How do I…?
Start from the outcome you need, run the smallest relevant Grid example, anticipate its common failure, and jump to the exact language contract.
How do I…?
Use this page when you know the outcome but not the syntax. Each path gives you something runnable, the failure most likely to mislead you, and the exact contract to read before expanding the pattern.
This index follows the 0.65.0 candidate documentation. Linked examples, guided builds, and films retain their own evidence version until recaptured.
If you have not completed one model run yet, begin with Get to a running model.
Pick an outcome
| I need to… | Start here |
|---|---|
| Clean and validate text | Normalize, extract, and reject bad text |
| Join, group, and window live tables | Build a relational result |
| Choose Array, List, Map, Set, or Deque | Choose a collection by behavior |
| Recover from errors without hiding them | Preserve the failure contract |
| Work with time, history, and revisions | Choose the correct coordinate |
| Call external functions safely | Bound the asynchronous work |
| Build a surface and bind writable state | Keep the model authoritative |
| Optimize, simulate, or solve symbolically | Choose the right analysis family |
| Train a model with a calibrated upper prediction | Keep the model version and endpoint evidence together |
| Govern a research workflow | Preserve artifact, policy, and workflow identity |
| Select distributed or disconnected execution | Choose the Backbone topology explicitly |
| Test, profile, and harden a production model | Turn a claim into evidence |
Clean and validate text
Run it. Complete Debug and harden a model, then open the Text and quality canonical model. Change one raw string and follow its normalized value, extracted parts, validation flag, and readable issue summary.
Common failure. Blank, empty string, and invalid text are different states. A broad IFERROR can turn malformed input into plausible output, while an unanchored regex can accept a valid-looking substring inside an invalid value. Preserve a validation result before adding a fallback.
Exact reference. Use Trim and clean, Regex, and the TRIM function entry. For metadata-backed input constraints, read Data validation.
Join, group, and window live tables
Run it. Complete Build a live revenue dashboard or inspect the smaller Live revenue KPI model. Make one source-row change and verify both the relation and bound surface update.
Common failure. Aggregation can silently change row identity, and a window without a deliberate partition and order can be deterministic yet answer the wrong question. Do not copy a snapshot into an Array when the result must remain attached to a live resident table.
Exact reference. Start with Joins, sets, windows, and subqueries, then check the native relational query recipe.
Choose Array, List, Map, Set, or Deque
Run it. Follow From arrays to data pipelines and compare its immutable array work with the Collection pipelines model.
Common failure. Choosing by familiar syntax instead of behavior creates awkward or invalid operations: Array for shaped spill computation, List for ordered growth, Map for keyed lookup, Set for uniqueness, and Deque for work at both ends. Converting repeatedly between them usually hides a missing ownership decision.
Exact reference. Use the Choosing a structure table and Arrays, spilling, and comprehensions.
Recover from errors without hiding them
Run it. Complete Debug and harden a model and compare the finished pattern with the Text and quality canonical model. Introduce the guided build's deliberate error, inspect the diagnostic, and apply the narrow repair before adding a fallback.
Common failure. IFERROR catches every error value. It can hide a broken reference, bad type, or authorization failure that should stop the result. Catch only the expected condition, and remember that an external pending state is not itself an error.
Exact reference. Read Error-handling operators and clauses, especially DEFAULT / ?? and TRY … ELSE.
Work with time, history, and revisions
Run it. Inspect the Temporal coordinates model and complete Explain and validate a decision. Compare a current value, a prior saved revision, and a simulated trajectory rather than treating them as interchangeable timestamps.
Common failure. TODAY() or NOW() answers model time now; it does not retrieve a saved revision. A history query has no trustworthy point-in-time answer until the relevant revision exists, and a simulation trajectory is a modeled path rather than an audit log.
Exact reference. Read Temporal references, Temporal predicates, and the simulation contract.
Call external functions safely
Run it. Complete External enrichment, safely or inspect the External enrichment canonical model. Exercise its successful, pending, denied-capability, and fallback paths.
Common failure. An external call has a status lifecycle, capability boundary, cache policy, and cost. Calling it eagerly per row can create unbounded fan-out; treating pending as failure can publish stale fallback data as if it were fresh.
Exact reference. Use Status lifecycle, Always add a fallback, and Capability requirements.
Build a surface and bind writable state
Run it. Complete Turn a model into an app and compare the CRM pipeline, Store map, Revenue dashboard, and Scenario console surface families.
Common failure. Reimplementing a calculation in the UI creates a second source of truth. A control cannot write a computed binding: the target must be declared writable input state, and its action must remain inside the model's capability boundary.
Exact reference. Choose from Built-in surfaces, follow Model bindings, and use Authoring custom UI only after the binding contract is stable.
Optimize, simulate, or solve symbolically
Run it. Use Optimize a pricing decision for a bounded decision, Run a warehouse simulation for state over model time, or the Symbolic mathematics model for exact transformation and proof-oriented results.
Common failure. These tools answer different questions. An unconstrained optimization can return a mathematically valid but unusable choice; an unseeded stochastic simulation cannot be reproduced; and a decimal approximation is not an exact symbolic certificate.
Exact reference. Compare Analysis and optimization, the simulation guide, and Exact solving.
Train and use a calibrated prediction
Run it. Start with Train and integrate predictive models. Train an exact candidate version, inspect MODEL_UNCERTAINTY, call PREDICTION_BOUND for an admitted endpoint, and pin the version only after review.
Common failure. A successful training job can still produce a point-only model. An unversioned formula can follow an older pin, and an upper P90 is a marginal coverage target rather than a per-row guarantee or full predictive distribution.
Exact reference. Read Train and use calibrated predictions and the calibrated model bounds formula contract.
Govern a research workflow
Run it. Use the Research Workspace workflow that matches the claim: reproducibility, use policy, bibliography, specialist validation, or scholarly release. Keep every status transition bound to the exact model revision and artifact identities.
Common failure. A completed UI step does not prove scientific validity, external execution, DOI issuance, repository deposit, or current policy after a competing mutation. Research services also require Pro product selection and deployment-owned governance authority.
Exact reference. Read Operate Research Workspace services for edition, authorization, backup, executor, and recovery boundaries.
Select a Backbone topology
Run it. Begin with ordinary in-process or v1 Backbone. Move to exact-node v2 or DDIL only when a reviewed deployment contract requires its trust, durable replay, recovery, or disconnected-authority behavior.
Common failure. Treating v2 as a transparent connector upgrade bypasses its node identity, trust, state, and recovery obligations. DDIL does not supply consensus, arbitrary merging, automatic failover, or exactly-once external effects.
Exact reference. Use Choose a Backbone topology and preserve every selected node, trust, state, and route identity in deployment evidence.
Test, profile, and harden a production model
Run it. Complete Verify a canonical example before Promote a model to production, then inspect the Treasury control plane canonical model as the larger operator-dependent case. Keep compile evidence, fixture inputs, assertions, and runtime receipts separate so each claim has the right proof.
Common failure. A successful compile does not prove a runtime output, /healthz does not prove Runtime Host readiness, and one observed value without its source hash, fixture, runtime identity, and revision cannot be replayed as a receipt.
Exact reference. Use Keep dependencies narrow, Troubleshoot Grid, and the Production checklist.
From recipe to larger model
Before promoting any pattern, record the outcome, prerequisites, complete source, deterministic checkpoint, deliberate failure and repair, Grid version, source provenance, compiler result, and runtime receipt when behavior is claimed. That same unit is suitable for a learner, a regression test, and a high-quality language-model training record.