Question and scope

This protocol measures how often a cross-team decision is reversed without a new decision record. A reversal means the team is now doing something different from what the recorded decision specified. The measure matters because an unrecorded reversal leaves the old decision in place while the work follows a new course, so anyone reading the log is misled. The output is a share, a median time to reversal, and a list of decisions that need a written correction.

Scope is one coordination team and a defined window, such as all decisions recorded in the last two quarters that affected more than one team. Include decisions recorded in meeting notes, a decision log, or a ticket, provided the record states the choice and the owner. Exclude routine scheduling changes that never altered a commitment. State the window and the source systems so the population can be rebuilt.

What counts as a reversal

A reversal is a change to the chosen option, the owner, the scope, or the date that a reasonable reader would treat as a different decision. A minor wording correction is not a reversal, and neither is a status update that reports progress. The test is whether the new instruction contradicts the recorded choice rather than adding detail to it. Write two examples of each before coding so the boundary is visible.

Distinguish a reversal from a refinement. A refinement narrows how the decision will be implemented while keeping the same choice and owner. A reversal changes the choice or moves the commitment enough that the old record no longer describes the work. The distinction is the largest source of disagreement, so resolve it in the dictionary rather than in the analysis.

What counts as a new record

A new record exists when someone has written the changed decision in the same place, with a date and an owner, and marked the earlier entry as superseded. A chat message that announces a new date is not a record unless it is linked from the decision log. A verbal agreement is not a record even when everyone in the room heard it. The standard is whether a new coordinator could find the current decision without asking a colleague.

Decide whether a separate document counts as a record. A revised project plan can count when it clearly states the decision change and identifies the approver. A slide deck without a date and owner does not. Record the location and the evidence type for every reversal so a reviewer can check the same source. When in doubt, treat the document as a candidate and verify the date and owner before counting it.

Unit of analysis and exclusions

The unit is one decision with one choice, one owner, and one effective date. Split a bundled decision when each part has its own owner or commitment. Keep one unit when several items share a single choice and a single review date. Use the record as it existed when the decision was made, not a version cleaned up during the audit.

Exclude duplicates created when a meeting note and a decision log both describe the same choice. Exclude decisions that were withdrawn before any work began, because there is nothing to reverse. Report the exclusion count and the reason for each category. A clean denominator is more valuable than a large one with unclear edges. Keep the removed identifiers in a short appendix so the count can be rebuilt.

Building the decision list

Start with the decision log and the recurring review notes. Search for phrases such as agreed, approved, decided, and going with, because many decisions are recorded in prose rather than a dedicated field. Add decisions that appear only in a ticket comment if they changed a cross-team commitment. The list should include the original choice, the owner, and the date, even when those details are scattered across two sources.

Give each decision a stable identifier and record where the evidence lives. Note whether the record includes a rejected alternative, because a decision without one is harder to interpret later. Freeze the list before reading the later work records so the sample is not selected around known reversals. Keep the frozen list with the analysis. Store the identifier in the working file, not only in the report, so the coding sheet stays traceable.

Evidence of the original decision

Start from the original record and write down the choice in one sentence. Add the owner, the effective date, and the commitment the decision affected. Then read the later records in date order and note the first point where the work diverges from that sentence. Use the earliest dated evidence, not the most dramatic message, because the measure is about when the change became real.

Where the original record is ambiguous, mark the decision as unclear rather than guessing. Count unclear units separately and exclude them from the main share. Report the unclear count so the reader can judge how much the result depends on interpretation. A smaller, clearer population is better than a larger one built on assumptions. Write the reason for each unclear unit in one line so the pattern is visible.

Coding the reversal

Code each decision as recorded, reversed with a new record, reversed without a new record, or reversed and later restored. Add the date of the first evidence and the date the record was corrected, if it ever was. A reversal without a new record that is corrected within the same week is a different problem from one that stays uncorrected for months. Report both the share and the correction delay.

Record who noticed the reversal and whether the discovery came from a routine review, an audit, or a complaint. This field shows whether the team's own checks catch changes before a customer does. Do not infer motive, and do not label a reversal as a failure when the original decision was simply wrong. The measure is about record integrity, not about whether the new choice was better.

Sampling and verification

Use a census when the decision list is small, which is common for a single coordination team. Above roughly one hundred decisions, draw a random sample and document the method. Stratify by decision type or by the team that owned the choice. Freeze the sample before reading the later work records. Record the export date and the filters used so another analyst can reproduce the frame.

Have a second reviewer code at least ten decisions independently. Compare the reversal classification and the date of first evidence, because those are the fields most likely to differ. Resolve disagreements by returning to the source record and write down the resolution. If the disagreement rate is high, revise the definitions and recode the full sample. Keep both versions of the coding sheet so the changes can be reviewed.

Decision rules before coding

Decide whether a reversal that was later reversed back to the original choice counts once or twice. Counting it once measures whether the work ever left the recorded choice, while counting twice measures churn. Write the choice down before coding and apply it consistently. Report the number of restored decisions separately so the reader can see the churn either way.

Decide how to treat a decision whose owner changed but whose choice stayed the same. A new owner is not a reversal unless the record fails to name the successor. Also decide whether a change announced by someone without authority counts as a reversal, or as an unauthorised instruction to be escalated. Both classifications are useful, but they answer different questions.

Analysis and reporting

Report the share of decisions reversed without a new record, the share reversed with a record, and the share unchanged. Add the median days between the first evidence of reversal and the correction. Show raw counts beside every percentage and suppress groups with fewer than five units. A single headline number hides whether the problem is frequency or delay. Show the denominator beside the numerator so a small population is obvious.

Break the result down by decision type, owner group, and review cadence. Look for a category where reversals cluster, because that may point to an intake problem or a decision that was never really agreed. Pair the share with the time to correction, since a fast correction is a much smaller risk than a silent change that persists. Use the priority dispute rulebook when a reversal is really a contested decision that needs a new hearing.

Limits and sources of error

The cited sources describe internal control and project management in general terms; they do not provide a benchmark for reversal rates in a specific decision log. A missing correction may mean the team forgot to update the record, or that the change was never treated as a decision. The measure also depends on whether the log preserves superseded entries instead of overwriting them. A system that edits in place will understate the problem.

Decisions made in conversation are undercounted when no one writes them down. Check a sample of meeting notes and messages to estimate that gap, and report it as a limitation. If the log is known to be incomplete, present the result as a range rather than a precise share. An honest caveat is more useful than a number the team will argue about.

Review cadence

Review the result at a fixed point and end with one change, one owner, and one check date. A practical first change is to require a superseded marker and a new date whenever a decision changes. Repeat the test after two full cycles before setting a target, because a single quarter may reflect an unusual reorganisation. Keep the definitions and the sample method with the result so the next reviewer can compare like periods.

Retire the measure if the log reliably captures reversals and the correction delay stays short. If the team keeps it, publish the share and the delay together, because a low share can still hide a few very old silent changes. Use the weekly blocked-work review to surface reversals before they become blockers. The purpose is a decision log that tells the truth, not a perfect count.

Sources