CORE_API.md

Core Framework API Reference

This reference tracks current exported behavior of @experiments/core in this repo.

1. Task Lifecycle and Adapters

The framework uses a standardized adapter pattern to manage the execution of diverse experimental tasks.

TaskAdapter (Interface)

All task adapters must implement this interface to be compatible with the unified shell and core lifecycle.

• readonly manifest: TaskManifest: Metadata about the task (ID, label, available variants).
• initialize(context: TaskAdapterContext): Promise<void>: (Optional) Called to set up the task, parse configuration, and prepare resources.
• execute(): Promise<unknown>: (Optional) Called to run the main task logic. Should return the task results.
• terminate(): Promise<void>: (Optional) Called after execution (success or failure) to clean up resources like global listeners or timers.

createTaskAdapter(options) (Factory)

Preferred way to define task adapters without per-task wrapper classes.

• manifest: TaskManifest: Task metadata.
• run(context: TaskAdapterContext): Promise<unknown> | unknown: Main task entrypoint.
• initialize?(context): Optional setup hook.
• terminate?(context): Optional cleanup hook.

The factory returns a TaskAdapter compatible with LifecycleManager and keeps task context management centralized in core.

TaskAdapterContext (Interface)

Context provided to task adapters during initialization and execution.

• container: HTMLElement: The root element for the task UI.
• selection: SelectionContext: Metadata about the current task, variant, and participant.
• coreConfig: CoreConfig: The full core framework configuration.
• taskConfig: JSONObject: The task-specific configuration (automatically resolved at participant scope).
• resolver: VariableResolver: A pre-configured resolver for handling block and trial scoped variables.

LifecycleManager (Class)

Orchestrates the execution of a TaskAdapter.

• constructor(adapter: TaskAdapter)
• run(context: TaskAdapterContext): Promise<unknown>: Executes the full lifecycle: initialize -> execute (or legacy launch) -> terminate.
• Note: run() automatically performs high-level variable resolution on context.taskConfig before calling initialize. Only participant scoped variables are resolved at this stage; block and trial scoped variables remain as tokens for the adapter to handle.

runCustomRtTrial(args): Promise<MultiPhaseTrialResult>

A generalized RT trial runner that supports arbitrary phase sequences. Each phase can have a custom render function.

SpatialLayoutManager (Class)

Provides utilities for generating non-overlapping spatial slots for stimuli.

• generateSlots(args): Point[]: Supports "circular", "grid", and "random" templates.

SceneRenderer (Class)

Standardized canvas renderer for SceneStimulus models. Handles rendering of shapes and supports slot-based positioning.

diffScenes(s1, s2): SceneDiff

Utility to identify identity changes between two structured scenes.

2. Selection and configuration

ConfigurationManager (Class)

Manages the loading, merging, and validation of experiment configurations.

• load(path: string): Promise<JSONObject>: Fetches and parses a JSON config file.
• merge(base, taskDefault, variantOverride, runtimeOverride?): JSONObject: Sequentially deep-merges configuration levels.
• resolve(config: JSONObject, resolver: VariableResolver): JSONObject: Recursively resolves variable tokens in the configuration using the provided resolver.

resolveSelection(coreConfig: CoreConfig): SelectionContext

Resolves task/variant/configPath/overrides/participant metadata from URL + JATOS.

When running under JATOS, URL-style parameters are resolved from:

1. window.location.search (if present)
2. jatos.urlQueryParameters (fallback, preserves launch params across Publix redirects)

Task/variant precedence:

1. JATOS (taskId, variantId)
2. URL (task, variant)
3. coreConfig.selection

Config-path precedence:

1. JATOS component input config (aliases configID / configId)
2. URL config
3. variant as a config alias — URL variant, or JATOS variantId / variant / variant_id — resolved task-scoped first (configs/<taskId>/<variant>.json)

Overrides precedence:

1. JATOS overrides
2. URL overrides

Accepted URL keys:

• task, variant, config, overrides, cc
• participant keys: PROLIFIC_PID, STUDY_ID, SESSION_ID, SONA_ID, participant, survey_code
• participantId is the canonical participant identifier after resolution; sonaId is only populated from the raw SONA_ID parameter when present
• auto-responder toggle: auto
• auto-responder jsPsych mode: auto_mode (visual or data-only)

resolveSelectionWithJatosRetry(coreConfig, maxWaitMs?): Promise<SelectionContext>

Retries selection resolution for JATOS launches where selection payloads may arrive slightly after app boot.

