Frozen evidence
11 frozen essays
A bounded observation tied to frozen public evidence; it is not automatically causal or general.
Public notebook · 30 essays · evidence snapshot
Read the chain.
Choose an entry point that matches your question, or follow the numbered sequence. Every essay remains available once, with its current public evidence state attached.
Public evidence snapshotAugust 6, 2026 at 6:02 AM UTC
Snapshot chronology / public boundary
Order tells the story.
State tells the boundary.
Essay numbers preserve research order, not publication dates. The state badge is the public evidence boundary for the essay; it does not upgrade a design or interim note because a later result exists.
Three clocks for one evaluation program. Snapshot intervals provide observation bounds, not exact process launch or exit timestamps. No account usage, credential balance, or model-specific dollar attribution is included. These clocks describe evaluation operations and do not support a strength, causal, or terminal DeepSeek claim.
Cutoff August 6, 2026 at 6:02 AM UTC · source identity d6b9d1726427…Evidence key / 30 essays
Three states.
One vocabulary.
Provisional
6 provisional essays
A dated or interim view; it is not a terminal result.
Protocol / design
13 design essays
A protocol, audit, or prospective method; it carries no outcome by itself.
Reading paths / no duplicate links
Start where the
question is.
These paths are thematic entry points through the same numbered chronology. Each public essay appears in one path only; every link keeps its canonical slug.
Path 01 / Measurement
Start with the measurement
Begin with the surprising result, then attach the full-game protocol before reading a rating as a property of a model. 4 essays · 4 frozen.
01
The anomaly
Two Moves Are Better Than Ten
With no candidate help, Luna struggled against a 1320-rated Stockfish opponent. Given ten strong moves to consider, it still struggled. Given only two, the combined system held its own against Stockfish at 2820
Frozen evidence
8 min read1,761 words
02The anomaly
Full Games, Not Chess Puzzles
A chess puzzle begins at the moment the puzzle setter finds interesting. The pieces are already arranged. The tactic exists. The task ends as soon as the right move is named.
Frozen evidence
7 min read1,497 words
03The anomaly
Elo Is a Property of a Protocol
Ask for a chess player's Elo and the question sounds ordinary. Ask for a language model's chess Elo and almost every noun becomes ambiguous.
Frozen evidence
10 min read2,086 words
04The anomaly
The Candidate-Set Paradox
More information should help.
Frozen evidence
9 min read2,020 words
Path 02 / Diagnosis
Find the failure inside the score
Trace candidate assistance, interface behavior, and the distinct mechanisms that one aggregate Elo number cannot separate. 7 essays · 4 frozen · 3 design.
05
Diagnosis
Is the Model a Searcher, a Critic, or a Policy?
A chess move arrives as one action, but producing it requires several different computations.
Protocol / design
9 min read1,980 words
06Diagnosis
Where Does the Game Fall Apart?
Two agents lose to the same opponent.
Frozen evidence
8 min read1,688 words
07Diagnosis
The Price of Thinking
A model can be cheap on a price sheet and still be impractical inside an agent.
Frozen evidence
12 min read2,713 words
08The instrument
Building a Chess-Native Reasoning Architecture
The easiest response to a weak chess-playing language model is to give it more chess text, more tools, more reflection, and more tokens.
Protocol / design
22 min read4,846 words
09The instrument
Candidate Curves as Model Fingerprints
Two chess agents can have the same unaided score and still be very different reasoning systems.
Frozen evidence
11 min read2,525 words
13Research discipline
Ideas or Vision? Why the Score Alone Cannot Tell
Suppose two agents find Stockfish's preferred move 31.5% of the time.
Protocol / design
10 min read2,244 words
16Laboratory operations
The Tool Call That Looked Like Causality
Here is a result that practically writes its own misleading headline.
Frozen evidence
12 min read2,626 words
Path 03 / Design
Make the intervention falsifiable
Follow the benchmark, pre-label audit, and preregistered decisions that define what a later outcome would mean. 5 essays · 5 design.
10
The instrument
An Open Benchmark for Tool-Mediated Chess
Release status: v0.1 draft. This document specifies a benchmark; it does not announce a terminal benchmark release. The Luna and DeepSeek ladders are exploratory pilot evidence, not v1 leaderboard submissions.
