Same evaluator
The scalar used during play and the decomposed receipt share one evaluation kernel. A disagreement is a software defect, not a difference of interpretation.
A chess engine that shows its work.
Some mistakes repeat.
Some are just noise.
Luxifare brings the difference to light.
Luxifare examines decisions across your recent games, proves every surviving finding with your own positions, and refuses to invent a pattern when the evidence does not hold.
Private preview · no account before value · every claim remains replayable
The obvious check Nf3+
Keep every piece. Stay winning.
+7.75 Luxifare found Rxd3+!!
Give the rook. Break the king.
+13.82 The forcing idea
- Rxd3+!!
- Bxd3
- Nf3+
- Kc2
- Qxe1
- Re2
- Nd4#
Material + piece squares+5.14
King safety+0.26
Weak squares+0.10
Pawn structure−0.12
The rook is the key. Taking it drags the bishop off e2. Then Nf3+ forks king and queen, and the attack ends in mate. The measured future is +6.07 better than the obvious check.
One decision, opened
First, see the idea.
Then inspect every number behind it.
Nf3+ keeps every piece and stays winning. Rxd3+ gives a rook for one pawn, pulls the bishop off e2, and turns the next knight check into a king-and-queen fork. Luxifare searched both moves to the same depth, then opened the difference concept by concept. Click either move to compare the futures.
Black to move. The attack is winning, but one move ends the argument.
Before the sacrificeBlack to move
Scroll the game continuation to watch the idea unfold.
comparison equal budgetdepth 8 eachcap 250,000 nodes eachscores exact, not estimates
Black to move
The attack is winning. Why throw in the rook?
Nf3+ keeps every piece and stays comfortably ahead. Luxifare found that Rxd3+ can give up the entire rook, pull the bishop off e2, and make the king and queen collapse together.
The continuation played in Luxifare's winexact at depth 8
01Rxd3+!!
Put the rook where it can be taken
The rook takes one pawn and checks. White has only two legal replies: Kc1 or Bxd3.02Bxd3
The bishop accepts the rook
Five points disappear for one pawn. The bishop is pulled off e2, exactly where it was holding the position together.03Nf3+
Fork the king and queen with check
The knight lands on f3 with tempo. The king must move while the queen on e1 is under attack.04Kc2
The king steps into the mating net
White answers the check, but cannot save the queen and keep the king safe at the same time.05Qxe1
Take the queen
The rook is still gone. Luxifare has converted the sacrifice into White's queen and an exposed king.06Re2
White blocks one route
The rook tries to cover the second rank. It cannot stop the knight from closing every escape square.07Nd4#
Finish with mate
The knight returns to d4. Every flight square is covered. The sacrifice ends in checkmate.The rook really stays gone. Its job is to drag the bishop off e2 so Nf3+ arrives with tempo against both king and queen. In the recorded game, Luxifare took the queen and finished with Nd4#.
Why the rook can disappear
The concepts settle the argument.
exact advantage+6.07
Material + piece squares+514cp
King safety+26cp
Weak squares+10cp
Conversion+10cp
Color complex+7cp
Pawn majorities+2cp
Threats+1cp
Safe checks−4cp
Piece mobility−7cp
Endgame king−8cp
Pawn structure−12cp
Named concepts 539cp+exact comparison remainder 68cp=decision margin 607cp
The mating continuation was played by Luxifare in a recorded engine game. The comparison was freshly re-run with the current champion. Both lines reached depth 8 under the same 250,000-node and 1,500ms caps. Both scores are exact completed search results. Stockfish 17.1 also ranks Rxd3+ first at depth 22.
Under every searched line
The explanation is not attached later.
It is inside the score.
At each evaluated position, one kernel measures 32 named concepts: material and piece squares, king safety, mobility, threats, pawn structure, outposts, weak squares and more. The fast path collapses them into one score for search. The trace path reads the same computation back out by name.
Resolved future after Rxd3+ wins the queen
sourcerecorded engine output
fen6k1/pp4pp/3pp3/2p1p2P/2P1q3/P2n2P1/1P1K3R/8 w - - 0 1
phase75 / 256
terms32 + 4 cp rounding = 1315 cp
build01c51a3ba3bd · efb58e872cc0
genome81988c93d18a
captured2026-08-21
ConceptContribution
Hover a concept. The board shows the squares it judges.