• Default max wait: 10000ms
• Stops early when JATOS selection becomes available
• Skips retry when URL already provides task

loadJatosScriptCandidates(candidates?): Promise<{ loaded, loadedFrom, attempts }>

Attempts to load JATOS runtime script from a candidate list (in order), returning:

• loaded: whether a candidate succeeded
• loadedFrom: source URL used when loaded
• attempts: all attempted candidate URLs

waitForJatosReady(timeoutMs?): Promise<void>

Waits for JATOS readiness via jatos.onLoad(...) when available, with polling fallback for componentJsonInput / studySessionData.

resolveRuntimePath(path: string): string

Normalizes relative-vs-absolute launch paths for local/dev and JATOS:

• preserves absolute URLs (https://..., data:..., etc.)
• rewrites leading-slash app paths to component-relative paths under JATOS
• keeps local paths local-friendly outside JATOS

Runtime path tokens supported by templated stimulus/config helpers:

• {runtime.assetsBase}
• {runtime.configsBase}

ensureEegBridgeReady(coreConfig, taskConfig?): Promise<void>

Performs optional EEG bridge preflight:

• resolves effective EEG config from coreConfig.eeg + taskConfig.eeg
• if enabled=true and requireBridge=true, checks {bridgeUrl}/health
• throws on failure (caller can block experiment launch)

resolveEegBridgeConfig(coreConfig, taskConfig?): ResolvedEegBridgeConfig

Resolves effective EEG bridge settings (enabled, bridgeUrl, requireBridge, event forwarding controls).

buildMergedConfig(base, taskDefault, variantOverride, runtimeOverride?)

Deep merge order:

1. base
2. taskDefault
3. variantOverride
4. runtimeOverride

loadJsonFile(path: string): Promise<JSONObject>

Fetches JSON and enforces object-only payload.

3. Experiment orchestration

runPromptScreens(container, screens): Promise<void>

Renders each prompt and waits for continue (button click or space).

runBlockTrialLoop(args): Promise<BlockTrialLoopResult>

Handles block envelopes and trial iteration.

Key behavior:

• Uses getTrials(block) when provided, else block.trials.
• renderBlockStart/renderBlockEnd returning non-null HTML triggers continue gates.
• Cursor policy is configurable:
  • default hides cursor during each trial
  • set hideCursorDuringTrial: false (or function) for mouse-first tasks.

runTrialTimeline(args): Promise<TrialTimelineResult>

Linear timed stage runner with optional timed response capture.

Returns:

• key, rtMs
• totalDurationMs
• stageTimings[]

runEpochSubTaskStandalone(context, spec): Promise<unknown>

Runs any SubTaskHandle (see the concurrent runner) as a standalone orchestrated experiment: each trial is a timed epoch during which scenario events are injected according to a config-declared schedule. The MATB standalone adapters (matb-comms, matb-sysmon, matb-resman) are thin EpochStandaloneTaskSpec objects over this runner.

The spec supplies the irreducibly task-specific pieces:

• createHandle — sub-task handle factory
• parseScheduleEntry(raw) / toScenarioEvents(event) — schedule parsing and scenario-event expansion
• trialRecordFields(result) / trialEndEventFields?(result) — per-trial CSV columns and trial_end event payload
• auxMetadataKey? / collectAuxRecords?(result) — auxiliary per-trial records accumulated into task metadata
• identity/config keys: taskId, mode, runnerName, csvSuffix, buttonIdPrefix, subTaskConfigKey, scheduleKey, stage size, instruction defaults, duration/ITI defaults

parseEpochStandaloneConfig(taskConfig, spec) handles the shared config shape: task.title, instructions, plan.blocks[] (label/phase/trials/trialDurationMs, before/after block screens, per-block schedule + sub-task config overrides), trialDefaults, and interTrialIntervalMs.

4. UI, keyboard, and jsPsych bridge helpers

normalizeKey(key): string

Lowercases and normalizes " " | "spacebar" | "space" -> "space".

captureTimedResponse(args): Promise<TimedResponse>

Captures first valid key in [startMs, endMs] window over totalDurationMs.

waitForContinue(container, html, options?): Promise<void>

Displays HTML screen with continue button and space shortcut.

toJsPsychKey(key): string

Maps canonical keys to jsPsych choices ("space" -> " ").

toJsPsychChoices(keys): string[]

Maps and deduplicates keys for jsPsych plugin choices.

resolveJsPsychContentHost(container): HTMLElement

Returns .jspsych-content host if present, else container.

pushJsPsychContinueScreen(...)

