decision points scanned). Its six evidence cells grade 2 proven, 4 mixed evidence and 0 unproven. The two proven cells are both on the short side: 7d, 0.8 R over 6 trades, and 30d, 0.7 R over 35 trades — small totals that won often, and the 7d sample is small, which is why the count is printed beside the grade. The long side, where most of the return comes from, is mixed evidence on every horizon. The dollar figures are compounded under the settle rule: each trade sized at open from settled equity only, booked at its horizon end, at most 7 open at once. The defense-study overlay for this card is temporarily not shown at all: in the sealed 2026-08-27 study its twelve cells came back 8 coverage-limited by book span and 4 no-contact by construction, that study was measured on the previous book vintage, the card's book has since been re-sealed with the early-era rows the 2026-08-31 depth restoration returned, which the study never graded, and mixing vintages is never shown — the overlay's own disclosure says so in place: temporarily not shown.

So: this is the founder's original thesis, it is the reason the system exists, it is measured at full rigor with no discount and no favour, and at the card level it has not earned a proven grade yet, while two of its cells have. We publish it at those grades. If the card never earns one, that publishes too.

True Edge won 59.7% of its sealed graded rows and carries 615.2 R. The Voice won 41.5% — under half — and still carries 411.5 R, because its winners are much larger than its losers. MODEL won 47.1%, also under half, and carries 168.7 R, 141.9 R of it from its 30-day long lane alone. LHCS / Curve won 49.6%, just under half, and carries 124.8 R. Those four pairs are the argument against reading a win rate on its own — it tells you how often, never how much.

The number that matters more than any of those is the one most systems never publish: how often the card declines to call. Both cards below are measured the same way: decision points scanned at card level, graded rows in the most active horizon lane. True Edge scanned 664 decision points and graded 520 of them in its most active lane, 2600 rows across five lanes, so on that lane 144 of 664 scans produced no trade. LHCS / Curve scanned 520 and graded 214 in its most active lane, 732 rows across four lanes — on that lane 306 of 520 scans stayed silent, which means silence is its normal output.

Abstention is scored here as integrity, not as missing data. A card that declines when its structure does not support a call is doing the thing you would want it to do, and a measurement that graded that silence as a failure would be measuring the wrong thing. It also means the honest way to read the win rates above is: this is the hit rate on the occasions the system chose to speak.

You cannot know it from a claim, so here is the machinery and here is its limit.

The honest starting point. Many configurations were explored over the life of this research — different geometries, different horizons, different conditioning. Any system that says otherwise is either new or not telling you the truth. Exploration is not the problem; unlabelled exploration presented as a discovery is. So the design here is to freeze, then measure, then publish the grade whatever it says.

What is actually controlled. The defense study carries a pre-registration seal minted before its first replay, and when a repair was needed the study was re-issued as a new artifact that supersedes the original rather than editing it — the superseded file's hash is recorded inside the successor, so the chain is checkable without trusting anyone. Its inference runs a hierarchical stepdown over a shared plan, and the family-wise error rate of that procedure was simulated rather than assumed: 1000 simulated studies over 31 hypotheses, with the outcome channel replaced by independent normal draws so denied and admitted rows are exchangeable by construction. Nominal alpha was 5%; the realized error rate came back 5.6%, inside the two-sided binomial tolerance of 6.4%. The verdict recorded is passed. Its own disclosure notes the simulation runs 1000 bootstrap replicates per simulated study against the sealed study's 10000, which changes the granularity of each simulated p-value and not the procedure.

What that does not buy. That simulation grades the defense study's inference procedure. It does not retroactively price the whole research history that produced the card designs in the first place, and no honest correction can, because the denominator of things tried before a freeze is not recoverable. The only clean answer to that question is data nobody has looked at, which is why a 2024-onward holdout was sealed unread before the study ran and remains unread. Grade on the machinery: proven that it exists and operated. Grade on immunity from selection: we do not claim it, and the holdout is the test we have pre-committed to.

14.16M price bars — 14159246 exactly — counted across 12 files from two independent exchanges, bitfinex and bitstamp. The count is a direct inventory of the files on disk, line by line, with a hash recorded per file, rather than a number carried forward in prose.