Material + Piece Squares
+1106
Piece Mobility
+73
King Safety
+27
Weak Squares
+27
Conversion
+25
Color Complex
+18
Pawn Majorities
+16
Endgame King
+15
Pawn Structure
+4
Bishop Pair
0
Dominance
0
Open Rook Files
0
Σ Total evaluation+1315cp
The 32 displayed integer contributions plus 4cp of per-term rounding equal the evaluation. No hidden strategic term. This is the engine's actual arithmetic.
Whole futures
Luxifare does not call the position after a move “the reason.” It follows the searched line to its resolved leaf, then compares that future with alternatives.
No fake certainty
Sometimes one concept dominates. Sometimes several small advantages decide the move together. Luxifare reports the difference instead of inventing a tidy story.
The safety boundary
Lux can speak.
She cannot change the truth.
Luxifare is the chess engine. It searches, scores, compares and qualifies its own evidence in deterministic code. Lux is the conversational character. She may turn a typed receipt into a fluent coaching sentence, but she cannot add a reason that is absent from the receipt or influence the move.
Luxifare · enginesearch · score · compare
→ typed receiptfacts · bounds · provenance
→ Lux · conversational layerphrase only
The improvement machine
Every improvement must prove itself
before it reaches your board.
Losses become evidence, not vibes. Repeated failure clusters nominate a bounded experiment. Candidates run behind shadow gates, then play external validation matches. Most ideas die. Only measured gains can become the next engine.
01Mine losses
Find recurring suites, phases and concept failures.
02Propose inside bounds
Change an approved parameter family, not arbitrary live code.
03Run in shadow
Ask where a shortcut would have cut, then verify what it missed.
04Play the evidence
Stockfish-first matches reject candidates that merely looked active.
05Promote or remember
Ship a measured gain, or store the failure so it is not rediscovered.
nonlinear evaluator candidate no reliable signal abandoned
stacked search trio −42 Elo rejected
search-memory candidate +10.0 Elo · 12,800 games promoted
An LLM may suggest a typed, bounded experiment. It cannot execute arbitrary evaluator code, edit the live search loop, or overrule the validation harness. The test has veto power. The model does not.
The public record
The rating went down as well as up.
Luxifare played 511 games in public under its own account, unattended, against whoever challenged it. The archive includes the wins, losses, ugly positions and rating drawdowns. It sits around 2400 Lichess rapid. Not world champion strength, and strong enough that its reasons have to survive real tactics.
The bot is paused while the work moves to analysis, so it will not answer a challenge today. Read the games move by move.
What exists, what opens next
No screenshots from the future.
The engine capabilities demonstrated above are real. The public product is not open yet. This ledger keeps engine-complete, product-complete and planned work separate.
Complete-game analysis Upload a PGN or fetch a Chess.com game. Every move gets a searched verdict, deterministic evidence and optional grounded coaching from Lux. Built · access soon
Interactive decision trace The engine protocol already emits the chosen line, re-searched alternatives, concept trajectories, depth history, threats and confidence. Product wiring is next. Engine complete
Player revelation Luxifare tests material decisions across your recent games, rejects false accusations, and admits only replayable proof. Lux then makes you face the same decision in a second game. Built · internal preview
Remembered transfer Your committed answer becomes an evidence-linked episode. On the next reading, Lux remembers whether you recognized the decision, missed it again, or left it unanswered. Built · internal preview
Longitudinal recurrence Compatible future games can test whether an admitted decision recurs, improves, or disappears without turning one short window into a personality verdict. Planned
Counterfactual board Move one piece and watch the accountable terms, not generated prose, change with it. Planned
Adversarial sparring A bot that steers toward the positions your fingerprint says you handle worst. Planned
Private preview
Request a private invitation.
We are opening in small, consenting cohorts while we verify the evidence, usefulness, and safety of every step. Leave your email and we will send your invitation when Luxifare is ready. No account is needed before value.
Invite-only · one invitation when ready · no account before value