Pushes a jsPsych call-function timeline node that renders a continue screen through core waitForContinue.

shouldHideCursorForPhase(phase: unknown): boolean

Returns true when the given phase label indicates a trial execution phase where the cursor should be hidden (matches fixation, blank, stimulus, response, or feedback). Safe to call inside jsPsych on_trial_start callbacks — never applies during instruction or continue screens.

extractJsPsychTrialResponse(data: Record<string, unknown>): { key: string | null; rtMs: number | null }

Normalizes the response and rt fields from a jsPsych trial data object into canonical form. Returns null for each field when the raw value is absent or non-finite.

TaskEnvironmentGuard (Class)

Encapsulates environment cleanup for tasks that install keyboard/scroll blockers or hide the cursor. Replaces the pattern of storing disposer references as module-level variables.

• installKeyScrollBlocker(allowedKeys: string[]): void — installs a key scroll blocker and registers its disposer.
• installPageScrollLock(): void — locks page scroll and registers its disposer.
• addDisposer(fn: () => void): void — registers an arbitrary cleanup function.
• cleanup(): void — shows the cursor (setCursorHidden(false)) and runs all registered disposers.

Typical usage:

const taskEnvironment = new TaskEnvironmentGuard();
// in onTaskStart:
taskEnvironment.installKeyScrollBlocker(allowedKeys);
// in terminate:
taskEnvironment.cleanup();

resolvePageBackground({ coreConfig?, taskConfig? }): string | null

Resolves outer shell background with precedence:

1. taskConfig.ui.pageBackground
2. coreConfig.ui.pageBackground
3. null (caller may use CSS default)

resolveInstructionPageSlots(instructions, defaults?): InstructionPageSlots

Shared instruction-slot coercion used by task adapters.

Returned shape:

• intro: string[]
• preBlock: string[]
• postBlock: string[]
• end: string[]

Accepted intro aliases (first key present wins):

• pages (preferred)
• introPages
• intro
• screens

Accepted pre/post/end aliases:

• pre: preBlockPages, beforeBlockPages, beforeBlockScreens
• post: postBlockPages, afterBlockPages, afterBlockScreens
• end: endPages, outroPages, end, outro

Behavior note:

• If a chosen key is explicitly present but blank (for example "" or [""]), the slot resolves to [] (intentional clear) and does not fall back to defaults.
• Blank entries inside arrays are ignored.

resolveWithVariables(value, resolver?, context?)

Shared helper that applies variable expansion across nested arrays/objects via resolver.resolveInValue(...) when a resolver is provided.

Useful for config fields that may be arrays/objects containing tokens (for example beforeBlockScreens: ["$between.preScreen"]).

createDrtPresentationBridge(config, adapter): DrtPresentationBridge

Mode-aware callback bridge for local renderer/audio DRT presentation without embedding renderer logic into core.

Returns:

• hasVisualMode
• hasAuditoryMode
• hasBorderMode
• onStimStart(stimulus)
• onStimEnd(stimulus)
• onResponseHandled()
• hideAll()

Typical usage:

• task-local DRT loop calls onStimStart/onStimEnd from engine/controller hooks
• task-local key handler calls onResponseHandled after a handled DRT response
• task cleanup calls hideAll

Task Modules and Extensions

The framework supports modular extensions that can be attached to specific scopes (task, block, or trial).

TaskModule (Interface)

• id: string: Unique identifier for the module.
• start(config, address, context): TaskModuleHandle: Called when a scope starts.

TaskModuleHandle (Interface)

• stop(): TResult: Called when the scope ends.
• step?(now: number): (Optional) Animation frame tick.
• handleKey?(key, now): (Optional) Keyboard event handler.
• getData?(): TResult: (Optional) Snapshot current active module data.
• controller?: unknown: (Optional) Runtime controller exposed for advanced task integrations.

TaskModuleRunner (Class)

Manages the lifecycle of active modules.

• constructor(modules?: TaskModule[])
• setOptions(options: { onEvent?: (event) => void })
• start({ module, address, config, context }): Starts a new module instance at the specified address.
• stop(address): Stops the module instance at the specified address and records the result.
• stopAll(): Stops all active modules.
• getResults(): TaskModuleResult[]: Returns all results from stopped modules.
• getActiveData(criteria?): Returns live data snapshots from active modules.
• getActiveHandle(criteria): Returns the active handle at an exact scope address.

