Restatement panels: keep original and revised observations
A restatement panel preserves multiple reported values for the same economic observation. It should support both what was reported at the time and what the latest eligible report says now. Replacing old values in place destroys that distinction and can make an AI-generated trend silently mix original and revised history.
SEC XBRL API documentation describes company concepts, units and facts across submissions. It also notes that frame data selects a last-filed fact fitting a calendar period. A convenient aggregate is therefore not a substitute for an explicitly selected original-versus-revised panel.
Define the observation independently of its version
Use entity, metric concept, dimensional scope, period and unit to identify the economic observation. Then attach accession or filing identity, publication date, reported value and revision relationship as version fields. The same period can appear in later comparative reporting without representing new economic activity.
Do not classify every changed tag or value as a restatement. Differences can arise from unit scale, duration, consolidation perimeter, taxonomy mapping or presentation. Inspect the applicable report’s explanation before assigning a reason. Keep unexplained differences unresolved rather than allowing AI to invent an accounting cause.
Build two explicit selections. The original panel uses the first appropriate reported observation under the defined rule. The revised panel uses the latest eligible comparable observation as of a stated cutoff. Preserve all candidate facts and the selection logic so a reviewer can challenge which version was chosen.
Hypothetical trend reversal
Assume fictional Company Moss originally reports revenue of 100 for Year 1 and 120 for Year 2. A later comparable disclosure revises Year 1 to 110 while Year 2 remains 120. Original-series growth is 20%; revised-comparison growth is about 9.1%.
Both calculations can be reproduced, but they answer different information questions. A memo discussing the original reaction may need the 100 value. A current comparison may need 110. Reporting 20% growth beside a footnote claiming all values are restated would be internally inconsistent.
Now suppose the later document’s 110 belongs to continuing operations while the original 100 covered a broader perimeter. The two cannot be labeled a clean numerical restatement until scope is understood. This caveat prevents the panel from laundering a definition change into a simple correction. All values are hypothetical.
Revision panel worksheet
| Observation key | Version fields | Selection fields |
|---|---|---|
| Entity and concept | Filing identity | Original eligibility rule |
| Period and duration | Publication date | Revised cutoff |
| Units and dimensions | Reported value | Comparable scope status |
| Perimeter | Stated reason or unresolved | Chosen version and rationale |
Calculate a revision delta only for compatible observations: revised value − original value. Store the percentage basis explicitly if a ratio is also reported. A difference divided by the revised value is not the same statistic as one divided by the original value.
Give AI one selected panel at a time
Provide the chosen values, version labels and scope notes before asking for narrative interpretation. Do not pass a flat pile of duplicate periods and ask the model to select the correct number by intuition. A model can produce a coherent story while selecting a different version in each sentence.
Check that totals and derived ratios use the same panel policy. Mixing revised revenue with original cost or original share counts can create combinations never reported together. If compatible revisions are unavailable, limit the comparison or show components separately. The final packet should make the revision’s effect observable without erasing the original information environment.
FAQ
Should the original figure be deleted after a correction?
No. Preserve it with its filing identity, while selecting the appropriate revised figure for the current question.
Does the latest value always make a comparable series?
No. Confirm period, unit, perimeter and metric meaning. Latest is a version rule, not proof of economic comparability.