Parameter Glossary & Canonical Registry — SOTA++ 2026¶

This is the collection's data dictionary. It exists to prevent semantic drift: the same name referring to different mechanisms, symbols reused without context, UI knobs mistaken for fundamental parameters, and backend aliases treated as new concepts.

Main rule:

\[ \boxed{ Name \neq Meaning } \]

Meaning is defined by the triple:

\[ \boxed{(Domain,Layer,Definition)} \]

Example: temperature has multiple incompatible meanings.

See CONVENTIONS for editorial and notation rules.

1. Canonical provenance taxonomy¶

Tag	Layer	Question
`ARCH`	architecture	how is the network built?
`MODEL`	checkpoint/config	what was fixed in the model?
`OBJ`	objective function	what does training optimize?
`DATA`	data	what distribution was used?
`TRAIN`	optimization	how were parameters updated?
`ALIGN`	post-training/alignment	how do preferences/rewards alter the policy?
`PEFT`	efficient adaptation	what parameter subspace is trained?
`DISTILL`	distillation	how is knowledge transferred?
`MERGE`	model merging	how are checkpoints/deltas combined?
`GEN`	generative process	AR, diffusion, flow, masked refinement, etc.
`COND`	conditioning	what input or representation conditions the prediction?
`INF`	inference	what operator/control runs at use time?
`STATE`	persistent state	what grows or persists?
`BACKEND`	runtime	how is computation implemented?
`PIPE`	pipeline	is it external to the model?
`INDEX`	retrieval index	how is the corpus indexed/searched?
`EVAL`	evaluation	how is performance measured?
`SEC`	security	what boundary/control reduces risk?
`UI`	frontend/vendor	friendly name without a universal definition?
`HEURISTIC`	empirical rule	guideline without a universal law?

2. Dimensional conventions¶

Symbol	Default meaning
\(B\)	batch size / number of sequences
\(T\)	temporal or sequence length; context must disambiguate
\(L\)	number of layers; loss uses \(\mathcal L\)
\(d\)	generic hidden/embedding dimension
\(d_h\)	dimension per attention head
\(H\)	image height or number of heads; avoid unsubscripted use in ambiguous formulas
\(W\)	image width or weight matrix; use context/subscripts
\(C\)	channels/conditioning depending on context
\(V\)	vocabulary size or value tensor; prefer \(\vert\mathcal V\vert\) for vocabulary when ambiguous
\(N\)	generic number of items/tokens/samples
\(K\)	top-K / retrieval K / clusters; always qualify
\(r\)	LoRA/low-rank rank
\(\theta\)	learned model parameters
\(\eta\)	learning rate in training context
\(\epsilon\)	noise/residual/tolerance; always qualify
\(\tau\)	contrastive/calibration temperature when explicitly stated
\(s\)	CFG/guidance scale in diffusion context

3. Critical semantic collisions¶

3.1 Temperature¶

Canonical name	Formula/use	Do not confuse with
Sampling Temperature	\(softmax(z/T)\)	contrastive temperature
Contrastive Temperature	\(\exp(sim/\tau)\)	LLM sampling
Distillation Temperature	softened targets via \(T\)	sampling
Entropy/Policy Temperature	regularization/exploration in RL	sampling
Calibration Temperature	post-hoc temperature scaling	generation randomness

Never write only “temperature” in a cross-domain document.

3.2 Rank¶

Rank	Meaning
LoRA rank \(r\)	dimension of the low-rank update
Retrieval rank	document position in a ranking
Matrix rank	algebraic dimension of row/column space
Rank correlation	statistical ordering, e.g. Spearman/Kendall

3.3 Steps¶

Term	Meaning
Diffusion/flow steps	integration updates along the solver trajectory; distinct from model-evaluation count
Training steps	optimizer updates
Agent steps	observe/act iterations
Reasoning/search steps	inference/search budget
Scheduler steps	temporal/noise-schedule discretization

3.4 Context / window¶

Term	Meaning
LLM context window	addressable tokens
Training sequence length	tokens processed per sample
Sliding attention window	local subset visible to each token
Video temporal window	frames/latents processed together
Retrieval context	chunks injected into the prompt
Agent context	conversation/state/tool trace

3.5 Alpha / Beta¶

These are generic symbols, not concepts.

