Scope: four package figures + one gallery-specific A113 figure + a separate E025 route figure. Every chart carries the evidence state that determines what it can—and cannot—support.
01 / LunaTerminal frozen
Terminal frozen, exploratory, and protocol-specific.
02 / DeepSeekInterim only
Historical interim, unfrozen, and provisional snapshots from the pre-terminal synthesis; terminal DeepSeek data and publication authorization are represented separately.
03 / MechanismNot yet causal
No causal candidate-count or cross-model claim.
The visual record
Four package figures. One gallery-specific A113 figure.
Each SVG is deterministic, checksum-pinned, and provider-free. Captions and assets are copied from a fixed allowlist; operational files never enter the site. E025 remains a separate route figure; this gallery intentionally does not duplicate its SVG. Open the separate E025 route figure ↗Open the public claim boundaries →
01
Luna / terminal evidence
A completed protocol curve
Terminal · frozen
luna-terminal-rating-curve. Frozen Luna joint rating estimates with profile 95 percent intervals for regular, top-10, top-5, top-3, and top-2 variants. Frozen Davidson-WDL estimates. Exploratory, Black-only, adaptive-anchor protocol; regular is not k=0 on one treatment curve.
Figure annotation
Point estimates are frozen Davidson-WDL local protocol ratings; profile 95% intervals show the uncertainty attached to each complete condition.
This is an exploratory, Black-only, adaptive-anchor systems curve. Regular is a separate interface condition—not k=0—and the separately generated, unpaired candidate conditions do not identify a causal candidate-count effect.
Text alternative
Frozen Luna joint rating estimates with profile 95 percent intervals for regular, top-10, top-5, top-3, and top-2 variants. Point estimates are frozen Davidson-WDL local protocol ratings; profile 95% intervals show the uncertainty attached to each complete condition. This is an exploratory, Black-only, adaptive-anchor systems curve. Regular is a separate interface condition—not k=0—and the separately generated, unpaired candidate conditions do not identify a causal candidate-count effect.
Artifact identity
Deterministic SVG · no network or provider calls · SHA-256 211baf118e2ffa15a1954b9d31d975bd976e3fe2a776be20fae9d759d1044b9e
02
DeepSeek / interim evidence
A curve still under measurement
Interim · unfrozen
deepseek-interim-rating-curve. Interim DeepSeek rating point estimates with working profile intervals for regular, top-10, top-5, top-3, and top-2 variants. Every value is provisional and unfrozen. Working intervals; adaptive anchors and repeated looks are not corrected. No terminal DeepSeek ranking or cross-model claim.
Figure annotation
Point estimates and working profile intervals are a dated historical synthesis from before the terminal freeze. Terminal data, a rating, and the joint E000/E001 qualification exist. The public DeepSeek cohort remains interim; public Elo and ranking are withheld, publication is unauthorized, and request-level billing attribution is unavailable.
Historical interim and unfrozen public projection. Terminal content-addressed data, rating, and joint qualification exist; public Elo is withheld, publication is unauthorized, and request-level billing attribution is unavailable. No public DeepSeek ranking or cross-model claim is supported.
Text alternative
Interim DeepSeek rating point estimates with working profile intervals for regular, top-10, top-5, top-3, and top-2 variants. Every value is provisional and unfrozen. Point estimates and working profile intervals are a dated historical synthesis from before the terminal freeze. Terminal data, a rating, and the joint E000/E001 qualification exist. The public DeepSeek cohort remains interim; public Elo and ranking are withheld, publication is unauthorized, and request-level billing attribution is unavailable. Historical interim and unfrozen public projection. Terminal content-addressed data, rating, and joint qualification exist; public Elo is withheld, publication is unauthorized, and request-level billing attribution is unavailable. No public DeepSeek ranking or cross-model claim is supported.
Artifact identity
Deterministic SVG · no network or provider calls · SHA-256 4d6267e4b50f38a8b86e90fe5a26f06ccff64a5d05047c6cb00c7f7bf2547b73
03
Runtime / accounting lenses
Four clocks, not four parts
Evidence boundary
deepseek-runtime-accounting. Horizontal bars compare suite-worker hours, reported reply hours, active wall hours, and five-way-equivalent hours. These accounting quantities overlap and must not be added. Accounting categories are alternative lenses, not additive components; worker-hours are not elapsed human wait time.
Figure annotation
Suite-worker, reported reply, active wall, and five-way-equivalent hours are alternative denominators for the same evaluation operations.
These accounting lenses overlap and must not be added. Worker-hours are not elapsed human wait time, and the snapshot does not support a strength, causal, or terminal DeepSeek claim.
Text alternative
Horizontal bars compare suite-worker hours, reported reply hours, active wall hours, and five-way-equivalent hours. These accounting quantities overlap and must not be added. Suite-worker, reported reply, active wall, and five-way-equivalent hours are alternative denominators for the same evaluation operations. These accounting lenses overlap and must not be added. Worker-hours are not elapsed human wait time, and the snapshot does not support a strength, causal, or terminal DeepSeek claim.
Artifact identity
Deterministic SVG · no network or provider calls · SHA-256 901988f21128c2cdd56264d1f09d3e2a4b84878767e14a3bbf55a9fa67895f8c
04
Inference / gate pipeline
A claim advances only with its evidence
Evidence boundary
evidence-gate-pipeline. Three evidence lanes show that frozen Luna results support protocol-specific description, unfinished DeepSeek results require terminal freezing, and causal candidate-count claims require a paired nested-set experiment with independent evaluation. A leaderboard pattern is evidence to design a causal test; it is not itself a causal result.
Figure annotation
The lanes show what each evidence state can unlock: frozen results support protocol description, while the public DeepSeek cohort remains interim and causal claims remain gated.
This is a gate map, not an outcome chart. A candidate-count mechanism claim still requires nested candidate sets, paired positions, and an independent evaluator; the paired, nested replication is preregistered but not represented here.
Text alternative
Three evidence lanes show that frozen Luna results support protocol-specific description, unfinished DeepSeek results require terminal freezing, and causal candidate-count claims require a paired nested-set experiment with independent evaluation. The lanes show what each evidence state can unlock: frozen results support protocol description, while the public DeepSeek cohort remains interim and causal claims remain gated. This is a gate map, not an outcome chart. A candidate-count mechanism claim still requires nested candidate sets, paired positions, and an independent evaluator; the paired, nested replication is preregistered but not represented here.
Artifact identity
Deterministic SVG · no network or provider calls · SHA-256 027abc4ecabd13c8c96bdd1115e9c6427e07460146b37172a9216bbe3d1ec171
05
A113 / terminal result figure
A paired engine-scaffold survival effect
Terminal · content-addressed
A113's route-specific aggregate figure: 319 complete source-puzzle pairs, +11.60 percentage points, and a +7.52 to +15.99 percentage-point source-game bootstrap interval.
Reading boundary
This figure attributes a local survival effect to the Luna model plus a fixed-budget Stockfish 18 proposer harness. It does not establish model learning, Elo, unassisted strength, general transfer, or a publication authorization.
A113 paired engine-scaffold effect: the scaffold improved complete-episode survival by 11.60 percentage points, with a 95 percent bootstrap interval from 7.52 to 15.99 points. The terminal public aggregate contains no puzzles, moves, prompts, responses, or provider identifiers.