Two exchanges matter because a single venue's history can carry its own outages, bad prints and gaps. Having the same asset from two sources is what makes a cross-check possible.

Two honest limits. This is an inventory of the tape held on disk; it is not a claim that every card consumed every bar, and different cards use different resolutions and spans. And it is one asset — this is a Bitcoin system, and the data provenance answer is a Bitcoin answer.

On three of the four cards the published book is graded on the geometry the card trades live; on the fourth the match is not established, and we say so.

The Voice card used to be the clear case of a split: its earlier sealed books were produced under a calibrated geometry while its live path runs an operator-ruled one — a bidirectional three-tier take-profit ladder at 4.0, 8.0 and 12.0 times risk with a sliding stop, carried in a named tune, voice_live_ladder_v1, that stamps itself onto every live signal it produces. Since the 2026-09-02 re-seal the Voice book published here is graded on that live ladder: its evidence tune reads voice_live_ladder_v1. The earlier calibrated-geometry books stay sealed beside it as a labelled vintage and are not re-written.

Across the four published books the tune check reads match on 3: LHCS / Curve, MODEL and The Voice. True Edge's reads unresolved evidence tune — its sealed book does not record which tune produced it, so the match with its live geometry is not claimed. The stamp on every live signal exists so nobody can later read a live post as though it came from the sealed book, or the reverse.

The general rule: a sealed number describes the configuration that produced it, stamped with its seal date. A live signal describes today's configuration, stamped with today's geometry. Where those differ, both are shown with their labels rather than blended into one comfortable figure.

Because that is the tape we own and the tape we have measured, and measuring one thing properly beat claiming four.

The price history behind every sealed number here is Bitcoin, 14159246 bars on disk, from two independent exchanges — bitfinex and bitstamp. The forward paper record is single-asset too, though its counts sit behind the authorization gate and are answered separately. Nothing on this page has been measured on equities, foreign exchange, or any other instrument, and no claim is made about them.

There are reasons the choice is a good one rather than merely convenient. The market trades continuously, so a time-marker conditioning study is not fighting session boundaries and holiday calendars. History is deep enough for a thirteen-year walk and short enough that one person can hold all of it. And the retail behaviour this system is trying to measure is unusually visible in this asset.

The honest cost of that choice: everything here is one asset's regime history, and a pattern that survives thirteen years of one market is a weaker claim than one that survives several. Extending to a second asset is real work and it has not been done.

Three layers, and only the first one has numbers on this page.

Layer one — sealed replay. Everything with a return, a drawdown, an R total or a win rate on this page comes from books sealed 2026-08-31 (True Edge) and 2026-09-02 (LHCS / Curve, MODEL and The Voice) over a frozen historical tape, with the dollar figures re-stated 2026-10-05 under the settle rule. It is replay of the past, computed once, hashed, and stamped with the engine commit that produced it.

Layer two — a locked forward paper record. Since 2026-07-19 the system has been writing its calls down before the outcome is known and resolving them afterwards. That record exists and is accruing. It is not published here: publication of forward-record material requires a hand-written authorization artifact that the founder has not written and that no automated process in this system is permitted to create, and every row in that store carries the producer's own publication flag set to false.

Layer three — live venue orders. Machinery tests on a personal account since 2026-08-26. No call from the system has been taken to the venue, and there is no graded live profit-and-loss record.

One more thing, and it is the part that decides whether the forward layer is worth anything later. Before any forward number was published, a sample floor was declared in writing: 100 resolved decisions in a card-and-horizon cell before any rate publishes at all, 200 for a pooled or portfolio-level figure, and the rule below the floor fixed as count + maturity date only, never a rate. The floor was set from a stated interval rule rather than chosen once the numbers were visible, and the current record is well below it. When somebody eventually shows you a forward hit rate from this system, the bar it had to clear was written before the number was.

Yes, and that is the part of this record that cannot be edited after the fact.

The oracle posts to a public account, @AetherNexus, under a standing arrangement where entries go out before the outcome is known and results follow afterwards. A post is a timestamped third-party record on somebody else's servers: it cannot be quietly revised, and a losing call that was posted stays posted.

That is deliberately the opposite of the usual arrangement, where a track record is assembled afterwards by whoever is holding the pen. Nothing about a public post makes a system good; it just makes the record harder to launder.