They may represent:

diffusion coefficients;
LoRA scaling;
optimizer betas;
DPO strength;
EMA coefficients;
statistical parameters.

Always use a qualified name.

4. LLM — architecture registry¶

Primary file: 02_LLM_TRANSFORMER_INFERENCE_DATASHEET.

Parameter	Tag	Canonical definition	Primary effect
`n_layers`	`ARCH`	number of Transformer blocks	capacity/compute/memory
`d_model` / hidden size	`ARCH`	residual-stream width	params/compute
`n_heads`	`ARCH`	query attention heads	attention partition
`n_kv_heads`	`ARCH`	K/V heads	KV memory
`head_dim`	`ARCH`	dimension per head	QK geometry
`ffn_dim` / intermediate size	`ARCH`	internal FFN dimension	params/compute
`vocab_size`	`MODEL`	vocabulary size	output matrix/tokenization
`max_position_embeddings`	`MODEL`	natively configured position limit	addressable sequence regime
`rope_theta`	`MODEL`	RoPE base/frequency parameter	positional phase
`sliding_window`	`ARCH/MODEL`	bounded local attention	compute/context locality
`num_experts`	`ARCH`	total MoE experts	storage/capacity
`experts_per_token`	`ARCH`	active top-k experts	active compute
`router_aux_loss`	`TRAIN/OBJ`	expert balancing term	routing balance

5. LLM — attention / state registry¶

Parameter	Tag	Definition
Q	`STATE`	query vectors
K	`STATE`	key vectors
V	`STATE`	value vectors
KV cache	`STATE/BACKEND`	K/V persisted for autoregressive decode
KV dtype	`BACKEND`	cache precision/quantization
KV offload	`BACKEND`	CPU/GPU placement of the cache
MHA	`ARCH`	one K/V head per Q head
GQA	`ARCH`	groups of Q heads share K/V
MQA	`ARCH`	one shared K/V set
MLA	`ARCH`	compressed latent attention state
attention mask	`ARCH/PIPE`	visibility matrix/rule
causal mask	`ARCH`	prevents looking into the future in AR

6. LLM — inference/sampling registry¶

Parameter	Tag	Definition	Conceptual unit/range
temperature	`INF`	divides logits before softmax	\(T>0\)
top-k	`INF`	keeps K highest-scoring tokens	integer
top-p	`INF`	smallest set with cumulative mass ≥ p	\(0<p\le1\)
min-p	`INF`	cutoff relative to the most probable token	\(0\le p\le1\)
typical-p	`INF`	filters by typicality/self-information	\(0<p\le1\)
top-n-sigma	`INF`	threshold based on the logit distribution	backend-specific
XTC	`INF`	exclusion/candidate-pruning sampler	backend-specific
adaptive-p	`INF`	reweights probabilities around an adaptive target, using an EMA of the original probabilities of selected tokens	backend-specific; llama.cpp implementation, record version
repetition penalty	`INF`	modifies logits of repeated tokens	ratio/factor
frequency penalty	`INF`	penalty proportional to frequency	scale
presence penalty	`INF`	penalty for binary presence	scale
DRY	`INF`	penalizes repeated sequences	multi-param
Mirostat	`INF`	feedback controller for surprise/perplexity	tau/eta/state
seed	`INF/BACKEND`	PRNG initialization	integer/state
max new tokens	`INF`	output budget	tokens
stop tokens	`INF/PIPE`	stop on token IDs	set
stop strings	`PIPE`	textual stop condition	strings
grammar	`INF`	allowed formal language	grammar/schema
reasoning budget	`INF/PIPE`	compute/token budget for reasoning	tokens/effort

7. Runtime registry¶

File: 09_LOCAL_AI_INFERENCE_RUNTIME_DATASHEET.

