Request: question, context, identity
Semantics: concepts, rules, versions
Data: snapshot, watermark, quality
Execution: plan, query, model
Output: hash, chart, limitation
A screenshot proves only what was displayed
Data corrections, semantic releases, authorization, model updates, and cache expiry can change an answer. A screenshot lacks input and calculation evidence.
Generate a response_manifest for each answer. Distinguish exact replay of old inputs, rerun on current data with old logic, and restatement under new logic.
Begin with request and identity
Persist normalized and original question, inherited conversation conditions, locale, timezone, organization, project, subject, role, and policy version.
Protect sensitive text through classification and encryption. Reopening another user’s answer still requires current authorization rather than inheriting the author’s scope.
Freeze semantic rules
Record concept IDs, metric versions, dimensions, enumerations, date roles, currencies, defaults, and document versions. Final SQL alone cannot explain why one business definition was selected.
Use immutable semantic releases or content hashes. If an old rule must be deleted, retain only permitted audit evidence and report loss of replay fidelity.
Describe data snapshots, not “latest”
Capture snapshot ID, partitions, watermark, quality state, and backfill version per critical dataset. Without retained history, logic may be rerunnable while numbers are not reproducible.
Keep original and restated snapshots distinguishable after source corrections, then attribute differences to data, semantics, or execution.
Persist plan and execution evidence
Store plan nodes, compiler version, SQL hash, parameters, engine, query ID, rows, grain, timeout, sampling, cache, and validation. Exclude credentials and sensitive literals from ordinary manifests.
A governed plan can be recompiled after table or dialect migration. Old SQL need not remain executable forever if equivalent semantics can be demonstrated.
Version model behavior and randomness
For parsing and narration, retain provider model identifier, prompt template, tool versions, decoding, and safety policy. Hosted models can change even at zero temperature.
Numbers should come from query results, not model memory. Narrative replay may allow wording variation while facts, limits, and citations remain invariant.
Hash the output under deterministic rules
Store schema, rows, ordering, precision, result hash, chart configuration, narrative version, and citations. Undefined order, floating precision, timezone, and rounding can create false differences.
Exports and shares reference manifest_id. A new renderer may restyle the answer but should declare that it is not the original presentation.
Balance retention with privacy
The manifest can contain sensitive prompts and results. Apply classified retention, access, encryption, and deletion. Metadata may outlive detailed rows when policy permits.
After deletion, an irreversible fingerprint may remain, but complete replay may no longer be possible. Report a reproducibility level instead of a blanket traceable label.
Validate with categorized differences
Test fixed data, late corrections, semantic change, revocation, model update, and engine migration through replay, rerun, and restatement.
Classify differences as input_data, semantic_rule, authorization, execution, model_summary, or rendering. A manifest that cannot localize change is incomplete.
AskTable.ai boundary
Require a unique manifest ID, complete versions and watermarks, controlled replay, current authorization, stable hashes, categorized differences, and enforceable retention.
AskTable.ai may be evaluated as the query and answer-record entry point. This article does not establish complete manifests, retained historical snapshots, or deterministic model replay in the current product; infrastructure support must be verified.
Public references
Ready to help your team start?
Talk through a real scenario and see how AskTable.ai can fit your business.
Book a demo