What this is not: a published forward performance number. The posts are individual entries and outcomes. The aggregate record they belong to is held behind the founder's authorization and behind a sample floor that was declared before anyone read the numbers.

The page is a rendering of a versioned artifact rather than hand-written prose, and the version is on it: 1.2.0.

Every number you see here is a pin — an artifact path, a pointer into that artifact, the value expected at that pointer, and the evidence grade that must ride with it. Before this page can deploy, a gate re-resolves every pin against its artifact and refuses the whole deploy if any value has moved, if a stated grade disagrees with the artifact's grade, or if a required disclosure is missing. A number cannot drift out of agreement with its source and survive to publication, because nothing here is typed twice.

The corpus carries a changelog, and every answer carries a stable anchor and its own version, so a link to an answer keeps pointing at the same question and a changed answer is visibly a changed answer rather than a silent substitution. When the underlying books are re-sealed, the answers re-derive from the new artifacts; they are not hand-updated to match.

The limit worth stating: this machinery makes the page agree with the artifacts. It cannot make the artifacts right. That is what the grades, the guards and the unread holdout are for.

Yes. On 2026-10-05 we re-stated every dollar figure on this site, and the True Edge headline went down.

What was wrong. The 2026-08-31 presentation, and the public headline built from it, compounded the sealed trades in the order their decision points occurred, booking each result in that order rather than when the trade actually settled. So a later trade could be sized off a result that had not settled yet. On a card that holds several trades open at once, that changes the compounded figure.

What the rule is now. The settle-at-horizon-end rule: each trade is sized when it opens, from equity that holds only trades already settled; its result is booked when it settles, at its horizon end — the latest it could have, because the sealed rows carry no earlier exit time; nothing compounds before it settles. The sealed trades themselves did not change: the same rows, the same R, the same fees. Only the order in which results reach the account changed.

True Edge, old and new, same book. CAGR at dynamic sizing 136% then, 102.2% now. Cumulative return +8,509,785.9% then, +1,164,983.9% now. Maximum drawdown −43.6% then, −42.1% now. At the flat 1x baseline: +1,043.7% then, 947.1% now, with drawdown −11% then, −9.8% now. The R total underneath did not move: 615.2 R then, 615.2 R now.

The other three cards moved too, for two reasons at once. Their books were also re-sealed on 2026-09-02, after the 2026-08-31 presentation was minted, so their old and new figures differ by both the newer book and the settle rule, and we do not split the two. CAGR at dynamic sizing, then and now: LHCS / Curve 3% to 8.6%, MODEL 0.1% to 8.2%, The Voice 30.9% to 27.2%. Some of those went up and one went down.

We publish the corrections that hurt us in the same breath as the ones that help. The old presentation is kept on disk as provenance and listed in the retired register as not citable; the old figures above are quoted from a census of it, never cited as current.

It did, and the number is on the site rather than in a drawer.

The study put one thing on trial: a fixed rule that would have stood a card aside when a risk signal fired, applied after the fact to already-sealed rows. Its own scope line says what was and was not being tested — NOT on trial: any card's predictive quality — and the thing explicitly not on trial is any card's ability to predict.

Across 64 cells the grades came back 29 insufficient samples, 24 coverage-limited by the book's span, 6 no contact by construction, and 5 scored but not promoted. The key for proven results is not merely zero — it is absent from the grade census entirely, which the artifact records as false, and we assert that absence rather than defaulting a missing key to zero.

What those grades mean in plain terms: on 59 of the cells the study records that the test could not reach a verdict, and on the remaining 5 it records a result measured against a bar set in advance and not cleared. Neither is a card failing a test. The result is underpowered, not absent.

The overlay's cost is also published. Applied across the full tape it shows −109.76 R against True Edge (negative means the Grid cost this card) with 27 false alarms. For the other three cards no overlay figure is shown at all, and each is treated the same way: the sealed study was measured on the previous book vintage, each card's book has since been re-sealed with rows the study never graded, and mixing vintages is never shown. The published overlay value reads not shown for The Voice and not shown for MODEL, and LHCS / Curve's overlay line in place is temporarily not shown; the study re-runs on the new vintage as a registered follow-on. The consequence in the live system is that True Edge is deliberately isolated from this overlay and LHCS / Curve is held with a hearing path rather than excluded.

