YUREI RESEARCH
Composition
9 published labnotes
2 inconclusive1 mixed5 negative1 positive
OPEN RESEARCH MAP →YUREI RESEARCH
Research library
Field dispatches, protocols, labnotes, and results from Yurei's experimental work. The stories may be lively; the evidence stays inspectable.
62 LABNOTES1 DISPATCHES9 RESEARCH AREAS
654ccfc93611 REVISION
01
Research areas
Browse published experiment families.
Multiview Cameras
6 published labnotes
2 mixed4 positive
OPEN RESEARCH MAP →Narrative Steering
2 published labnotes
1 mixed1 positive
OPEN RESEARCH MAP →Prosody
18 published labnotes
4 inconclusive3 mixed8 negative3 positive
OPEN RESEARCH MAP →Vehicle Navigation
2 published labnotes
1 mixed1 positive
OPEN RESEARCH MAP →Voxel Guidance
6 published labnotes
4 mixed1 negative1 positive
OPEN RESEARCH MAP →Zenith
17 published labnotes
4 mixed8 negative5 positive
OPEN RESEARCH MAP →02
Dispatches
Human-readable debriefs around the work. Formal labnotes remain the evidence record.
dispatch-001
FIELD DISPATCH / published
After-action report: the truck drove, and the cameras kept receipts
We let the game’s own recorded route take the wheel in a headless test. The truck moved, the waypoint loop wrapped, and four software-rendered camera feeds left a time-stamped trail.
03
Labnotes
Published records ordered by research date.
multiview-cameras-006
completemixed
Operational vision through damage-aware wireframes and normal MCP capture
Can the read-only Voidwright interface expose a damage-aware structural view and an inspectable HoloView image through ordinary Codex MCP tools?
multiview-cameras-005
completepositive
Client-visible markers for virtual camera poses
Can a connected Space Engineers client display a virtual camera's position and orientation as GPS markers that track a moving vehicle?
multiview-cameras-004
completepositive
Live multiplayer observation and moving-grid camera tracking
Can Voidwright observe a connected human and a human-driven moving vehicle in a shared Space Engineers world, keep a grid-relative virtual camera attached, and return provenance-bound HoloView captures?
game-session-observation-001
completemixed
V3 human-driven valley recording and server evidence
What can a bounded desktop recording and dedicated-server logs establish about a human-driven V3 valley session?
vehicle-navigation-002
completemixed
V3 Red Truck depth-guided GPS attempt
Can the V3 Red Truck's depth-guided controller reach the End Position GPS within a bounded low-speed operation?
vehicle-navigation-001
completepositive
Repeated V2 color-zone waypoint baseline
Can the V2 Blue Utility Truck repeatedly reach the four color-zone GPS targets and return to Safe Spawn on the open proving-ground platform using bounded wheel guidance?
multiview-cameras-003
completemixed
Sparse terrain-height reconstruction over a stitched vehicle run
Can receipt-backed front-depth captures be registered to a long Voidwright drive trace to produce a useful, explicitly sparse terrain-height overlay?
multiview-cameras-002
completepositive
Paired HoloView capture during a bounded vehicle approach
Can the current Voidwright server controller bring a land truck to a defined approach radius while paired onboard and overhead HoloView feeds record the complete bounded trial?
multiview-cameras-001
completepositive
Bounded HoloView capture in a headless dedicated server
Can a headless Space Engineers server export a bounded camera-frame sequence with pose/tick provenance and renderer-side scene context, without native framebuffer access?
voxel-guidance-006
completemixed
Temporal synthetic spatial-perception calibration
Do short grayscale frame sequences plus ego-motion improve local spatial estimates over a matched single-frame baseline on held-out procedural geometry?
meta-research-001
completepositive
Grounded build-time labnote summarization
Can a versioned AI apparatus produce concise labnote orientations that preserve findings, limitations, and negative results without introducing unsupported claims?
composition-009
completenegative
Readiness-gated bounded-span repair
Can suffix-triggered readiness plus bounded-span repair recover the deferred-composition idea without freezing incorrect surrounding structure?
composition-008
completenegative
Forced deferred-hole composition pilot
Does deferring one locally uncertain span until right-hand context exists improve bounded composition over direct generation or unrestricted two-pass revision?
prosody-018
completepositive
conversation-conditioned two-speaker pilot
Can conversation-aware turn direction produce a convincing two-speaker exchange where isolated line direction does not?
composition-007
completenegative
Matched deep-thinking composition campaign
Does enabling extended model thinking improve bounded composition enough to justify its latency and generation overhead?