Parameter	Tag	Definition
bpw	`BACKEND`	effective bits per weight
weight dtype	`BACKEND`	weight representation
compute dtype	`BACKEND`	multiplication/accumulation dtype
activation dtype	`BACKEND`	activation representation
GPU layers/offload	`BACKEND`	fraction/layers placed on GPU
batch size	`BACKEND`	sequences/tokens processed as a batch
microbatch	`BACKEND/TRAIN`	smaller physical batch unit
flash attention	`BACKEND`	memory-efficient attention-kernel family
mmap	`BACKEND`	memory-mapped weight loading
pinned memory	`BACKEND`	host pages pinned for faster transfers
TTFT	`EVAL`	time to first token
TPOT	`EVAL`	time per output token
prompt TPS	`EVAL`	prefill token throughput
decode TPS	`EVAL`	generation token throughput
peak VRAM	`EVAL`	maximum device memory used
peak RAM	`EVAL`	maximum host memory used

8. Quantization registry¶

Term	Definition
weight quantization	quantizes persistent weights
KV quantization	quantizes K/V cache
activation quantization	quantizes activations
group size	number of values sharing quantization parameters
scale	dequantization factor
zero point	offset in asymmetric quantization
symmetric quantization	levels centered around zero
asymmetric quantization	uses zero point / shifted interval
mixed precision	tensors/layers at different precisions
calibration set	data used to estimate quantization parameters
quantization error	difference between original and quantized values

Q4 alone is an insufficient description.

9. Image diffusion/flow registry¶

File: 04_IMAGE_GENERATION_DIFFUSION_FLOW_DATASHEET.

Parameter	Tag	Definition
latent width/height	`STATE`	latent spatial resolution
VAE scale factor	`MODEL`	pixel ↔ latent-resolution ratio
latent channels	`ARCH/MODEL`	channels in latent space
timestep \(t\)	`GEN`	trajectory coordinate
sigma \(\sigma_t\)	`GEN`	noise level
alpha \(\alpha_t\)	`GEN`	signal coefficient
integration steps	`INF`	number of solver updates along the integration grid
NFE_model	`INF`	model evaluations per generated sample, including separate guidance branches
scheduler	`INF`	locations \(t_i/\sigma_i\) on the integration grid
solver/sampler	`INF`	numerical stepping rule between points
CFG scale	`INF`	conditional/unconditional extrapolation
guidance	`INF/MODEL`	strength/form of conditioning; family-specific definition
denoise strength	`PIPE/INF`	start point / amount of added noise in img2img
flow shift	`INF/MODEL`	schedule transformation in flow models
seed	`INF`	PRNG initial state
negative prompt	`PIPE/INF`	alternative conditioning in compatible pipelines
clip skip	`MODEL/PIPE`	text-encoder layer selection; legacy/model-specific
Hires Fix	`PIPE`	second-pass upscale/refinement

10. Solver versus scheduler registry¶

Concept	Question
Scheduler	where along the trajectory do we evaluate?
Solver	how do we advance between evaluations?
NFE_model	how many model evaluations are performed per generated sample, counting each guidance branch?
Ancestral step	does it inject a stochastic component during the trajectory?
Deterministic ODE step	does it integrate the field without extra noise injection?

Here, NFE means NFE_model: each conditioning/guidance branch counts separately even when branches are concatenated into one batched forward call. Integration steps, guided-field evaluations, and model evaluations can differ. For example, the EDM Heun sampler uses 35 model evaluations for 18 integration steps, without additional guidance branches. This count tracks model work; batching, tensor shapes, and precision affect FLOPs and wall time, which must be reported separately.

Never use sampler/scheduler as generic synonyms.

11. Video registry¶

File: 05_VIDEO_GENERATION_SPATIOTEMPORAL_DATASHEET.

Term	Tag	Definition
frames \(T\)	`STATE`	observable temporal extent
latent frames	`STATE`	temporal extent after compression
temporal compression ratio	`MODEL`	frames ↔ latent frames
spatial compression	`MODEL`	pixels ↔ latent spatial grid
fps	`PIPE/MODEL`	temporal sample/display rate
frame stride	`PIPE`	spacing between conditioning frames
temporal window	`ARCH/PIPE`	window processed jointly
overlap	`PIPE`	frames shared across chunks
motion bucket/score	`MODEL/UI`	conditioning specific to some model families
STG	`INF`	spatio-temporal guidance
audio guidance	`INF/MODEL`	audio-specific guidance in joint AV models
continuation length	`PIPE`	video-extension horizon

12. Audio/speech registry¶

File: 06_AUDIO_MUSIC_TTS_GENERATION_DATASHEET.

