Strojno učenje

std-data-collector

pure buffer-accumulation atom, parameterized for any ML domain. Collects data points into a buffer and fires BUFFER_READY with the full accumulated buffer once it reaches config.bufferSize, then resets for the next cycle. No ML compute, no model call — state machine + array operators only. Use ahead of a training loop, a batch inference call, or any downstream rung that needs a fixed-size window of accumulated observations.

Entiteta: DataBuffer (runtime)

Lastnosti

Nastavitve

std-graph-builder

builds node/edge structures from an entity collection for downstream GNN pipelines; pure array/object transform, no ML compute.

Entiteta: GraphStructure (runtime)

Lastnosti

Nastavitve

std-ml-classify

structured LLM verdict from a closed category set, via the llm.classify service seam. Bottom of the fall-through: always answers or fails, never abstains.

Entiteta: ClassifierVerdict (runtime)

Lastnosti

Nastavitve

std-ml-exact-check

normalizes and compares a candidate value against a key value; abstains on no match.

Entiteta: ExactCheck (runtime)

Lastnosti

Nastavitve

std-ml-infer

the learned rung (R4) of the intelligence ladder, and the seam between a .lolo circuit and a trained model deployed on masar (or off-the-shelf HuggingFace weights) reached through the ml service. Sends the caller-supplied input straight to call-service ml infer, then guards the raw output before it ever becomes an effect: validates it against the declared outputContract, clamps it into range, and only then decides. Emits INFERRED when the output satisfies the contract and confidence clears confidenceFloor — carrying the clamped output, confidence, and an empty violations list. Otherwise abstains with INFER_ABSTAINED, whose reason is one of: low confidence (contract satisfied but below floor, violations empty), contract violation (violations populated via contract/violations for diagnosis), or unavailable (the service call itself failed — a masar cold start or a down endpoint, never surfaced as a user-visible error). Use when a numeric or tensor-shaped model answer needs a licensed guard before becoming a decision — classification scores, regressions, embedding-derived predictions — and the next rung (R5 generative) should take over on any abstain.

Entiteta: Inference (runtime)

Lastnosti

Nastavitve

std-ml-label-capture

records a labeled example (input, verdict, source rung, observed outcome) so that expensive top-rung answers become training data for a cheaper rung later. Generic: it knows nothing about grading or students. sampleRate gates what fraction of examples are kept (1 = capture all); sink tags which named destination the example is attributed to.

Entiteta: LabeledExample (persistent)

Lastnosti

Nastavitve

std-ml-lookup

finds the candidate value in a list of accepted values; abstains when nothing matches.

Entiteta: KeyLookup (runtime)

Lastnosti

Nastavitve

std-ml-posterior

the statistical rung (R2) of the intelligence ladder. Holds a Beta(alpha, beta) belief over an unknown success rate and updates it from boolean observations (Bayesian updating, knowledge tracing, Thompson-sampling prior). After each observation it draws samples from the posterior, computes a credible interval, and decides: MASTERY_REACHED when the mean clears the floor with a tight interval, REMEDIATION_NEEDED when the mean sits at or below the ceiling with the same confidence, or EVIDENCE_INSUFFICIENT — the abstain — when the interval is still too wide to trust. Use when a hand-tuned score needs to become a belief that can say 'not enough evidence yet' (mastery, skill, risk, preference, difficulty).

Entiteta: Posterior (runtime)

Lastnosti

Nastavitve

std-ml-similarity

embed a candidate against references and gate on cosine floor + margin; abstains when not confident.

Entiteta: SimilarityComparison (runtime)

Lastnosti

Nastavitve

std-tokenizer

a text-preprocessing atom for tokenization. Splits raw text into a token sequence using a configurable method (whitespace, character, or arbitrary delimiter), normalizes and truncates the sequence, then maps each token to a vocabulary id (falling back to unkId when the token is absent from config.vocab). Feeds downstream embedding or sequence-model atoms. Pure circuit logic: str/* and array/* operators only, no ML compute, no model call.

Entiteta: TokenSequence (runtime)

Lastnosti

Nastavitve

std-weight-validator

a constraint-checking atom for model weights, applying the same discipline as the zero-warning validator to weights: constrained and checked, not trusted (Almadar_Intelligence_Ladder.md §7.5). Checks a candidate weight tensor against three pure predicate constraints — magnitude bounds (every element's absolute value must clear maxMagnitude), forbidden regions (excluded [min,max] bands no weight may land in even within the magnitude bound), and a regression threshold against a baseline tensor (per-element drift must not exceed regressionThreshold; 0 disables the check) — and reports every violated check with its detail. Emits WEIGHTS_ACCEPTED when all three checks clear, or WEIGHTS_REJECTED carrying the full violations list otherwise. Pure circuit logic: predicates and comparisons only, no ML compute, no model call.

Entiteta: WeightValidation (runtime)

Lastnosti

Nastavitve

std-escalating-decision

the intelligence ladder itself, as one reusable composition. One request enters at DECIDE and falls through the rungs in order — R0 exact check, R1 table lookup, R3 retrieval, R4 learned inference, R5 generative classification — each rung handing off to the next by abstaining, until one proves the answer. Exits DECIDED, tagged with which rung answered (rung, confidence, matched, reasoning), or DECISION_FAILED if even the generative floor could not serve it. The escalation is a chain, not a hub: each rung listens directly to the rung above it and projects its own inputs out of the carried request, so the trait graph is a DAG and no event ever returns to a rung that already ran. Domain-agnostic: it knows about candidates and categories, not students or quizzes — grading, routing, matching, moderation, and duplicate detection are all this same molecule with different knobs. R2 (the Beta posterior) is deliberately not in this chain: it is the 'decide what to do next' half of a turn and already lives in std-knowledge-tracing.

Entiteta: LadderRequest (runtime)

Lastnosti

Nastavitve

std-knowledge-tracing

turns one true/false observation ('did the learner get this right?') into a tracked belief about mastery, and captures the same observation as a labeled training example in the same motion. Composes std-ml-posterior (the R2 statistical rung — a Beta-Bernoulli belief over an unknown success rate) with std-ml-label-capture: every posterior verdict — MASTERY_REACHED, REMEDIATION_NEEDED, or the abstain EVIDENCE_INSUFFICIENT — becomes a persisted labeled example (the input skill, the posterior's own verdict, and the ground-truth outcome that produced it), so a cheaper rung can later be trained to reproduce the same call without running the statistics. Domain-agnostic: it knows about skills and observations, not students, quizzes, or courses. One instance tracks one belief (e.g. one learner × one skill); the host composes as many instances as it has beliefs to hold.

Entiteta: KnowledgeState (runtime)

Lastnosti

Nastavitve