Shared task helpers

• resolveScopedModuleConfig(raw, moduleId): read module overrides from either modules.<id> or task.modules.<id> with local precedence.
• resolveUniformBlockScopedModuleConfig({ trialConfigs, moduleId, coerce, warnLabel? }): resolve the single block-scoped config for a module across all trial configs in a block; warns and uses the first resolved config when trials disagree.
• maybeExportStimulusRows({ context, rows, suffix }): centralized exportStimuliOnly gate and export finalization path.
• parseSurveyDefinitions(entries): parse mixed survey configs (preset or inline questions[]) into SurveyDefinition[].
• collectSurveyEntries(config, { arrayKey, singletonKey }): collect survey candidate entries from surveys arrays/scoped arrays and optional singleton aliases.
• runSurveySequence(container, surveys, buttonIdPrefix): run sequential surveys with standardized button id handling.
• attachSurveyResults(record, surveys): attach survey run payloads onto a trial/result record.
• findFirstSurveyScore(surveys, scoreKey): read first finite numeric score for a key across survey runs.
• renderSimpleInstructionScreenHtml(ctx, options): simple standardized instruction renderer for intro append + optional block label behavior.
• createInstructionRenderer(options): renderer factory for common task instruction patterns (intro append, block-label policy, optional summary-card section rendering, page resolver hook).
• buildTaskInstructionConfig(args): build standardized task instruction config (intro/pre/post/end, block intro template, block label policy) from raw task instructions + defaults.
• applyTaskInstructionConfig(taskConfig, instructions): apply standardized instruction config onto task instruction surfaces for orchestrator flows.

DRT as a Task Module

The DrtController provides a static helper to use the DRT engine as a task module:

• DrtController.asTaskModule(config): Returns a TaskModule instance configured for DRT.

Canvas helpers

Common display helpers used by tasks:

• computeCanvasFrameLayout
• drawCanvasTrialFrame
• drawCanvasFramedScene
• drawCanvasCenteredText
• drawCenteredCanvasMessage
• createScaledCanvasHost
• mountCanvasElement
• ensureJsPsychCanvasCentered
• renderCenteredNotice

createCanvasPhaseDrawers(scene, drawStimulus): CanvasPhaseDrawers

Builds the standard fixation/blank/stimulus/feedback canvas renderers used by framed RT tasks (stroop, flanker, go_no_go, sft). Every phase draws the shared cue-and-frame scene; only the stimulus content differs per task and is supplied as the drawStimulus(frame, trial) callback.

scene fields:

• layout: a CanvasFrameLayout from computeCanvasFrameLayout
• frameBackground / frameBorder / cueColor
• cueText: static string or (trial) => string provider for per-trial cues (e.g. sft rule cues)
• fixation: { text?, color?, fontSizePx?, fontWeight? } — glyph defaults to "+"

Returned drawers: drawFixation(canvas, trial?), drawBlank(canvas, trial?), drawStimulus(canvas, trial), drawFeedback(canvas, feedback, view) (delegates to drawTrialFeedbackOnCanvas).

5. Scheduling and randomization

buildScheduledItems(args): T[]

Supports:

• weighted (default)
• sequence
• quota_shuffle
• block_quota_shuffle (alias)

Schedule options include withoutReplacement and without_replacement.

Manipulation planning helpers

Core also exports generic helpers used by task adapters for block-level manipulation assignment:

• createManipulationOverrideMap(value):
  • converts [{ id, overrides }] into an id -> overrides map.
• createManipulationPoolAllocator(value, seedParts):
  • creates a participant-seeded pool allocator from a config object like:
  • { "poolA": [ ["manipA"], ["manipB"] ] }
• resolveBlockManipulationIds(blockLike, allocator?):
  • resolves manipulationPool, manipulation, and manipulations into an ordered id list.
• applyManipulationOverridesToBlock(blockLike, manipulationIds, overrideMap, errorContext):
  • deep-merges referenced manipulation overrides into a block object.

These are intentionally task-neutral primitives. Whether a task allows one manipulation per block vs multiple is controlled by the task adapter.

RNG utilities

• hashSeed(...parts): number
• createMulberry32(seed): () => number
• SeededRandom (next, int, shuffle)

Stimulus pool helpers

Core now exports task-neutral stimulus pool primitives:

• Source loading:
  • coerceCsvStimulusConfig(value)
  • loadCategorizedStimulusPools({ inlinePools, csvConfig, resolver?, context? })
  • loadTokenPool({ inline?, csv?, normalize?, dedupe? })
• Draw planning:
  • collectPoolCandidates(pools, categories, excludedCategories?)
  • createPoolDrawer(candidates, rng, drawConfig?)
  • createCategoryPoolDrawer(pools, categories, rng, options?)
• Config coercion:
  • coercePoolDrawConfig(value, defaults?)
  • coerceCategoryDrawConfig(value, defaults?)