Term	Tag	Definition
sample rate	`MODEL/PIPE`	samples per second
hop length	`MODEL`	STFT/mel-frame stride
FFT size	`MODEL/PIPE`	number of FFT points (`n_fft`); may exceed window length through zero padding
window length	`MODEL/PIPE`	analysis-window length in samples (`win_length`), distinct from FFT size
mel bins	`MODEL`	mel-spectrogram dimensionality
codec frame rate	`MODEL`	latent/codec frames per second
codebooks	`ARCH/MODEL`	code-vector dictionaries; in RVQ, successive stages quantize the residual left by earlier stages
bitrate	`MODEL/PIPE`	represented bits per second
duration	`INF/PIPE`	generation horizon
speaker embedding	`STATE/COND`	conditioning representation of speaker identity
similarity	`UI/EVAL`	vendor knob or metric; qualify explicitly
stability	`UI`	vendor abstraction with no universal definition
VAD threshold	`PIPE`	speech/non-speech decision threshold
endpointing	`PIPE`	rule for detecting end of turn
chunk size	`PIPE/BACKEND`	streaming window
lookahead	`ARCH/PIPE`	permitted future context
WER	`EVAL`	word error rate
RTF	`EVAL`	real-time factor

13. ASR-specific registry¶

Term	Definition
blank token	CTC symbol with no emitted output
CTC collapse	removes repeats/blanks to form a transcript
acoustic encoder	maps audio to states
prediction network	label history in a transducer
joint network	combines acoustic and prediction state
beam size	number of hypotheses retained during decode
timestamp token	output tied to time
diarization	who spoke when
forced alignment	aligns known text to audio

14. 3D registry¶

File: 07_3D_GENERATION_REPRESENTATION_DATASHEET.

Term	Definition
vertices \(V\)	mesh positions
faces \(F\)	mesh connectivity
SDF	signed distance field
density field	volumetric/NeRF-like density
Gaussian mean \(\mu\)	splat position
covariance \(\Sigma\)	Gaussian shape/orientation
opacity \(\alpha\)	volumetric contribution
SH/features	directional appearance/color
isosurface threshold	level set used to extract a surface
marching cubes resolution	grid used during extraction
poly count	faces after generation/decimation
watertightness	closed-surface property
manifoldness	local topological validity
UV resolution	texture resolution
PBR maps	material channels such as albedo/roughness/metallic

15. Multimodal/VLM registry¶

File: 10_MULTIMODAL_VLM_OMNI_DATASHEET.

Term	Tag	Definition
patch size	`ARCH/MODEL`	pixels per visual patch
vision tokens	`STATE`	tokens/states derived from an image
projector	`ARCH`	maps encoder dimension → language dimension
resampler	`ARCH`	compresses/adapts token count
Q-Former	`ARCH`	query-based bridge between modalities
early fusion	`ARCH`	modalities combined early
joint fusion	`ARCH`	shared multimodal processing
late fusion	`ARCH/PIPE`	representations/decisions combined late
multimodal RoPE	`ARCH`	position encoding across space/time/modality
visual token budget	`STATE/BACKEND`	vision contribution to context
frame sampling	`PIPE`	temporal subset sent to the model
grounding	`OBJ/EVAL`	bind output to an actual region/time span

16. Retrieval/embedding registry¶

File: 11_EMBEDDING_RETRIEVAL_RERANKER_DATASHEET.

Term	Tag	Definition
embedding dimension \(d\)	`MODEL`	vector dimensionality
pooling	`MODEL/PIPE`	sequence → vector
L2 normalization	`PIPE/MODEL`	\(z/\Vert z\Vert\) for nonzero z; numerical implementations must specify an epsilon or zero-vector policy
cosine similarity	`INF`	normalized angular similarity
dot product	`INF`	inner product
contrastive temperature \(\tau\)	`TRAIN/OBJ`	sharpness of the contrastive loss
chunk size	`PIPE`	size of an indexed fragment
chunk overlap	`PIPE`	duplicated content across chunks
retrieval K	`INF`	number of retrieved candidates
rerank K	`PIPE/INF`	candidates passed to the reranker
HNSW M	`INDEX`	connectivity of the HNSW graph
efConstruction	`INDEX`	search width during index construction
efSearch	`INDEX/INF`	search width at query time
IVF nlist	`INDEX`	number of coarse clusters
IVF nprobe	`INDEX/INF`	number of clusters searched
PQ subquantizers	`INDEX`	vector partition for product quantization
Recall@K	`EVAL`	fraction of relevant items retrieved
NDCG@K	`EVAL`	discounted ranking quality

17. Reward/verifier registry¶

File: 12_REWARD_VERIFIER_JUDGE_DATASHEET.

