STRATEGY DIFF ANALYSIS: 5,690 CHANGES, 21 REAL KEYS
A strategy diff analysis cut 5,690 changed values to about 21 causal keys, found one switch costing 2,086.80, and reproduced the best cell bit for bit.
A strategy diff analysis found about 21 meaningful changes inside 5,690 noisy ones. The operator handed the campaign a new base setfile, and the honest first move wasn't to run mutations. We diffed it against the champion first.
The raw diff produced 5,690 changed values. A blind batch against that surface would have been a coin flip with extra steps: thousands of changed keys, no idea which ones were load-bearing, and every "improvement" hard to separate from noise. The measurement that followed shows how we found the lever and how much of the rewrite was decoration.
The funnel went from 5,690 numeric differences to roughly 21 causal keys. Between them sat the inert wiring that changed nothing and carried no behavioral difference. The same funnel then made a prediction, measured it, and confirmed it.
The method: diff, classify, and measure the classification. Our bit-identical control supplies the proof standard for that last step.
#Strategy diff analysis starts by counting 5,690 changed values
The diff itself was arithmetic. Comparing the new base against the champion key by key produced 5,690 real numeric differences. No interpretation, no ranking, just the count.
The raw count set a hard limit on the story we could tell. Five thousand changed values weren't a tuning change; they were a rewrite. A rewrite needs classification before it needs a batch.
The classification runs on behavior, not names. A key is inert if changing it back produces a run bit-identical to the base: the same profits, drawdowns, and trade counts on all four ranges, to the cent.
A key is causal only if flipping it moves a number. Everything else is a hypothesis waiting for its run.
#The funnel found 3,947 inert keys and about 21 causal keys
The classification came back lopsided when run against all four ranges. The sell-side wiring (3,772 keys) and the buy-side wiring (175 keys) were behaviorally inert.
That left 3,947 changed values doing nothing at all. The new base and the champion traded identically on every range no matter what those keys said. The whole behavioral difference was about 21 keys: five of the six levers carried by the campaign's best-known cell, plus one switch those levers depend on. The measurement follows the clean separation demanded by our three-phase testing protocol.
| Funnel stage | Keys | What the runs showed |
|---|---|---|
| Numeric diffs, base vs champion | 5,690 | The raw count — a rewrite, not a tune |
| Sell-side wiring | 3,772 | Inert — bit-identical on all four ranges |
| Buy-side wiring | 175 | Inert — bit-identical on all four ranges |
| The causal remainder | ~21 | The entire behavioural difference between the two files |
The inert majority wasn't wasted work. It was the reason the search could concentrate. A wave that mutated the new base's differences would have spent its four probes per axis fishing in 3,947 dead keys.
The inert keys were inert on these four ranges, with these logics. The classification describes a measured surface, not a law of nature. The registers carry it as a dated finding, and a new range set would reopen the question.
#Why the inert majority was not obvious
The diff predicted the shape before any run. The numerical base-diff described the new base as the champion plus the sell-side wiring changes, with no exclusive keys of its own. That was a testable claim, and the funnel's job was to test it.
Two early waves read the claim wrong in an instructive way. They mutated mode and identity keys on disabled logics, concluded the mode paths were blocked, and wrote it down. The next wave found the paths alive on a dormant slot with a live reference, but unbankable for a different reason: hazard lots.
The register entry was corrected rather than deleted. The modes were never blocked; the carriers were wrong.
A diff tells us which values changed; a classification tells us which changed values can matter. The second claim earns its place with runs.
The 3,947 inert keys weren't the ones that merely looked inert. Several had dramatic names, like hedge scale and reverse scale. They came back bit-identical when mutated, which is the only definition of inert we accept.
Skipping the step has a standing cost. A wave that fishes for improvements among dead keys spends its four probes learning nothing, and the record fills with "no effect" rows that hide the two rows that mattered.
#The new base's own surface sums to 74,665.36 across 3,297 trades
With the classification done, we had mapped the new base's own behavior. It became the reference surface every later cell was measured against. All eight direction and mode master switches ship on, making this the first base in the campaign where the sell-side hedge and reverse paths were reachable at all.
| Range | Profit | Drawdown | Trades (buy / sell) |
|---|---|---|---|
| R1 | 28,165.27 | 12,533.04 | 1,209 (1,126 / 83) |
| R2 | 24,124.74 | 27,931.71 | 842 (742 / 100) |
| R3 | 9,040.20 | 10,361.51 | 342 (270 / 72) |
| R4 | 13,335.15 | 47,019.11 | 904 (830 / 74) |
| SUM | 74,665.36 | 97,845.37 | 3,297 (2,968 / 329) |
Reachable wasn't the same as useful. Hedge and reverse could only add on a slot that contributes zero today, and the base had exactly five such dormant slots.
One dormant slot, enabled with its shipped preconfiguration, fired 9,189 trades on the first range (7.6× the control) and then hit margin exhaustion. FreeMargin fell to −2,981.68, and the run terminated at bar 160 of 276 with a final balance of 31,155.93. Its −68,844 was a floor, not the full-window loss.