Supported draw modes:

• ordered (source order, loops)
• with_replacement (independent random draw)
• without_replacement (shuffle/consume/recycle)
• category drawers also support round_robin

These helpers are used by PM/NBack for participant-seeded deterministic pool behavior.

Prospective memory helpers

Core now exports additive PM utilities in prospectiveMemory.ts:

• generateProspectiveMemoryPositions(rng, { count, minSeparation, maxSeparation })
• resolveProspectiveMemoryCueMatch(context, rules)

Cue-rule primitives support:

• category_in
• text_starts_with
• stimulus_color
• flag_equals

Keyboard Arbitration Policy

Current shared policy for concurrent keyboard modules (primary task + DRT):

1. Primary task keys remain task-owned and are handled in task runtime order.
2. DRT uses controller-level capture listeners and only consumes configured DRT response key.
3. DRT ignores repeated keydown events for its response key until a keyup is observed.
4. If keys overlap by configuration, overlap is allowed and task adapters must explicitly prevent default/propagation where needed.
5. Recommended practice is non-overlapping key maps per task/module pair.

6. Staircase API

QuestBinaryStaircase

Methods:

• nextStimulus()
• update(response: 0 | 1)
• estimateMode()
• exportPosterior()

Helpers:

• buildLinearRange
• luminanceToDb
• dbToLuminance

7. Data and lifecycle

recordsToCsv(records)

Converts object rows to CSV with escaping.

• primitive cells are written directly
• object/array cells are JSON-serialized (instead of [object Object])

CSV row-shaping helpers

• toPrimitiveCell(value): coerce any value to a CSV-safe primitive (objects/arrays JSON-serialized, null/undefined → null).
• nonEmptyString(value): trimmed string or null.
• flattenUnknown(value, prefix, out): flatten nested objects into prefix_key columns; arrays are JSON-serialized rather than expanded.
• pruneEmptyUnknown(value): recursively drop null/blank/empty entries; returns undefined when nothing remains.

finalizeTaskRun(args): Promise<{ submittedToJatos: boolean; redirected: boolean }>

Note: Task adapters do not call finalizeTaskRun directly. It is invoked internally by TaskOrchestrator as part of the session completion flow. This function is documented here for core reference only.

Behavior:

1. Local save when coreConfig.data.localSave !== false, using coreConfig.data.localSaveFormat:
   • default: "csv" (CSV only)
   • "json": JSON only
   • "both": CSV + JSON CSV uses explicit args.csv.contents when provided, else inferred tabular rows from payload where possible. Optional args.extraCsvs allows downloading additional CSV files in the same finalize pass.
2. Submit to JATOS when available.
   • When a core data sink handles JATOS incrementally, finalization does not overwrite streamed result data with a second full-payload submit.
3. endStudy() unless endJatosOnSubmit === false.
4. Resolve and apply redirect template if enabled.

Core Data Sinks

TaskOrchestrator can emit task/session data through a core-level TaskDataSink.

• Default behavior now installs a JATOS JSON-lines sink when JATOS is available.
• Session lifecycle events and trial results are emitted incrementally as envelopes.
• Local CSV/JSON save remains available for testing and debugging.
• Task adapters should not implement JATOS submission directly.
• The final task payload's blocks[] entries contain blockIndex and label, plus an optional semantic blockType copied from the planned block.

TaskOrchestrator Instruction Defaults

TaskOrchestrator.run(args) now auto-derives instruction UI from config unless explicitly overridden:

• task-level pages: instructions.introPages, instructions.endPages
• block-level defaults: instructions.preBlockPages, instructions.postBlockPages
• block intro template: instructions.blockIntroTemplate
• block flags: instructions.showBlockLabel, instructions.preBlockBeforeBlockIntro
• repeat behavior flag: instructions.skipBeforeBlockScreensOnRepeat (skip block-level beforeBlockScreens on retry attempts)
• block page merging: global pre/post pages are merged with block.beforeBlockScreens / block.afterBlockScreens

For adapters that parse/normalize instructions before orchestration, use args.instructionDefaults to provide the same surfaces once (instead of manually wiring introPages, endPages, and getBlockUi per task).