Both, and they are counting different things. The confusion is real enough that it gets its own answer rather than a footnote.

Nine of ten is True Edge's own evidence cells in its sealed book: ten cells, 9 graded proven and 1 mixed evidence. That is the card measured against its own pre-registered bar on the frozen tape.

Zero of sixty-four is a different object entirely: a separate study of a risk overlay — a rule that would have stood cards aside when a signal fired — applied after the fact to those same sealed rows. Its grade census carries 24 coverage-limited, 29 insufficient-samples, 6 no-contact and 5 scored-not-promoted cells, and the proven key is false from that census.

So one number grades the cards, and the other grades a defensive rule bolted on top of them. The study's own scope line says it in the artifact: the thing not on trial was any card's predictive quality.

Across all four sealed cards there are 32 evidence cells in total: 11 proven, 17 mixed evidence, 4 unproven. Per card, same form: True Edge 9 proven / 1 mixed / 0 unproven; LHCS / Curve 2 proven / 4 mixed / 0 unproven; MODEL 0 proven / 5 mixed / 3 unproven; The Voice 0 proven / 7 mixed / 1 unproven. That distribution is the fair summary of the estate, and it is a long way from every cell proven.

Four things, and you can check three of them yourself.

1. The rules were frozen before the answers were read. The defense study carries a pre-registration seal minted before its first replay, and its repair was issued as a new artifact that supersedes the original rather than editing it, with the superseded file's hash recorded inside the successor.

2. Every card sits behind a regression tripwire contract. Each of the four sealed cards has a contract holding a byte-exact fingerprint of its expected replay result. Any engine change that moves a graded number trips the guard, and the fingerprint can only be replaced through a governed re-seal carrying a frozen before-and-after receipt. True Edge's contract records 97 verification runs and 13 tripwire sweeps; LHCS / Curve's, 83 and 30; MODEL's, 61 and 28; The Voice's, 11 and 3 — it joined the protected set last, and we publish the thin history rather than padding it.

3. The fingerprint is a hash, not an opinion. A fingerprint is accepted only after cold runs of the same sealed inputs produce byte-identical output. Warm or partial runs are explicitly not accepted as evidence, and verification runs execute one at a time because two overlapping runs can invalidate each other's evidence.

4. Every number publishes with its grade, including the ones that lose. Of the 32 evidence cells across the four sealed cards, 11 are proven, 17 mixed evidence and 4 unproven. MODEL's short side sits just below zero — −0.1 R at 3d, −1 R at 7d and −1.2 R at 30d, the last with a −1.3 R drawdown. The Voice's 30-day short side is −8.1 R, graded unproven. The defense overlay shows −109.76 R against True Edge across the full tape (negative means the Grid cost this card). The founding card, LHCS / Curve, has no proven cell on its long side, where most of its return comes from. Those sit on the same pages, in the same type size, as the best number here.

One honest limit: none of this is a third-party audit, and reconciliation of the guard pins across the four cards is not finished — it is closed for True Edge, The Voice and LHCS / Curve, and still open for MODEL with a dated hearing on 2026-08-30. Anyone telling you all four cards are fully reconciled is ahead of the record.

Not one commit — 2, and the receipt says which card sits on which.

True Edge's sealed book was minted at 7368993c9f61e09f3d83bf89400387f8d3dec61a. LHCS / Curve, The Voice and MODEL were re-minted together on 2026-09-02, all three at the same later commit: ea05df409489128d30a54035a954f006ce9f692e, ea05df409489128d30a54035a954f006ce9f692e and ea05df409489128d30a54035a954f006ce9f692e. Earlier vintages were minted across three different commits; the 2026-08-31 full re-replay of all four cards brought that to one, and the 2026-09-02 re-mint of the three siblings moved them to a later head while True Edge's book stayed where it was sealed. The commit recorded on a book is the commit that minted the book; the calls inside the three re-minted books come from the 2026-08-31 full re-replay. So this answer states two commits rather than describing two as one engine.

On currency between reseals: engine files move, and that is the normal condition of a system still being built. What matters is whether a change moved a graded number. Each card's guard contract records a governed rebaseline ceremony for exactly this, and on the most recent one for each card the expected replay result is recorded as unchanged — True Edge true, LHCS / Curve true, MODEL true, The Voice true. For LHCS / Curve the moving edit is named on the receipt as commit 2078f317, bracket rl34pr-20260922, and the bracket recorded that the engine's byte-baseline changed while the expected replay result did not.