Term	Tag	Definition
reward \(r\)	`OBJ/EVAL`	scalar utility proxy
chosen/rejected	`DATA/ALIGN`	preference pair
ORM	`MODEL`	outcome reward model
PRM	`MODEL`	process reward model
generative verifier	`MODEL/PIPE`	reasoning/critique → verdict
judge	`PIPE/EVAL`	evaluator model/human
rubric	`EVAL`	explicit evaluation criteria
calibration	`EVAL`	match between confidence and empirical correctness
position bias	`EVAL`	preference dependent on ordering
verbosity bias	`EVAL`	preference for longer responses
best-of-N	`INF/PIPE`	sample N + select

18. Agent registry¶

File: 08_AGENT_ACTION_CONTROL_SYSTEM_DATASHEET.

Term	Tag	Definition
tool schema	`PIPE`	formal action interface
tool choice	`INF/PIPE`	action/tool selection
strict schema	`INF/PIPE`	constrained syntax
retry budget	`PIPE`	maximum re-executions
step budget	`PIPE`	maximum agent-loop iterations
tool budget	`PIPE/SEC`	maximum tool calls/cost
state	`STATE`	agent working memory
HITL	`SEC/PIPE`	human approval before an action
idempotency	`SEC/PIPE`	repeating an action does not multiply its effect
rollback	`SEC/PIPE`	revert a change
policy hook	`SEC`	validates an action at runtime
capability	`SEC/PIPE`	operation granted to the agent

19. World model / VLA registry¶

File: 13_WORLD_MODEL_VLA_EMBODIED_AI_DATASHEET.

Term	Definition
state \(s_t\)	representation of the environment
action \(a_t\)	applied command
transition model	\(p(s_{t+1}\vert s_t,a_t)\)
observation \(o_t\)	partial measurement of the state
belief state	distribution/latent over hidden state
rollout horizon \(H\)	imagined steps
action chunk	block-predicted action sequence
diffusion policy	trajectory denoising in action space
flow policy	vector field over an action trajectory
MPC horizon	model-predictive-control optimization horizon
success rate	task completion rate
collision rate	physical-safety events

20. Time-series registry¶

File: 14_TIME_SERIES_TABULAR_FOUNDATION_MODELS.

Term	Definition
context length	observed history
horizon	predicted future
patch length	points grouped into each token/patch
stride	shift between patches
frequency	sample interval/rate
covariates	auxiliary variables
quantile \(\tau\)	level of the predicted distribution
prediction interval	probabilistic range
MAE	mean absolute error
RMSE	root mean squared error
MAPE	relative percentage metric with a zero-value caveat
CRPS	scoring rule for a cumulative predictive distribution

21. Tabular registry¶

Term	Definition
feature tokenization	column/value → representation
categorical embedding	embedding of a category
missingness indicator	explicit representation of absence
row attention	relationships across records
column attention	relationships across attributes
in-context training set	tabular examples supplied as context
PFN	Prior-Data Fitted Network family
AUROC	binary ranking metric
log loss	probabilistic classification loss

22. Graph registry¶

File: 15_GRAPH_FOUNDATION_MODELS_DATASHEET.

Term	Definition
node	element \(v\in V\)
edge	relation \((u,v)\in E\)
degree	number of connections
adjacency	connectivity structure
message passing	neighbor → node aggregation
neighborhood \(\mathcal N(v)\)	neighboring nodes
aggregation \(\oplus\)	sum/mean/max/attention, etc.
graph depth	message-passing layers
oversmoothing	node states become too similar
oversquashing	distant information compressed through a bottleneck
Laplacian PE	spectral positional encoding
random-walk PE	position/structure through walk statistics
link prediction	infer edges
graph generation	generate topology/attributes

23. Training registry¶

File: 16_TRAINING_ALIGNMENT_ADAPTATION_DATASHEET.