To avoid per-task orchestrator wiring, core also exposes:

• applyResolvedTaskInstructionSurfaces(taskConfig, surfaces)

This hydrates normalized instruction surfaces back onto taskConfig.instructions, so TaskOrchestrator can consume them without adapter-level instructionDefaults.

TaskOrchestrator Staircase Phase

TaskOrchestrator.run(args) supports a standardized pre-main staircase phase:

• args.staircase.run: async callback executed after intro flow and before block/session execution
• args.staircase.enabled (optional): explicit enable/disable override
• if args.staircase.enabled is omitted, orchestrator runs staircase only when taskConfig.staircase.enabled === true

This keeps staircase as a core lifecycle slot (instead of task-specific intro hooks), so any task can use the same orchestration surface.

TaskOrchestrator Module Auto-Start Filter

TaskOrchestrator.run(args) supports:

• args.shouldAutoStartModule(ctx)
• args.resolveModuleContext(ctx)

This predicate is evaluated for each configured task module at block/trial scope before startScopedModules is invoked. It enables selective opt-out of orchestrator auto-start for specific modules without task-level raw config mutation.

Module config resolution is now layered centrally:

1. rawTaskConfig.task.modules (task-level baseline)
2. block.modules / block.task.modules (block-level override)
3. trial.modules / trial.task.modules (trial-level override)

Merging is deep and scope-aware (config.scope still determines whether a module starts at block or trial boundaries).

8. Events, outcomes, feedback

createEventLogger(selection)

Provides .emit(eventType, eventData?, meta?) and accumulated .events.

DRT runtime

• DrtEngine: pure timing/scoring engine for DRT probes (presented/hit/miss/false_alarm, event log export).
• DrtController: browser runtime wrapper over DrtEngine with:
  • scoped start()/stop()
  • keyboard listener lifecycle
  • requestAnimationFrame stepping
  • sampler-based ISI generation via shared core createSampler specs.
  • independent probe displayDurationMs and responseWindowMs
  • responseTerminatesStimulus control
  • optional online parameter transforms (parameterTransforms) that consume drt_response events and emit per-update estimates via onTransformEstimate.
  • transform persistence control (transformPersistence):
    • "scope": reset transform state at each DRT scope boundary (default)
    • "session": persist transform state across all DRT scopes within one task run/session
  • row-level export linking each drt_response to transform output (exportResponseRows()), including:
    • estimate: primary estimate object for the response (or null)
    • transformColumns: flattened scalar columns for long-format analysis (drift_rate, threshold, t0, CI bounds, etc.)
    • estimates: full estimate list (kept for backward compatibility)
  • built-in presentation modes:
    • visual (default: top-center red square anchored to task display area when available; otherwise viewport)
    • auditory (WebAudio tone)
    • border (flash outline around target display element only; does not fall back to full-screen viewport border when target bounds are unavailable)

Continuous tracking helpers

Core now exports reusable tracking primitives (tracking.ts):

• TrackingMotionController:
  • waypoint motion (sampled destinations + linear traversal)
  • chaotic motion (heading jitter + wall reflections)
• TrackingBinSummarizer:
  • accumulates per-window sample counts
  • stores insideCount, outsideCount, and boundary-distance moments for weighted aggregation
• geometry helpers:
  • computeTrackingDistance(point, target) for circle/square boundary distance with inside=0 convention.

Online parameter transforms

• OnlineParameterTransformRunner: generic runtime for event-driven, online parameter estimation modules.
• OnlineParameterTransform: minimal interface (observe, reset, exportState) for reusable model adapters.
• Included first transform type: wald_conjugate:
  • Moving-window analytic shifted-Wald fit from RT observations.
  • Configurable priors (mu0, precision0, kappa0, beta0), window sizes, and credible interval bounds.
  • Non-decision-time (t0) supports:
    • t0Mode: "mix" (default): averages three robust estimators (method-of-moments, conservative min-RT fallback, and Gauss-Legendre MAP node selection over feasible t0 values).
    • t0Mode: "fixed": uses t0 as constant milliseconds.
    • t0Mode: "min_rt_multiplier": uses t0 = t0Multiplier * minObservedRtMs, where minObservedRtMs is tracked across all finite observed RTs for that transform instance (not just the moving fit window).
  • Drift posterior now uses the Gamma-Normal conjugate node moments directly (TMSA), replacing the earlier student-t-based approximation.
  • With transformPersistence: "session", that min-RT tracking persists across DRT scope boundaries in a run; with "scope", it resets each scope.
  • Optional trial-varying prior mean shift mode (priorUpdate.mode: "shift_means") matching the provided R/Python pattern.
  • Transform configs are object entries in parameterTransforms[] (for example { "type": "wald_conjugate" }), not string shorthands.

evaluateTrialOutcome(args): TrialOutcome

Standardized correctness evaluation output used by task adapters.