The honest limit: that machinery proves an edit did not move the pinned replay projection between reseals. The event that refreshes these numbers end to end is a full re-replay — the most recent one ran 2026-08-31; the three sibling books were re-minted from it on 2026-09-02, and the dollar figures were re-stated on 2026-10-05 under the settle rule without changing any sealed row. Every figure here is stamped with that vintage rather than presented as though it were computed this morning.

41 book vintages sit across 5 retired registers, each by name, each with the reason it was superseded, and 2 earlier dollar presentations are retired the same way — the most recent of them on 2026-10-05, when the dollar figures were re-stated under the settle rule. The first register's law is old books get truth states, never deletion — so a retired book stays readable as evidence of the engine that produced it, and simply may not be cited as current capability.

This is the mechanism that keeps an old flattering number from quietly surviving on a page after the engine underneath it moved. It also means that if you find an older, better-looking figure of ours somewhere, the register is where you check whether it is still allowed to be quoted. Several of the retired entries are more flattering than what replaced them.

It said that, on every book, and we publish it rather than a green check.

The first sealing ceremony, on 2026-08-22T14:33:06Z, recorded 10 books. On 10 of them the provenance field read not resolved, and on 0 did it read resolved. The reasons are named on the receipt: for 6 the source reference inside the book was ambiguous, for 3 the book declared no seal state of its own, and for 1 no machine-readable source reference could be found at all.

Where the four books published today stand: 3 read sealed and eligible and 1 reads sealed over its own refusal. LHCS / Curve and The Voice were re-sealed on 2026-09-02 after their source artifacts were sealed first, and are eligible outright. MODEL was re-sealed the same day and admitted through a written decline that its card publishes: graded on 7 of 9 designed inputs. True Edge's book, sealed 2026-08-31, still carries the refusal seal provenance refused: no resolvable source reference — it is published with that label rather than hidden behind the other three.

What that field is about: an internal back-pointer from a sealed book to its own source declaration. What it is not about: whether the numbers are the numbers. Every book's bytes were hashed before sealing, the hash is on the receipt, and you can reproduce it. The ceremony's own authority line is the sentence that does the most work on this page: a seal is provenance, never a grant.

We could have shipped the rest of this receipt and left this field out. Publishing the whole receipt including the part that looks bad is the only version of this page worth having.

No. Not one.

There is no auditor, no fund administrator, no prime broker, no exchange attestation beyond a personal account's own order history, and no independent quant shop that has reviewed these books. Nobody outside this project has checked these numbers, and any impression to the contrary from the volume of machinery on display would be wrong.

What exists instead is a system built to be checked rather than believed: every sealed book is hashed and its hash is on a ceremony receipt; every card's per-trade rows are published so the aggregates can be recomputed; the rules were registered before the answers were read and the supersession chain is inside the artifacts; 32 evidence cells publish with their individual grades, of which only 11 are proven; and the machinery blocks its own publication until a human authorizes it, with every authority flag hard-false.

That is self-auditing with receipts. It is genuinely stronger than an unaudited claim, and it is genuinely weaker than an audit. A serious diligence process should treat the receipts as the thing to test, and the right test is to recompute something of ours and see whether it lands where we said it would.

The company holds no client funds, manages no money, takes no custody and has no trading authority. Every sealed artifact carries those flags false: autonomous trading false, broker execution false, capital false, deployment false, model promotion false.

The founder, personally, on his own account at a regulated spot exchange, has placed graduation and test venue orders only, beginning 2026-08-26. The record of that session shows a first real fill at minimum venue size with a native protective stop resting, and a live kill drill run by hand. Individual order sizes are not on this page.

The sentence that matters most on this page: no oracle call has been taken to the venue. There are no live fills driven by the system's own signals. What happened was machinery proof — can the software place, protect, tighten, cancel and reconcile a real order on real rails without losing track of it — and it is not, in any reading, a performance record. Anyone describing this as the oracle live-trading is describing something that has not happened.