Parameter	Tag	Definition
learning rate \(\eta\)	`TRAIN`	base magnitude of optimizer updates
batch size	`TRAIN`	samples per logical update
microbatch	`TRAIN`	samples per physical forward/backward pass
gradient accumulation	`TRAIN`	microsteps accumulated before optimizer step
warmup	`TRAIN`	increasing-LR phase
weight decay	`TRAIN`	weight regularization
gradient clipping	`TRAIN`	bounds gradient norm/magnitude
optimizer betas	`TRAIN`	EMA coefficients in Adam-like optimizers
sequence length	`TRAIN`	tokens per sample/window
packing	`DATA/TRAIN`	combining samples into a window
epochs	`TRAIN`	passes over the dataset
steps	`TRAIN`	optimizer updates
tokens seen	`TRAIN/EVAL`	actual processed token volume

24. SFT / alignment registry¶

Term	Definition
SFT	supervised fine-tuning
completion-only loss	prompt tokens masked out of the loss
preference pair	\((x,y_w,y_l)\)
reward model	preferences → scalar
DPO beta	strength/log-ratio scaling in DPO
KL coefficient	limits policy drift
GRPO group size	samples per prompt for relative advantages
reward std	signal diversity within the group
rollout length	generated response/action horizon
clipping epsilon	probability-ratio clipping band in the PPO surrogate objective; does not guarantee a hard bound on parameter changes or KL
GSPO sequence ratio	likelihood ratio normalized by sequence length
RLVR	reinforcement learning with verifiable reward

25. PEFT registry¶

Term	Definition
LoRA rank \(r\)	low-rank dimension
LoRA alpha	scale of the adapter update
target modules	matrices receiving the adapter
LoRA dropout	dropout on the adapter path
QLoRA	quantized frozen base + trainable LoRA
DoRA	magnitude/direction decomposed adaptation
PiSSA/EVA/LoftQ init	initialization/adaptation strategies
adapter merge	folds the delta into the base weight
modules_to_save	additional modules trained/saved

26. Distillation registry¶

Term	Definition
teacher	model supplying the target/signal
student	model trained to imitate it
distillation temperature	smoothing of teacher/student distributions
hard targets	discrete labels/answers
soft targets	logits/probabilities
sequence distillation	teacher outputs become training text
reasoning distillation	teacher trajectories/rationales become supervision
on-policy distillation	student generates states/trajectories evaluated by teacher
dataset distillation	compresses information into a compact synthetic dataset

27. Model merging registry¶

Term	Definition
linear merge	weighted average of parameters
task vector	\(\tau=\theta_{ft}-\theta_{base}\)
task arithmetic	weighted sum of task vectors
SLERP	spherical linear interpolation
TIES	trim + sign resolve + merge
DARE	drop-and-rescale task deltas
density	fraction of deltas retained
merge weight	contribution of each source model
base model	common reference for task vectors

28. Evaluation registry¶

File: 17_EVALUATION_BENCHMARK_EXPERIMENTATION_DATASHEET.

Metric	Definition
accuracy	correct / total
exact match	equality after normalization
precision	TP/(TP+FP)
recall	TP/(TP+FN)
F1	harmonic mean of precision/recall
pass@k	probability of at least one correct solution among k
perplexity	exponential mean NLL
Brier score	squared probabilistic error
ECE	calibration-bin gap
WER	(S+D+I)/N
MRR	mean reciprocal rank
NDCG	normalized discounted cumulative gain
MAE	mean absolute error
RMSE	root mean squared error
TTFT	time to first token
TPOT	time per output token
TPS	tokens per second
p95/p99	latency tail percentile
CI	confidence interval
effect size	practical magnitude of a difference

29. Statistical registry¶

Term	Definition
sample mean \(\bar x\)	estimate of the mean
sample variance \(s^2\)	dispersion estimate
standard error	standard deviation of a statistic's sampling distribution
estimated standard error of the mean	\(s/\sqrt N\) for N iid observations with finite variance; s is the sample standard deviation (NIST); dependent observations require a design-appropriate variance estimator
confidence interval	interval estimate
bootstrap	resampling with replacement
paired comparison	same unit evaluated by A and B
permutation test	null distribution from random reassignment
McNemar	paired test for binary outcomes
power	probability of detecting a real effect at a given size
saturation	benchmark loses discrimination between systems
contamination	evaluation evidence appears in training/development

30. Security registry¶

File: 18_SECURITY_ROBUSTNESS_FAILURE_MODES_DATASHEET.