prosody-017
completemixed
multi-utterance prosody candidate pool
Can a small multi-utterance candidate pool supply natural repeated readings for a held-out fingerprint-retrieval test?
prosody-016
completepositive
acoustic fingerprint pilot
Do lightweight acoustic fingerprints recover perceived emphasis and expressive delivery in the Prosody 015 corpus?
prosody-015
completemixed
emergent prosody annotation
What coherent readings actually emerge when direct lexical-focus control fails?
prosody-014
completenegative
contrastive-focus instruction-form ablation
Does removing the competing focus word from Qwen's instruction restore directional contrastive emphasis?
prosody-013
completenegative
instruction-conditioned contrastive focus
Can reference-free natural-language instructions produce natural, directionally correct contrastive focus?
prosody-012
completenegative
reference-conditioned contrastive-focus transfer
Can whole-utterance reference conditioning reproduce an explicitly demonstrated contrastive focus without the artifacts of post-hoc DSP?
voxel-guidance-005
completenegative
Semantic namespace token ablation
Can a semantic namespace token replace the full field dictionary when a local model interprets compact behavioral state?
voxel-guidance-004
completemixed
Machine-native state serialization baseline
Can two-minute behavior windows be represented compactly without loss, and does one local model preserve task recognition across prose, JSON, and typed tokens?
voxel-guidance-003
completemixed
Current-task feature-family ablation
Which observable behavior families carry the 120-second held-seed task-inference signal?
voxel-guidance-002
completepositive
Blinded current-task inference
Can the current Minecraft task be inferred from causal behavior while holding out an entire seed block?
voxel-guidance-001
completemixed
Minecraft behavioral-vector pilot
Can repeated, task-balanced Minecraft sessions produce an interpretable behavioral vector whose within-condition variation is smaller than its uncertainty?
narrative-steering-002
completemixed
Fingerprint compiler and blinded scorer calibration
Which pilot steering coordinates are stable across story states, and can a blinded local model scorer reproduce the human measurements well enough to reduce future review burden?
narrative-steering-001
completepositive
Matched-state narrative steering vector pilot
Can blinded ratings of continuations from matched narrative states recover a repeatable directional steering vector for each model, distinct from story and prompt effects?
zenith-017
completenegative
swamp localized panel holdout
Does snowy-Hero successor candidate `3d37e24ca4e8b57038724062fe87a370c98ec208d5bc0f1663aeb907e40b50f8` generalize to a fresh balanced swamp holdout?
zenith-016
completepositive
snowy Hero development fold
Does adding independently collected Hero-only Hoelbrak development data repair the snowy Hero weakness exposed by Labnote 015 without regressing Inventory?
zenith-015
completemixed
snowy localized panel holdout
Does deterministic localized candidate `bc155807a28aee1a2d7d10fb18a96b24d40b9fe300fc3e01650f70e5d61cc649` generalize to a fresh balanced holdout captured during the snowy Claw of Jormag encounter?
zenith-014
completepositive
deterministic localized transfer
Does deterministic, class-balanced full-batch multiple-instance training retain Hero generalization after adding varied Inventory-only position data?
zenith-013
completemixed
varied Inventory position transfer
Does adding new Inventory-only development batches staged at left, right, and center positions repair the localized candidate's position dependence?
zenith-012
completenegative
localized panel frozen holdout
Does the localized panel candidate that passed development cross-validation generalize to a separately frozen, balanced live Small-UI holdout?
zenith-011
completepositive
localized panel transfer
Can class-specific localized evidence avoid the environmental confounding seen by whole-tile transfer learning while preserving development-only evaluation?
zenith-010
completenegative
corrective panel transfer
Does adding an approved Inventory-only and closed-panel development batch fix the cross-batch false-positive behavior of the frozen MobileNet V3 Small panel classifier without spending a fresh holdout?
zenith-009
completenegative
panel transfer baseline
Does a frozen ImageNet-pretrained MobileNet V3 Small backbone provide enough visual structure to recognize Hero and Inventory across development batches without spending another holdout?
zenith-008
completenegative
first visual panel baseline
Can a compact CPU ridge classifier trained on masked Small-UI panel tiles generalize from two development batches to a recognition-disabled frozen holdout captured during a live world boss?
zenith-007
completemixed
Small-UI panel recognition
Can Zenith Vision recognize a concrete open Guild Wars 2 UI panel in a messy live scene without retaining the full frame or emitting unrelated screen text?
zenith-006
completepositive
live one-shot observation
Can Zenith Vision execute its complete live privacy and authority boundary once against Guild Wars 2 without allowing an optional OCR failure to suppress valid MumbleLink telemetry?