Order-level receipts are not published, because they describe orders on a personal exchange account and this system's own publication rules bar account and order identifiers from any public file. They are available in a founder walkthrough.

Nothing that moves money, changes a model, or publishes to the world.

The flags are carried on the artifacts themselves, not in a policy document: autonomous trading false, broker execution false, capital false, deployment false, model promotion false. The sealing ceremony states the principle directly: a seal is provenance, never a grant.

In practice that means: no software here places an order without a person, no model gets promoted because it scored well, nothing deploys itself, and no result is published externally without a hand-written authorization artifact that no automated process is permitted to create.

That is a constraint on speed, and it is the point. A system that could promote its own model on its own score is a system whose scoreboard you cannot trust.

Venue. A regulated spot exchange, on the founder's personal account. We name the class rather than the venue on a public page, because this system's own publication rules bar account identifiers and venue identity from public files, and naming the venue beside a personal account is a step toward identifying it. The venue is named in a founder walkthrough. Spot, unleveraged: the sealed books' leverage is a sizing convention applied in measurement, not borrowed money at a venue.

Jurisdiction. Canada. Who may actually invest is a counsel question: exemption, accreditation, and the working rule that the person is qualified before any money is taken. No cheque closes on the strength of this page. Legal characterizations of what is or is not prohibited here route to counsel, like the sibling entity and tax rows.

Is this an offer? No. This page is not an offer to sell securities. It is a personal record of research and machinery tests. There are no managed funds, and nothing here is investment advice. The authority flags on every artifact say the same thing in machine-readable form: false for capital authority, false for broker execution.

Entity structure, incorporation and the instrument itself are answered by the founder directly; they are not in the record and this page does not guess at them.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

The locked forward paper record, as of the 2026-08-29 vintage census: 524 locked rows across 41 distinct observed days from 2026-07-19 to 2026-08-29, of which 387 have resolved and 137 are still waiting on their outcome window, all on BTC/USDT.

Every row in the store still carries the producer's publication flag set to false — 0 rows permit publication and 0 rows permit a claim — and no part of this system overrides a producer's flag from the outside. The founder's written authorization is on file; it authorizes publishing these counts, not a rate.

What this does not publish, and will not even with that authorization: a hit rate or a realized return. Those are performance claims, every rate-class metric here sits well below the sample floor declared before these numbers were read, and a reader must not be able to infer one from published components. The counts answer whether a locked forward record exists, how big it is and what it covers. They do not answer whether it is working, and neither does anything else on this page.

This answer is held. There is no graded live profit-and-loss record yet — no oracle call has been taken to the venue — and the live-money book that would grade one has not been written. What is public: machinery-test venue orders since 2026-08-26, on the founder's personal account, are a wire proof, not a performance record. LYFE has not taken any oracle calls.

Small enough to be embarrassing if you mistake it for a track record, which is exactly why the numbers are here.

The first real order was a minimum-size machinery test on 2026-08-26: 0.00005 BTC ≈ $3.95, filled with a native protective stop resting behind it. The largest was a full three-tier bracket at the smallest notional the venue's own arithmetic permits — 0.00015152 BTC ≈ $11.98 — which entered, rested its stop, held three tiers at their authored allocations, tightened once, and flattened on a receipt. Total fees for the entire live graduation: ~$0.25.

Every one of those orders was afterwards field-verified against the venue's own history, and the verification record reads zero field mismatches.

These are machinery tests on a personal account at a regulated spot exchange. They prove the wire works end to end on real rails with real money. They prove nothing whatsoever about whether the system makes money, and no reader should let the precision of the figures suggest otherwise.

One person. That is the honest headline and it is a real risk, not a modest one.

The system is designed so that nothing runs without a human — every authority flag reads false, no model promotes itself, nothing deploys itself, nothing publishes itself. That is a deliberate safety property and it is also the bus factor: the person who has to be there is the same person every time.

What survives him: the evidence. The sealed books are static files with hashes on a ceremony receipt; 32 graded cells and every per-trade row are already published; the retired-vintage register records what was superseded and why. Somebody else could pick up the record and check it without him, because that is what the receipt discipline was for.

What does not survive him: the build. The engines, the campaign machinery and the judgment that decides what gets measured next are in one head, and no amount of documentation makes that untrue today. Anyone funding this is funding a single operator with an unusually well-documented workbench.