Term	Tag	Definition
prompt injection	`SEC`	content alters intended instruction/control flow
indirect injection	`SEC`	hostile instruction comes from external content
jailbreak	`SEC`	bypass of behavioral policy
data poisoning	`SEC/DATA`	deliberately contaminated training data
backdoor	`SEC/MODEL`	trigger → hidden malicious behavior
RAG poisoning	`SEC`	manipulated corpus/index
excessive agency	`SEC`	functionality/permissions/autonomy beyond what is needed
goal hijack	`SEC`	operational objective is diverted
tool misuse	`SEC`	valid tool used dangerously
memory poisoning	`SEC`	malicious persistent state
supply-chain attack	`SEC`	compromised artifact/dependency/provider
HITL	`SEC/PIPE`	human approval gate
sandbox	`SEC/BACKEND`	execution isolation
least privilege	`SEC`	minimum permission scope
circuit breaker	`SEC/PIPE`	automatic stop on a condition
ASR	`EVAL/SEC`	attack success rate — do not confuse with Automatic Speech Recognition

The ASR collision is particularly dangerous in this collection. In audio it means Automatic Speech Recognition; in security it means Attack Success Rate. Always spell it out on first use.

31. UI knob registry — treat with healthy suspicion¶

Knobs such as:

Creativity;
Stability;
Quality;
Motion;
Stylization;
Guidance;
Strength;
Fidelity;

may map to different mechanisms in different products.

Rule:

\[ UIKnob \xrightarrow{document} UnderlyingParameter(s) \]

Without that mapping, do not promote the knob into a law in a datasheet.

32. Backend alias examples¶

Context¶

Possible names:

n_ctx;
context_length;
max_context;
num_ctx.

Canonical concept:

runtime context/token capacity.

Repetition¶

repeat_penalty;
repetition_penalty.

GPU offload¶

n_gpu_layers;
gpu_layers;
percent offload.

The registry should map aliases to one concept rather than proliferating new entries.

33. Unit registry¶

Memory¶

Prefer:

bytes;
MiB/GiB for binary memory.

\[ 1GiB=2^{30}bytes \]

Use GB only when the source/vendor uses decimal units, and make that explicit.

Bandwidth¶

Use GB/s or GiB/s according to the source; do not mix them without conversion.

Compute¶

FLOP/s and tokens/s measure different things.

Time¶

Duration:

ms;
s.

Rates (counts per unit time):

frames/s;
samples/s;
tokens/s.

34. Canonical naming policy¶

Prefer:

Sampling Temperature
Contrastive Temperature
Distillation Temperature

over three entries all named “Temperature.”

Prefer:

Diffusion Steps
Training Steps
Agent Steps

over “Steps.”

35. Parameter card template¶

Every important new variable should be registrable as:

### Canonical Name

- Symbol:
- Domain:
- Provenance tag:
- Type/unit:
- Valid mathematical range:
- Typical implementation range:
- Governing equation:
- Direct effect:
- Coupled variables:
- Physical cost:
- Failure surface:
- Backend aliases:
- Do not confuse with:
- Source / snapshot:

36. Law versus implementation range¶

Distinguish:

Mathematical domain¶

Example:

\[ T>0 \]

for sampling temperature.

Backend accepted range¶

May be:

0..2

in a UI.

Empirical operating range¶

May be:

0.7..1.2

for a specific model.

These are three different things.

37. Native/model parameter versus user control¶

Example:

n_kv_heads is ARCH.

You normally do not change it at inference time.

KV dtype is BACKEND.

You can change it at runtime.

temperature is INF.

You can change it per request.

This classification prevents looking for a nonexistent slider for an architectural property.