zenith-005
completepositive
Normal profile blind human validation
Does the frozen Normal UI crop profile agree with independently drawn human boxes on fresh gameplay frames?
zenith-004
completemixed
frozen Normal profile seeded validation
Does the frozen Normal UI crop profile remain visually acceptable across a fresh, independently collected gameplay batch?
zenith-003
completenegative
UI-scale stratification
Does one generic default-layout geometry describe the visible Guild Wars 2 HUD across the Small, Normal, and Large UI profiles?
zenith-002
completenegative
batch holdout geometry confirmation
Does the default-layout geometry baseline recover operator-reviewed visible HUD regions across a larger private gameplay holdout?
zenith-001
completenegative
default-layout geometry on a real holdout
Can the deterministic default-layout proposal serve as a UI-region detector, or is it only suitable for conservative crop routing?
prosody-011
completeinconclusive
Reproducible synthetic conversational-audio rerun
Does the negative Labnote 003 matched-versus-shuffled result persist when every audio clip is generated locally by a known, fingerprinted synthesizer and exact transcript fidelity is verified before downstream generation?
prosody-010
completenegative
native stress prominence gate
Does native lexical stress create reliable directional discourse prominence?
prosody-009
completemixed
native Kokoro stress pilot
Can native Kokoro lexical stress produce natural audio and validate the intended focus mapping?
prosody-008
completenegative
audible focus-control gate
Can post-hoc focus control create audible, natural emphasis before spending human-review labor?
composition-006
completemixed
Targeted residual-defect repair
A second model pass may justify its cost when restricted to repairing one explicit residual defect in an otherwise complete direct rewrite, even though unrestricted optional editing showed no preference advantage in Labnote 005.
composition-005
completenegative
Stochastic deduplicated optional editor campaign
A modest nonzero sampling temperature can produce enough distinct optional editor outcomes to support a useful conditional blind review, provided exact duplicate answer pairs are removed before operator intake and remain visible as multiplicity telemetry rather than independent samples.
composition-004
completeinconclusive
Changed-output optional editor campaign
Optional editor access may be useful specifically when it chooses to change a direct rewrite, even if most assignments appropriately remain unchanged. A larger campaign can estimate both overall usefulness and preference conditional on a genuine editor-caused change without requiring the owner to review identical texts.
composition-003
completenegative
Optional editor-access campaign
Giving a model optional access to constrained revision after a matched direct rewrite may improve judged output without forcing a mutation. Finalizing the initial candidate unchanged is a first-class treatment outcome.
composition-002
completeinconclusive
Seeded mandatory-revision campaign
Schema-constrained revision can reliably transform a seeded draft through real virtual-buffer mutations and produce outputs suitable for blinded comparison with direct rewriting.
composition-001
completepositive
Prompted and constrained editing baseline
An existing instruct model can produce deterministically applicable virtual-buffer operations reliably enough that fine-tuning is not yet justified.
prosody-007
completeinconclusive
LLM-directed bounded speculation subscriptions
Can a conversational model recognize a predictable semantic trajectory once per completed turn and request a bounded runtime-owned subscription that selects future speculative work more effectively than a stateless numeric readiness gate?
prosody-006
completenegative
Familiarity-gated speculative cognition
Can a runtime use a sub-millisecond readiness policy to select the minority of unfinished conversational turns where private speculative response work has positive net value, while abstaining everywhere else?
prosody-005
completenegative
Speculative semantic branch and turn-state prediction
Can semantic branch prediction during an unfinished human turn safely hide language model latency, while a turn-state estimator reduces premature interruption and supports optional collaborative completion?
prosody-004
completeinconclusive
Context-conditioned prosody IR inference and synthesis
Can a language model infer an intended, backend-independent prosodic structure from discourse context when the literal target utterance is held fixed? Separately, can a Kokoro compiler render that structure so a listener can recover the intended reading?
prosody-003
completeinconclusive
Amy-LM synthetic prosody downstream-response pilot (v0.3.1)
Does adding delivery cues derived from the correct audio improve an LLM's response to flattened conversational text relative to text-only and shuffled-cue controls?
prosody-002
completenegative
Shakespeare monologue downstream-response pilot (v0.3.1)
Does adding delivery/prosody information to a transcript improve or meaningfully change an LLM's conversational response?
prosody-001
completepositive
Selected Poems real-media ingest (v0.1.0)
Can the packaged file and simulated-stream ingestion paths produce consistent prosodic metadata from real spoken audio when transcription remains an external step?