Yes. A holdout covering 2024 onward was sealed before the defense study ran and has not been opened. Its file is present on disk — true — and this system's own tooling records that it was not opened, parsed, or hashed to answer this question: false. The presence check is a filesystem stat and nothing more.

That is deliberate and it is the most valuable asset on this page. Every other number here was produced by people who could, in principle, have looked at the answer first. This one cannot be, because nobody has looked.

What we will not do is tell you what it says, hint at what it says, or open it early because a conversation would go better if we had. When it opens, the result publishes at whatever grade it earns — including the case where it contradicts the sealed books and those grades come down.

Written down here, dated, before the tests that would trigger them.

One — the holdout. When the sealed 2024-onward data opens, if the cells that currently grade proven do not survive on it, those grades come down and the affected book is retired to the register with a truth state. That is the single most likely way this ends, and it is the test we chose to build.

Two — the forward record failing to clear its own bar. A sample floor was declared before the forward numbers were read: 100 resolved decisions in a card-and-horizon cell before any rate publishes at all, and the rule below that floor is fixed as count + maturity date only, never a rate. If the forward record reaches that floor and the result sits meaningfully below the sealed book at the same convention, the sealed book is describing a world that no longer exists and we say so.

Three — a tripwire that cannot be explained. If a guard fingerprint moves without a governed ceremony and a frozen before-and-after receipt, the affected card's verification record is broken and its grades are void until it is rebuilt from scratch.

Four — the ordinary one. Money. The build is founder-funded and the runway question is answered by the founder directly, not by this page.

What is deliberately not on this list: a bad month, a drawdown inside the sealed range, or a card grading unproven. Those are expected outputs of an honest system, not falsifications of it.

The mechanics are already built, which is the only reason this answer can be specific.

The affected book does not get deleted and does not get quietly replaced. It gets a truth state and moves into the retired-vintage register, whose own law reads: old books get truth states, never deletion. A retired book stays readable as provenance of the engine that minted it and becomes uncitable as current capability. The registers already hold 41 retired book vintages across 5 registers from previous rounds of exactly this, so the path is worn, not hypothetical.

Then the affected cells' grades change on every surface that carries them, because those surfaces compute from the sealed artifacts rather than from copied prose. And the result publishes — a failed holdout is a finding, and findings are what this page exists to carry.

They have no current sealed book, and the absence is the answer.

The sealed presentation covers four cards — LHCS / Curve is the first of them, and the set is True Edge, LHCS / Curve, MODEL and The Voice. Merged is false present in it and Consensus is false present in it. The reason is plain: the blend work that would produce a post-overhaul book for them has not run yet.

There is an older Merged figure on the site from a previous vintage. It was correct on the engine that produced it and it stays visible rather than being deleted, labelled as a prior vintage that is not a current-engine book. What we will not do is compute a plausible-looking Merged number today so the table has no gap in it.

The short list, in one place, because it is faster than making you find it.

No custody of anyone's money and no authority to trade it — every authority flag on every sealed artifact reads false. No third-party audit of any kind. No live fills driven by the system's own calls; the venue orders that exist are machinery tests on a personal account. No graded live profit-and-loss record, and no forward rate published at all, because every rate-class forward metric is below the sample floor declared before those numbers were read — 100 resolved decisions per cell is the bar and the rule below it is count + maturity date only, never a rate. No sealed book for the Merged or Consensus cards. No proven cell on the founding card's long side, where most of its return comes from: LHCS / Curve's 2 proven cells are both on the short side and on small samples, and its long side carries 4 mixed-evidence cells and nothing proven. No result yet from the sealed holdout. No automatic failover and no cloud redundancy — a second machine now reproduces the verification runs for the three replay-capable cards byte-for-byte (certified 2026-09-01 at a pinned code version), but the live system itself still runs on one computer. No revenue.

Each of those has its own answer on this page, with numbers. This list exists so that the gaps are as easy to find as the results.

We do not know, and we have not measured it. That is the whole honest answer, so here is what is behind it.

The sealed books measure per-trade R on a 2% shared risk budget. R is a unit normalized to the amount risked: it says how many multiples of that amount a trade returned, and it is deliberately independent of account size. That property is what makes the books comparable across cards, and it is exactly what makes them silent about capacity. There is no market-impact model in the replay, no slippage curve beyond the fees already inside each row, and no order-book depth study.

