Skip to content
← All writing
Technical · 2 min read

The 'diff every number' rule: catching stale carryover in cyclic reports

Recurring regulatory reports fail QC at the copy-paste step. Here's the paragraph-diff workflow I use to catch it — and the small tool that automates it.

Data QACompliance reportingAutomation

On this page

A lot of environmental work runs on cyclic reports — the same compliance document, produced every quarter or every six months, for years. Post-closure landfill monitoring is a classic example: a semiannual report on groundwater, surface water, and landfill gas, built on a controlled template that mostly stays the same while the numbers underneath it change.

That structure is efficient. It’s also where quality quietly fails.

Where cyclic reports break

The failure mode isn’t dramatic. It’s a sentence from the last cycle that should have been updated this cycle and wasn’t:

  • A value carried over from the previous event because someone updated the table but not the paragraph that references it.
  • A “no exceedances were observed” line that’s still true — or quietly isn’t.
  • A date, a well ID, or a trend statement that no longer matches the current data.

None of these throw an error. The document still reads cleanly. It just isn’t true anymore — and in a regulatory submittal, that’s the failure that matters.

The rule: diff every number against its source

The discipline I hold to is simple to state and tedious to do by hand: every value in the narrative gets checked against its source for this cycle — not assumed correct because it was correct last time. Tables against lab reports, narrative against tables, trend statements against the multi-year record.

Done manually across a long report, that’s exactly the kind of repetitive task where attention erodes near the end — right where carryover errors hide.

Automating the boring part

So I built a small Python tool that does the first pass for me. It diffs the new report against the previous cycle’s template paragraph by paragraph and color-codes the result:

  • Unchanged text — carried over from last cycle (verify it should be unchanged).
  • Updated text — changed this cycle (verify the change is correct).
  • New / pending text — added or flagged for review.
For each paragraph in new_report:
  match = closest_paragraph(prior_report)
  if identical(paragraph, match):      tag = "carryover  → confirm still true"
  elif similar(paragraph, match):      tag = "updated     → verify change"
  else:                                tag = "new/pending → review"

It doesn’t decide whether the report is right — judgment stays with the engineer. What it does is make the carryover visible, so the review focuses on the lines most likely to be stale instead of re-reading the whole document with fading attention.

Why it matters

A compliance report has to survive project-manager review, client review, PE certification, and a regulator. The cheapest place to catch an error is before any of them see it. “Diff every number” is the habit; the tool just makes the habit faster and harder to skip.

The broader point: a lot of engineering quality lives in unglamorous, repetitive review steps. Knowing enough to automate the repetition means you spend your attention where judgment is actually required.

Related writing