This page is generated from the project’s canonical curriculum state. Research depth and story coverage are tracked separately: a subject can be well mapped before a story is commissioned.
Evidence reviewed through State version 1.11.0
A research badge reports how far the territory has been explored. Inside each subject, stories are grouped by Beginner, Advanced, and Professional reasoning level; only populated levels appear.
Where does work run, what do boundaries promise, and how can behavior evolve safely?
Where does work run, wait, contend, and remain alive after cancellation?
What does each participant actually promise across a boundary?
How can behavior change without losing intent?
Which state is authoritative, how is it found, and what survives replay or partial failure?
Where is truth stored, which version is authoritative, and how is it found?
Who may act after partial failure, and what ordering can observers rely on?
How does meaning survive delivery, replay, time, and multiple engines?
How do platforms remain observable, economical, recoverable, and authorized?
Which boundaries let teams change, operate, and recover independently?
How does useful work degrade, fail, recover, and become diagnosable?
Which resource limits useful work, and what does a successful outcome really cost?
Who or what may cause an effect, and what evidence establishes trust?
How do probabilistic components become dependable, bounded products?
How does data become a decision that remains valid in a real workflow?
13 defined content areas
Which model behaviors can the application constrain, measure, or afford?
16 defined content areas
How do models, tools, state, protocols, media, and people complete bounded work?
20 defined content areas
How do we prove an AI system is dependable, contained, observable, and governable?
All content areas below are research-ready. Story status shows which parts of the subject have been turned into public teaching material.
Define the population, unit of decision, desired outcome, observable proxy, action set, intervention owner, time horizon, baseline, and cost of each error before selecting a model.
No story has been commissioned for this area.
Specify how an outcome becomes an authoritative, versioned label, including observation delay, adjudication, correction, missingness, censoring, and the effect of intervention.
Build reproducible features and semantic representations from only the evidence available at decision time, with the same meaning across training and serving.
No story has been commissioned for this area.
Evaluate whether scores, rankings, classes, or generated signals support the intended decisions across relevant slices and uncertainty regions.
No story has been commissioned for this area.
Convert model evidence into a deterministic, versioned policy for action, non-action, evidence gathering, abstention, review, escalation, and appeal.
No story has been commissioned for this area.
Co-design threshold, prioritization, deferral, and reviewer capacity so the deployed system remains timely, learnable, and safe under the actual score and uncertainty distribution.
Evaluate and learn when actions determine which outcomes occur, which outcomes become visible, and which counterfactuals remain unknowable.
No story has been commissioned for this area.
Define and operate decision quality for materially different populations, contexts, and harms, with a usable path to challenge or correct a decision.
No story has been commissioned for this area.
Detect loss of decision validity and identify whether the cause is population, data, feature, label, model, calibration, policy, threshold, reviewer, intervention, or outcome change.
No story has been commissioned for this area.
Compare decision systems without exposing users or operations to uncontrolled interventions or evaluating a shadow system on outcomes it did not cause.
No story has been commissioned for this area.
Reproduce and change the complete decision system across data, features, labels, preprocessing, model, calibration, policy, threshold, reviewer workflow, and environment versions.
No story has been commissioned for this area.
Use LLM or multimodal outputs as versioned, evaluated decision evidence without treating structured generation, explanation, self-confidence, or a model judge as authoritative truth.
No story has been commissioned for this area.
Defend a production decision system whose population, labels, features, models, policies, reviewers, interventions, and governance controls change on different clocks.
No story has been commissioned for this area.
All content areas below are research-ready. Story status shows which parts of the subject have been turned into public teaching material.
Define supported tasks, inputs, segments, outputs, failure states, and the deterministic authority that may accept a generation.
No story has been commissioned for this area.
Reproduce what the model received and test representation, position, language, stop, cost, and evidence-loss effects.
No story has been commissioned for this area.
Choose generation controls while keeping business invariants outside sampling assumptions and seeds.
No story has been commissioned for this area.
Use constrained decoding for structure while enforcing semantic and domain rules before authority; bound repair and expose unsupported output.
Allocate context and reasoning by task value, quality evidence, deadline, cache, and shared-capacity impact.
Choose prompting, examples, SFT, PEFT, preferences, reward adaptation, or no training from the behavior, evidence, update rate, and rollback contract.
No story has been commissioned for this area.
Preserve source rights, generation lineage, coverage, filtering, deduplication, diversity, and frozen evaluation boundaries.
No story has been commissioned for this area.
Translate intended behavior into bounded feedback without mistaking a proxy or verifier for the task objective.
No story has been commissioned for this area.
Change one capability without silently erasing another; operate compatibility, coexistence, and rollback.
No story has been commissioned for this area.
Route only within an eligible set; distinguish route, cascade, fallback, retry, and abstention; evaluate rare slices and full-path cost.
Decompose prefill and decode, latency, sequence lengths, concurrency, warm-up, and admission against useful-output service objectives.
Balance reuse against queue fairness, load, eviction, trust boundaries, and in-service-level accepted outcomes.
No story has been commissioned for this area.
Treat serving optimizations as behavioral releases validated on the real runtime and workload.
No story has been commissioned for this area.
Compare colocated, partial, and separated execution using interference, KV transfer, locality, compatibility, and workload evidence.
No story has been commissioned for this area.
Operate the complete behavior tuple across canaries, aliases, mixed versions, provider change, and rollback.
Defend a fleet whose behavior, routing, serving, adaptation, evidence, and tenant commitments change independently.
All content areas below are research-ready. Story status shows which parts of the subject have been turned into public teaching material.
Place each decision in deterministic code, an explicit workflow, a constrained model decision, or a bounded agent loop.
No story has been commissioned for this area.
Survive loss, deployment, waiting, and failover while separating history, checkpoints, workflow state, and business truth.
Convert probabilistic intent into an attributable, retry-safe or reconcilable operation.
Bound time, cost, steps, fan-out, authority, and blast radius across a task tree—and verify what actually stopped.
Prove who acts, on whose behalf, for which target, and under which current effect-specific approval.
Reconstruct decisions, checks, effects, and progress; evaluate both outcome and path across repeated trials.
Choose a coordination topology only when its benefit exceeds added cost, authority, and failure modes.
No story has been commissioned for this area.
Operate tool and remote-agent protocols as lifecycle, trust, authorization, cancellation, and reconciliation contracts.
No story has been commissioned for this area.
Define what may be remembered, for how long, with what provenance, and how retraction propagates.
No story has been commissioned for this area.
Roll out, roll back, govern, and incident-manage an agent whose independently versioned parts outlive releases.
No story has been commissioned for this area.
Place media edges, session owners, reconnect state, and regional boundaries deliberately.
No story has been commissioned for this area.
Choose speech architecture from latency, evidence, control, accessibility, and operating constraints.
No story has been commissioned for this area.
Model conversational timing without confusing stopped speech, cancelled work, and revoked effects.
Resume a conversation without replaying speech, model work, tool execution, or business effects.
Reconcile long-running results with the user’s current goal, target, and approval intent.
Bind interpretations and actions to evidence from the correct moment and modality.
No story has been commissioned for this area.
Allocate latency and capacity across media, perception, generation, tools, and response.
No story has been commissioned for this area.
Separate short-lived media access from current authorization to cause business effects.
No story has been commissioned for this area.
Preserve consent, identity, disclosure, and an equivalent completion path across modalities.
No story has been commissioned for this area.
Evaluate task success, timing, accessibility, policy, and causal evidence across every interaction plane.