The live evidence does not help either, and pretending otherwise would be absurd: the venue orders placed so far are machinery tests measured from under a dollar to about twelve dollars.

The asset itself is deeply liquid, and the horizons here are days rather than milliseconds, so the honest expectation is that capacity is not the first binding constraint. But an expectation is not a measurement, and we do not have the measurement. A capacity study is a real piece of work that has not been done.

Who wrote it. The founder, working with AI coding assistants under his direction, over an intense build period. A substantial fraction of the lines were produced by models; the architecture, the laws the system enforces on itself, and every adjudication about what counts as evidence are his. Pretending otherwise would be silly given how this was built, and it would also miss the interesting part: the reason this estate has so much verification machinery is precisely that AI-produced work needs it.

Third-party code. There is a standing policy against adopting third-party runtime components: external material is read as source, security-reviewed, and re-authored in house rather than installed. Foundational dependencies that cannot be avoided are inventoried instead of ignored, and that register currently holds 24 rows. The consequence for a buyer is a small dependency surface and no third-party widget in the published pages.

Ownership. The legal question — who holds title to work produced by a model under direction, and how that flows into an entity — is answered by the founder with counsel, not by this page. It is on the list of things this record does not contain.

Most of this estate runs on one computer. That is a single point of failure and there is no honest way to dress it up.

What that actually costs: the verification runs, the replay campaigns and the live executor all share that machine, which is why they are serialized rather than run in parallel — two overlapping verification runs can invalidate each other's evidence, so one-at-a-time is a correctness rule here and not only a capacity one. A second machine is now kept synced by a receipted deploy ceremony and is certified to reproduce the verification runs for all three replay-capable cards byte-for-byte (certified 2026-09-01 at a pinned code version; any code change to a card's engine re-suspends that card's certification until the second machine re-verifies at the new version). Heavy verification defaults to the second machine, which takes that load off the primary — but it is not failover for the live system. There is no automatic failover and no cloud redundancy.

What is genuinely protected: the evidence. The sealed books, the ceremony receipts and the published per-trade rows are static files under version control with hashes recorded, and copies exist off the machine. A dead machine costs time and momentum; it does not cost the record.

Fixing this is money and a weekend, not research. It has not been prioritized ahead of measurement, which is a defensible choice today and would not be after money arrives.

For build velocity: heavily. For the evidence: not at all, and the difference is the whole answer.

The build runs on commercial AI coding models, and the pace of the last months is not reproducible without them. If pricing changes or a model is retired, the build slows down. That is a real business exposure and it is managed the boring way — work is routed to the cheapest model that can do it and only escalated when the task genuinely needs it, because the founder is paying for this personally.

The sealed artifacts, though, are static files. Re-deriving any published number requires the data and the code, and no AI vendor at all. Nothing on this page depends on a model being available to check it.

And no model has authority here. Every authority flag reads false for autonomous trading and false for model promotion. No AI in this system can place an order, promote a model, or publish a result. The AI writes code and analysis under direction; it does not get a vote on what is true.

Not every one, and that gap is disclosed rather than rounded away.

A machine-readable execution ledger exists: 278 hash-chained rows, 30 distinct order identities, 16 graduation-tier rows at validate and 8 at live. Individual identifiers stay off this page.

The residual gap is narrower than 'no ledger'. During the live graduation at least one venue order does not carry a complete machine write-path receipt. The order itself is visible in the venue's own history and was covered by the field verification pass; what is missing is this system's own internal record of having written it, which is a different and stricter thing than the venue knowing about it.

The exact count of write-path gaps is still being reconciled by the executor's own follow-on work, and this answer deliberately carries no write-path-gap number until that reconciliation produces one that can be pinned. Publishing an approximate count of our own receipt gaps would be precisely the kind of comfortable-sounding number this page exists to avoid.

Why it is here at all: a system whose whole claim is receipt discipline has to publish the places its receipt discipline did not hold.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.

This one is not in the record, so it is not answered here. The founder answers it directly, and the answer arrives with its own receipt rather than as prose. Nothing on this page estimates it, brackets it, or fills it with a placeholder.