Trial feedback module

Core exports shared feedback helpers used by task adapters:

• parseTrialFeedbackConfig(value, fallback, defaults?)
• resolveTrialFeedbackView({ feedback, responseCategory, correct, vars?, resolver?, resolverContext? })
• drawTrialFeedbackOnCanvas(ctx, layout, feedback, view)

Supported feedback config fields:

• enabled
• durationMs / duration_ms
• messages.correct|incorrect|timeout|invalid
• messages.byResponseCategory / messages.by_response_category
• style.correctColor|incorrectColor|timeoutColor|invalidColor (snake_case variants also accepted)
• style.byResponseCategoryColors / style.by_response_category_colors
• style.fontSizePx|fontWeight|canvasBackground|canvasBorder (snake_case variants also accepted)

resolveTrialFeedbackView supports {placeholder} interpolation from vars and core variable resolver context.

9. Generic condition planning

enumerateConditionCells(factors): ConditionCell[]

Builds full-factorial cells from named factors and levels.

buildConditionSequence(args): ConditionCell[]

Builds an exact-quota condition sequence using full-factorial cells + optional weights + optional adjacency constraints:

• maxRunLengthByFactor
• maxRunLengthByCell
• noImmediateRepeatFactors

Useful for balanced block construction in any trial task.

parseQuotaMap(raw, labels, expectedTotal, options?): Record<string, number>

Parses a per-condition trial quota map from config. When no quota is provided for any label, trials are split evenly across labels (remainder distributed from the first label onward). When provided, the quotas must sum to expectedTotal or an error is thrown; pass options.taskName to prefix the error message (for example "Stroop").

10. Semantic indexing helpers

buildSemanticIndex(labelsToTerms, options?): Map<string, string>

Creates normalized term -> label mappings from grouped vocabularies.

createSemanticResolver(indexLike, options?): SemanticResolver

Resolver API for normalized semantic label lookup:

• resolve(term): string | null
• has(term): boolean

CSV dictionary helpers

• parseCsvDictionary(csvText, keyColumn, valueColumn, options?)
• loadCsvDictionary(spec)
• loadSemanticIndexFromCsvColumns(csvPath, keyColumn, labelColumns, args?)
• loadTokenListFromCsvColumn(path, column, args?)

These support generic dictionary/lexicon ingestion from CSV-backed assets.

11. Response semantics

createResponseSemantics(categoryToKeys, options?): ResponseSemantics

Creates a normalized key-to-category resolver for RT tasks where physical keys map to abstract response categories.

Capabilities:

• allowedKeys(categories?) for scoped key sets (for example block-specific key availability)
• responseCategoryFromKey(key) with built-in timeout and invalid categories
• expectedCategoryFromKey(key, fallback?) for key-coded expected responses
• expectedCategoryFromSpec(spec, fallback?) for expected responses that may be either:
  • a physical key (mapped to category), or
  • a category label directly (including omission categories like timeout)
• keyForCategory(category) to resolve canonical key output for a category

This is used by current tasks (PM, Stroop, SFT) to avoid task-local key classification logic.

12. Color token registry

createColorRegistry(tokenToColor, options?): ColorRegistry

Creates a normalized token -> CSS color registry with validation and fallback support:

• resolve(token): string | null
• has(token): boolean
• entries()

Helpers

• normalizeColorToken(token)

13. Task Hooks (Extension Hooks)

Core now exposes a generic extension hook runtime intended for cross-task overlays (for example embedding an auxiliary N-back stream into another primary task).

Registration and state

• prepareTaskHooks(hooks, options?)
  • filters disabled hooks (enabled: false)
  • resolves stable IDs
  • orders hooks by priority then declaration order
• createHookStateStore(initial?)
  • shared mutable state map for hook instances (get, set, update, delete, entries)

Lifecycle and event channels

• runTaskHookLifecycle(args)
  • async lifecycle fanout for task_start | task_end | block_start | block_end | trial_start | trial_end
• emitTaskHookEvent(args)
  • async custom event fanout (TaskHookEvent) for task-specific signals
  • suitable for per-trial/per-stage side channels (audio onset, cursor stream ticks, etc.)