A sibling activation summed −257,685.07. Both registered dead. The cheap dormant grids were inert rather than small: one added 3 trades on the first range and 1 on the fourth. The tight grid was starved by the side it trades, not poisoned by its spacing.
This is the risk frontier: a reachable path can be live and still fail on margin.
#Eight wiring families accounted for 3,772 inert sell-side keys
The 3,772 sell-side keys weren't anonymous. The numerical diff named them, and the families matched wiring the campaign had already been circling: hedge scaling, reverse scaling, compounding type, break-even mode, partial-close percentages and balances, grid type, and reverse-enabled flags.
These families had never been individually measured before this campaign. They were present on the sell logics of the new base, and measured inert here.
| Wiring family | Keys | Verdict on this base |
|---|---|---|
| Partial-close balances (four slots) | 1,260 | Inert — bit-identical on all four ranges |
| Compounding type | 420 | Inert |
| Break-even mode | 420 | Inert |
| Partial-close percent | 420 | Inert |
| Hedge scale | 418 | Inert |
| Grid type | 380 | Inert |
| Reverse scale | 399 | Inert |
| Reverse-enabled flags | 98 | Inert |
#One reverse-enable switch cost 2,086.80 on the second range
Of the roughly 21 causal keys, one switch did the most measurable damage on its own. The reverse-enable key, sitting at 1 on the new base, cost the second range 2,086.80 and the fourth range 1,070.05 against the champion's best cell.
This wasn't a rounding error or a tuning dial. It was a mode switch, and the campaign's measured position on mode switches was already on the record: a mode conversion on a live logic lost 68,255.98 on the sum, because a mode replaces a logic's grid instead of adding to it.
The funnel paid for itself on exactly this key. Flipping it back to 0 and applying the sixth lever, the one lever the best-known cell was missing, reproduced the champion's best cell bit for bit.
The identity was replicated twice. The first time as a reconstruction, the second through the proper builder. Two paths to the same bit-exact numbers turned "the whole advantage is ~21 keys" from a diff story into a measurement.
#The clean classification predicted that the champion's grid ladder would not transfer
A classification this clean makes predictions, and the campaign wrote one down before running it. If the base really was the champion plus inert wiring plus a handful of causal keys, the champion's tuned axes should not transfer. A grid ladder whose peak was measured on the champion's values had no reason to peak on a base with a different behavioral core.
The prediction was recorded, then measured.
| Grid ladder on the new base | Direction | Measured result |
|---|---|---|
| One step tighter | Down from the base value | −204,623 — poison |
| Base value | — | Peak on this base, on this base alone |
| One step wider | Up from the base value | −6,525 |
| Sibling grid, one step tighter | Down from the base value | −107,489 |
| Sibling grid, base value | — | Peak |
| Sibling grid, one step wider | Up from the base value | −566 |
Every step was worse in both directions on both ladders. The champion's registered peak for the sibling grid was a cliff here, while the champion's peak for the first grid was already the base's own value.
The champion's ladder did not transfer. The numerical diff predicted that before a single cell ran, and the cells measured it to the cent. A base isn't a champion with extra keys. A base is its own surface.
#The funnel narrowed the search but did not finish the campaign
Three things are now defensible. The base's own surface sums to 74,665.36 on 3,297 trades (2,968 buy, 329 sell) against the 78,111.46 the best cell on the champion's grid stack had reached.
The causal core is about 21 keys, and flipping one of them back reproduces the best-known cell exactly. The tuned axes are local to their base, so the next wave starts from measurement, not inheritance.
The funnel's last deliverable is honest scope. The sell-side machinery is measured, named, and mostly inert. The causal core is 21 keys and reproducible, while the axis ladder is local to the base.
The funnel did not finish the campaign. The second range cleared its deposit bar by 928.35 at best. The third was 14,716.56 short, and the fourth 10,308.80 short. You can compare that scope with the published DAAVILEFX results.
Every value ladder on this base was exhausted. What remains is whole-mechanism work that has never been run, plus one blocker no diff can move: the drawdown guard, which is an operator ruling rather than a key. The funnel got the search to the right battlefield. It cannot win the war.
- Count the changes before changing them: 5,690 diffs classified before any mutation ran.
- Test the causal core against the base's own measured surface: about 21 keys, one of which reproduces the best-known cell bit for bit.
- An inert classification is a dated finding on four ranges; a new range set reopens the question.
Open the quantum dashboard to run your own diffs against live wave data.
#Questions we get asked
#How many strategy changes actually mattered?
About 21 keys. The diff narrowed from 5,690 changed values through 3,947 measured-inert keys to about 21 causal keys.
#Which key cost the most?
The reverse-enable switch at 1. It cost 2,086.80 on the second range and 1,070.05 on the fourth, making it a mode change rather than a small tuning tweak.
#Did flipping the switch back reproduce the best result?
Yes, exactly. With the sixth lever applied, the base reproduced the champion's best cell bit for bit, and the identity was confirmed twice.
#Did the strategy diff analysis finish the campaign?
No, it narrowed the search. The second range cleared its deposit bar by 928.35 at best, while the third was 14,716.56 short and the fourth 10,308.80 short.