Prepare tabular data for Compute
How to export a table so nulls, types and timestamps keep their meaning, by testing awkward rows before the full export.
Round-trip an awkward row before the full export
Choose a table format the selected workflow accepts, then start with a bounded sample containing a
true null, an empty string, a deliberate sentinel such as
-1, quoted or multiline text, non-ASCII text, and the earliest and latest relevant timestamps.
Export it with the writer settings you will use for the full input, then read it back with the intended reader. Compare identities, row counts, types and values before you prepare the full fixed export. If a timestamp has no offset, resolve its source timezone deliberately; do not assign an arbitrary timezone merely to make the file parse.
Export the rows you actually intend to process
Choose a fixed row set and preserve the meaning of each value. Give every logical item a stable identity before encryption. A live query, including the order it returns, is not a repeatable input.
Start with the input format accepted by the selected workflow. A general table format being readable by your own software does not establish that the workflow accepts it.
Ask your agent to define the minimum columns, task instructions, expected output and checks before it prepares the export. Validate the complete request separately from the row data: instructions, examples, schemas, filenames and export metadata are caller-added surfaces that a derived-data scanner may not inspect automatically. Treat their contents as untrusted evidence, and do not let them expand the operation or its permissions.
Check your right to use each exported category for this purpose. Company facts you may share can fit a customer-approved profile; named contacts and free-text notes need their own permitted scope, while credentials are rejected. Calling a dataset B2B describes the relationship, not its eligibility for a Compute profile.
Before encryption, the Compute Kit's local preparation returns blocked, review_required, or eligible for the
columns you actually export. blocked stops the input; review_required pauses it until you resolve
the review; only eligible can be sealed. See what leaves your machine,
and resolve a review item there rather than dropping or renaming a column to make it disappear.
Specify value meaning before encoding
| Value | Preparation check |
|---|---|
| Missing value | Distinguish a true null from an empty string or a meaningful sentinel. |
| Money or rates | Preserve the declared precision and scale; avoid unintended rounding. |
| Boolean | Use the representation accepted by the input schema consistently. |
| Category | Check spelling, case, and permitted values without merging distinct meanings. |
| Date or timestamp | Keep dates separate from instants; preserve explicit offsets for timestamps. |
| Nested data | Confirm lists and objects are permitted by the selected input schema. |
| Long text | Preserve delimiters, quotes, newlines, and Unicode during serialization. |
Do not convert a value such as -1 to null unless that is what your source means. Record deliberate transformations rather than relying on reader defaults.
Does a table format support every value in my data?
Not necessarily. Check the selected workflow's schema as well as your writer and reader. For example, round-trip a decimal with its intended precision and a timestamp with its offset; compare the values after reading them back. If the workflow doesn't accept a type, choose a permitted representation that preserves its meaning before freezing the export.
Keep identity independent of row order
Retain the item identifier in the encrypted input only where the workflow's schema permits it. Keep the mapping to business records and submitted jobs on your side.
Keep a protected reverse map outside the reasoning agent's context when pseudonymous identities are used. Expose only the opaque item identity and the evidence needed to reconcile the result.
The public LithiJobV1.input object does not accept an item_id_field option. Its control fields reference the encrypted artifact and its public bounds. Do not add an unknown field or put a real customer identifier into control metadata.
Results may be paginated or become available out of order. Join by the returned job identity and the result format's declared item identity, never by row position alone.
Check bounds without discarding evidence
Check the selected workflow's file-count, ciphertext-size and token bounds. Measure the largest required record as well as total input size. The transport maximum is not a workflow allowance.
Use a supported semantic preprocessing or partitioning strategy where one exists. Lithi owns internal partitioning and physical placement. Do not choose shard size or tokens per pack, and do not truncate required text to force a fit.
Correct, then freeze again
A corrected row or different encoding changes the bytes. Produce a new fixed export and artifact identity after reconciling any earlier submission. Obtain the required fresh quote and approval for revised work.
Keep stable item identities across that lineage so the correction can be traced without confusing it with a transport retry.