Trial planning

• applyTrialPlanHooks(trial, context, hooks?)
• applyTrialPlanHooks({ trial, context, hooks, state?, options? })
• applyTrialPlanHooksAsync(args)

Hooks can transform trial plans and access shared hook state (state) plus per-hook IDs (hookId). Sync API throws if a hook returns a Promise.

Outcome evaluation

• evaluateTrialOutcomeWithHooks(args)
• evaluateTrialOutcomeWithHooksAsync(args)

Hooks can patch:

• inputs before evaluation (beforeEvaluate)
• output after evaluation (afterEvaluate)

Sync API throws if a hook returns a Promise. Both APIs support HookExecutionOptions:

• continueOnError (default false)
• onError(error, context) callback

14. Reaction-time trial helpers

computeRtPhaseDurations(timing, options?): RtPhaseDurations

Computes common RT phase durations from a timing config:

• fixation
• blank
• pre-response stimulus
• response window
• post-response stimulus

options.responseTerminatesTrial (default false) forces postResponseStimulusMs to 0.

runBasicRtTrial(args): Promise<BasicRtTrialResult>

Runs a generic single-trial RT lifecycle with:

• normalized timing decomposition
• shared response capture window
• user-provided render hooks for fixation/blank/stimulus
• optional responseTerminatesTrial phase shaping (for fixed-trial tasks keep this false)

resolveRtTaskConfig(options): ResolvedRtTaskConfig

Resolves an RT task config from:

• a required baseTiming
• optional override object (enabled, responseTerminatesTrial, timing.*)
• default flags (defaultEnabled, defaultResponseTerminatesTrial)

Useful for task-level defaults.

mergeRtTaskConfig(base, override?): ResolvedRtTaskConfig

Merges a partial override onto an already-resolved RT config. Useful for per-block/per-condition overrides in plan-driven tasks.

15. Block summary helpers

coerceBlockSummaryConfig(value): BlockSummaryConfig | null

Parses instructions.blockSummary into a normalized config. Supports:

• enabled
• at: before_post/after_post (aliases for block-end insertion slots)
• title
• lines:
  • string or string[] (shared summary stats for all lines)
  • line object or line object[] with:
    • text (or line / template) template
    • optional where filters (field -> value or array; dotted paths supported; string values support * wildcard and regex: prefix)
    • optional metrics overrides (correctField, rtField, metricField)
• when filters (blockIndex, blockLabel, blockType, isPractice)
• where trial-result filters (field -> value or array of values; supports dotted paths, * wildcard strings, and regex: string patterns)
• metrics.correctField, metrics.rtField (supports dotted paths)

buildBlockSummaryModel(args): BlockSummaryModel | null

Builds a computed block summary from block metadata and trial results. Template variables include:

• {blockLabel}, {blockIndex}, {blockIndex1}, {blockType}, {isPractice}
• {total}, {correct}, {incorrect}, {accuracyPct}, {meanRtMs}, {validRtCount}
• {blockSpawned}, {blockCleared}, {blockDropped}, {blockPoints}
• {experimentSpawned}, {experimentCleared}, {experimentDropped}, {experimentPoints}

renderBlockSummaryCardHtml(model): string

Renders a simple HTML card from a summary model for tasks that use custom waitForContinue screens.

computeBlockSummaryStats(args): { total, correct, accuracyPct, meanRtMs, validRtCount }

Computes filtered summary stats from trial results using where + metrics. Both where keys and metric field names can use dotted paths (for example game.stats.cleared). Useful when non-UI control flow (for example, retry logic) should use the same scoring semantics as block summary screens.

16. Block repeat helpers

coerceBlockRepeatUntilConfig(value): BlockRepeatUntilConfig | null

Parses a block-level repeatUntil object into normalized form. Supports:

• enabled
• maxAttempts
• minAccuracy (0..1) and minAccuracyPct (0..100 alias)
• minCorrect, minTotal
• maxMeanMetric, minMeanMetric (mean absolute value of metrics.metricField)
• where trial-result filtering (supports dotted paths)
• metrics.correctField, metrics.metricField (supports dotted paths)

evaluateBlockRepeatUntil(args): BlockRepeatEvaluation

Evaluates pass/repeat decisions for one block attempt from trial results. Returns:

• passed
• shouldRepeat
• reason (threshold_met, threshold_not_met, max_attempts_reached, disabled)
• attempt-local stats (total, correct, accuracy, meanMetric)