38. Causality strength¶

We can classify levers as:

Direct control¶

Explicit mathematical transformation:

temperature;
top-p;
CFG;
solver;
LR.

Indirect conditioning¶

prompt;
reference image;
retrieval context.

Frozen state¶

weights;
tokenizer;
architecture.

Environmental/system control¶

tool permissions;
sandbox;
index.

This classification should accompany causal interpretations.

39. Parameter-coupling notation¶

When useful, record the local sign:

\[ \frac{\partial Y}{\partial x}>0 \]

but only if the relationship is monotonic in the regime being discussed.

Otherwise:

\[ Y(x)\text{ is non-monotonic} \]

is more accurate than drawing a misleading arrow.

40. Snapshot field¶

Implementation-dependent entries should include:

snapshot: YYYY-MM-DD
backend/model: ...

Permanent mathematical equations do not need “2026 version” in their meaning.

41. Deprecated/legacy registry¶

Never silently delete an important old term.

Mark it:

status: legacy
superseded_by: ...

Potential examples:

SVD-specific knobs;
Clip Skip outside families where it still makes sense;
removed UI parameters.

42. Semantic drift detector¶

Before adding a new term, ask:

does a canonical concept already exist?
is it a backend alias?
is it a UI abstraction?
is it family-specific?
does the equation change, or only the name?

If the definition does not change:

\[ NewName\rightarrow Alias \]

not a new entity.

43. Registry maintenance rule¶

Add an entry when at least one condition holds:

the parameter appears in more than one datasheet;
the name has a semantic collision;
the knob is frequently misunderstood;
physical cost depends on it;
it is required to reproduce an experiment;
it is an important architectural axis.

Do not register every argument from every CLI.

44. Document map¶

Domain	Document
ontology	01_GUT_GENERATIVE_AI_UNIFIED_ENGINEERING
LLM	02_LLM_TRANSFORMER_INFERENCE_DATASHEET
code	03_CODE_GENERATION_AND_AGENTIC_CODING_DATASHEET
image	04_IMAGE_GENERATION_DIFFUSION_FLOW_DATASHEET
video	05_VIDEO_GENERATION_SPATIOTEMPORAL_DATASHEET
audio/speech	06_AUDIO_MUSIC_TTS_GENERATION_DATASHEET
3D	07_3D_GENERATION_REPRESENTATION_DATASHEET
agents	08_AGENT_ACTION_CONTROL_SYSTEM_DATASHEET
runtime	09_LOCAL_AI_INFERENCE_RUNTIME_DATASHEET
multimodal	10_MULTIMODAL_VLM_OMNI_DATASHEET
retrieval	11_EMBEDDING_RETRIEVAL_RERANKER_DATASHEET
reward/judge	12_REWARD_VERIFIER_JUDGE_DATASHEET
world/VLA	13_WORLD_MODEL_VLA_EMBODIED_AI_DATASHEET
time series/tabular	14_TIME_SERIES_TABULAR_FOUNDATION_MODELS
graph	15_GRAPH_FOUNDATION_MODELS_DATASHEET
training	16_TRAINING_ALIGNMENT_ADAPTATION_DATASHEET
evaluation	17_EVALUATION_BENCHMARK_EXPERIMENTATION_DATASHEET
security	18_SECURITY_ROBUSTNESS_FAILURE_MODES_DATASHEET

45. Final rule¶

When someone says:

“increase X”

first ask yourself:

\[ \boxed{ Which\ X? } \]

Then:

in what domain?
at what layer?
which equation?
which backend/model?
what physical cost?
which observable changes?
what failure surface?

Only then touch the slider.

Implementation references¶

This registry is derived from the collection's datasheets. For versioned implementation definitions, consult the snapshot references in the relevant document.

Particularly useful references for aliases/current implementation state:

llama.cpp — https://github.com/ggml-org/llama.cpp
Hugging Face Transformers — https://huggingface.co/docs/transformers/
Hugging Face PEFT — https://huggingface.co/docs/peft/main/index
Hugging Face TRL — https://huggingface.co/docs/trl/index
Hugging Face Diffusers — https://huggingface.co/docs/diffusers/
mergekit — https://github.com/arcee-ai/mergekit