Truthound · All articles · 23/24
Evidence-led Truthound Core guidance on data contracts, check inputs and operational limits.
480 articles
OpenTelemetry Rust JSON: checking exemplar values separately from metric values
A readable metric does not prove that its exemplar wire contract survived. Test the two independently.
jq delpaths: agreeing on deletion targets before mixed-sign paths change an array
Hand over deletion targets resolved against the original snapshot, rather than positions in an array that is changing during deletion.
jq large-integer subtraction: keeping raw numbers in the team handover
A small difference can still be wrong when its inputs are large integers. Preserve the original numbers because precision may be lost before subtraction.
fast-csv headers: distinguishing field selection from renaming
Explain field selection separately from output-title changes so different teams do not assign different meanings to the same headers option.
DSPy OutputField constraints: checking the generated schema and returned score separately
Structured-output experiments should separate schema generation from post-response business-rule validation.
Docling video export: checking picture source lists before handing JSON to another team
A completed conversion is different from a reusable document. Video and document teams should share the reload contract.
Dask and Zarr parallel writes: inspect shard ownership before trusting a readable file
Overlapping partial updates to one shard can separate successful completion from pixel preservation.
Investigating Dask string truncation at compute with a first-block dtype control
A sufficiently wide declared dtype does not establish preservation of the final concatenated strings.
Check key ownership before treating Dask partition groupbys as final results
A key spanning partitions can produce separate partial aggregates that are mistaken for a complete record.
Dask left-merge row loss: inspect shuffle buckets at the 128 boundary
Partition scaling tests need both bucket representation boundaries and preservation of left-hand rows.
Checking valid boundary keys when Dask divisions contain NaN
Missing-key join semantics and disappearance of valid keys require separate decisions.
Dask map meta differs from the computed dtype: include unmatched keys
A lookup that can create missing values requires separate checks of declared metadata and the computed result type.
Dask left-merge index provenance: do not invent a right-side ID for an unmatched row
Identify the source role of an output index before treating its display value as a business identifier.
Dask isin with string and integer codes: identical display text can select different entities
Compare mixed-object filtering against the original combination of type and value.
Investigating Dask task fusion when independent chains receive the same key
Preserve the mapping between logical and executable work before relying on output-file existence.
Checking Dask divisions length separately from correct row values
The relation between partition count and boundary count is not itself a verdict about data loss.
Dask date-column assignment changes an integer column: check the scope of the change
Adding a date column must preserve the types and values of columns outside the requested change.
Dask versus NumPy cumulative sums: define an error budget before judging chunk boundaries
Interpret floating-point cumulative-sum differences against an independent reference and an agreed error budget.
Validate Dask complex Cholesky with conjugate-transpose reconstruction residuals
Check complex matrix factors against the original equation rather than copying a transpose check from a real-valued example.
Dask merge_asof followed by drop raises KeyError: preserve intermediate dependencies
A column absent from the final output may still be required until the join has finished.