Protocol / design
16 min read3,497 words
11Research discipline
The Experiment We Refused to Run
The dangerous moment in an experiment is not always when the result appears. Sometimes it is the moment just before an expensive job starts, when the code looks complete, the sample is frozen, and everyone is t
Protocol / design
8 min read1,694 words
12Research discipline
Why Not Skip Straight to Higher Elo?
The first games looked absurdly easy.
Protocol / design
9 min read1,896 words
17Laboratory operations
The Control Arm Was Full of Chess
The most misleading condition in our next experiment is called no_information.
Protocol / design
8 min read1,763 words
18Laboratory operations
The Analysis That Existed Before the Results
We had already designed the intervention.
Protocol / design
8 min read1,661 words
Path 04 / Audit
Protect the evidence path
Follow the byte-level, transport, label, and no-result gates that stop a working protocol from becoming a stronger claim than its evidence. 5 essays · 5 design.
19
Laboratory operations
The Bytes We Could Prove
A105 began with an apparently simple promise: change whether a model can read a board panel or a complete legal-action panel, and change nothing else.
Protocol / design
6 min read1,357 words
20Laboratory operations
The Request That Failed Before the Network
The previous systems audit ended at an intentionally artificial boundary. A105-R2 proved that our repository adapter could carry 128 canonical treatment cells into two in-memory serialization profiles without a
Protocol / design
7 min read1,416 words
21Laboratory operations
The Patch That Did Not Erase the Failure
A failure discovered before outcomes creates a temptation: fix the code, rerun the test, and describe the system as if the failure never happened.
Protocol / design
6 min read1,326 words
22Laboratory operations
The Engine That Had to Wait
It is easy to describe position labeling as housekeeping. Collect some chess positions, ask a strong engine whether each one is tactical, ambiguous, or quiet, and balance the final benchmark across those catego
Protocol / design
9 min read1,912 words
25Laboratory operations
The Result That Is Still Zero
There is a satisfying kind of research progress that produces no number.
Protocol / design
4 min read847 words
Path 05 / Chronology
Read the later snapshot without upgrading it
Track operations, shuffled inputs, transport repairs, and later results in numbered order; the evidence badge remains the public boundary. 9 essays · 3 frozen · 6 provisional.
14
Laboratory operations
Five Workers, One Clock: Why LLM Chess Runs Feel Slow
At one point in our DeepSeek chess run, two numbers seemed impossible to reconcile. The token counter was reported as 19.277 million. The runtime was described as roughly 117 hours. Yet the experiment had not b
Provisional
9 min read1,951 words
15Laboratory operations
The Result That Refused to Exist
At the 2026-08-04 drafting boundary, the most important output of our DeepSeek chess evaluation did not exist.
Provisional
11 min read2,404 words
23Laboratory operations
The Curve That Runs Backward
The most tempting graph in this project slopes the wrong way.
Provisional
8 min read1,659 words
24Laboratory operations
The Shuffle That Looked Like a Signal
The naive experiment is easy: show one list, shuffle it, ask twice, and count changes. It is also wrong. A model can change its answer even when the prompt is byte-for-byte identical. Without measuring that bas
Frozen evidence
2 min read446 words
26Laboratory operations
The Result That Arrived Late
We tried to measure a prompt-guidance effect twice, paid for sixteen chat intents, and refused to read either attempt as an experiment.
Frozen evidence
4 min read833 words
27Laboratory operations
The Prompt Did Not Whisper
E026's live transport aborted before any chat return; E027's chat transport succeeded for ten returns, but metadata attestation failed on the immediate lookup. E028 then ran as a fresh response cohort with zero
Frozen evidence
5 min read955 words
28Laboratory operations
The Single-Source Oracle Was Exact. The Protocol Wasn't.
An exact answer from a single frozen source does not make an experiment exact.
Provisional
5 min read1,128 words
29Laboratory operations
Two Exact Moves Changed the Submitted Policy
E029 asked a narrower question than a game ladder can answer. If Luna receives the same board but also sees two unranked moves drawn from an exact Syzygy policy, does its submitted move preserve more 50-move-ru
Provisional
4 min read695 words
30Laboratory operations
Two Unranked Engine Moves Helped Luna Survive
The sharp question was not whether an engine can play chess. It was whether a small engine-generated action scaffold changes what the same language model can do when everything else stays fixed.
Provisional
3 min read673 words