← Back

Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice

← All guides

How to Handle ADMT Data Correction Requests: From Intake to Re-Run Decision

The short answer: A correction request says: data about me that the tool used is wrong — fix it, and look at my decision again. The duties behind it: Colorado — tell the person how to correct their data and act on corrections they send. California — corrections surface through access and review routes, with the answer built from the same record. Connecticut — for housing, correct the data and reevaluate the decision on the corrected data (since July 1, 2026). The workflow has five steps — intake, locate, verify the fact, correct and re-run, record — and the record step is the one that proves the other four happened.

Check your duties and dates — free Build your first notice — free

What is a correction request? It is a person's notification that specific data used in an automated-assisted decision about them is factually incorrect, with the correct fact supplied or identifiable. It is not a general appeal (that is human review) and not an opt-out (that stops future automated use). The three routes are named side by side on the kit's notices and explanations, and they often travel together — a correction frequently triggers a re-run, and a review often begins with a correction.

Intake: a contact that works, a log that starts immediately

Corrections arrive through the correction contact printed on the notice and the explanation. Intake needs three properties: the contact is monitored by a named owner; each request is logged on receipt with its date and the decision it concerns; and the requester gets an acknowledgment that names what happens next. The most common correction failure is not refusal — it is a request sitting unread in a mailbox that a departed employee set up.

Ask for the specific fact. A request that says 'your data is wrong' needs one clarifying exchange: which item, what should it be, and what shows that? The explanation's data summary is the shared reference — the person can point at the line that is wrong. This is another reason the data summary must be written so the person it describes can find an error in it, as covered in How to Write an Adverse-Outcome Explanation.

Locate and verify: find the fact, check the fact

Locate the item in the decision's data summary and its source. Corrections divide into three kinds: a transcription error (your system holds it wrong), a source error (the supplier's file is wrong about this person), and a matching error (another person's record attached to this file). The fix differs by kind — edit, dispute upstream, or unmerge — but the record should note which kind it was, because patterns of one kind point at a systemic problem worth more than any single case.

Verify the correct fact with proportionate evidence: the person's document, the source's confirmation, or your own records. Correction is not automatic acceptance of any asserted fact — it is acting on corrections, in Colorado's phrasing — but the verification standard should be one a person with a genuine error can actually meet. Record what was checked, briefly, beside the correction itself.

Worked example: Teton Ridge Apartments

Example — fictional business, real duty structure. Teton Ridge is a fictional housing operator screening tenants in Colorado and Connecticut. One quarter of correction traffic, handled through the five-step workflow:

Act and re-run: the correction changes the decision, or you say why not

Once corrected, the data feeds the decision again. Connecticut states the endpoint expressly for housing — the decision is reevaluated using the corrected data — and the same logic is the honest reading of Colorado's 'act on corrections they send': a correction that changes nothing, not even the record the next decision will use, has not been acted on. Re-run the decision on the corrected inputs; if the outcome changes, the person gets the new outcome; if it does not, they get that too, with the reason the correction did not move it.

Both versions stay in the record: original decision on original data, re-run on corrected data, with dates. Deleting the original destroys the audit trail of what the person originally questioned. The kit's hash-chained record is built for exactly this — corrections and re-runs append to the chain rather than rewriting it. Connecticut's housing version of this duty is detailed in Connecticut Profiling Opt-Out Rights.

Timing and communication: no invented clocks, no silent waits

The kit's data states no fixed day count for completing a correction in Colorado or California, and this guide does not invent one. (Connecticut's 45 days governs opt-out answers; Colorado's 30 days governs explanations.) What the workflow owes the person without a statutory number is still concrete: acknowledgment at intake, a decision on the correction without unexplained delay, and a closing communication that says what was corrected, whether the decision was re-run, and what the outcome is.

Measure your own times from the log and publish them internally. A correction queue with a median of six days and a visible tail is manageable; a queue nobody measures discovers its tail from complaints. The monthly correction report — received, corrected, re-run, outcomes changed — is four numbers and one honest paragraph, and it doubles as evidence the duty is being acted on, not just offered.

Patterns: corrections are quality data about your pipeline

Aggregate corrections monthly by kind and by data source. Ten matching errors from one supplier is a supplier problem; repeated stale balances point at a refresh cadence; transcription errors cluster where manual entry survives. Each pattern fixed upstream prevents a class of wrong decisions, not just a class of complaints — the correction path is the cheapest audit your decision pipeline will ever get.

Feed the patterns into the places that already ask for them: California's risk assessment has a risks element that should name your observed error patterns, and its safeguards element should name the upstream fixes. Handled this way, corrections, assessments, and reviews stop being three compliance chores and become one quality system with three doors. Confirm the correction duties for your state in the free checker.

  1. Put a monitored, named correction contact on every notice and explanation — and log every request on receipt
  2. Ask for the specific fact: which item, what it should be, what shows it — against the explanation's data summary
  3. Classify each error (transcription, source, matching), verify proportionately, and record what was checked
  4. Correct the data and re-run the decision; keep both versions; tell the person the outcome either way
  5. Review corrections monthly by kind and source, and fix the upstream pattern, not just the case

Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.

Build your first notice — free

Related guides

See all ADMT guides for the full set.

Frequently asked questions

What is a data correction request?

A person's notice that specific data the tool used about them is factually wrong, with the correct fact supplied or identifiable. Colorado's duty is to tell people how to correct their data and act on corrections they send.

Must the decision be re-run after a correction?

In Connecticut housing, expressly yes: the person can correct inaccurate data and have the decision reevaluated using the corrected data. Elsewhere the kit treats re-running on corrected data as the working standard — a correction that changes nothing has not been acted on.

Is there a deadline for handling corrections?

The kit's data states no fixed day count for corrections in Colorado or California. Connecticut's 45-day standard governs opt-out answers, and Colorado's 30 days governs explanations — different duties, different clocks.

What if the asserted correction is wrong?

Verify proportionately. If the data is accurate, keep it, record what was checked, and tell the person what was verified and what happens next (including the review route).

Should the original decision be deleted after a re-run?

No. Keep both versions in the record: the original decision on the original data, and the re-run on the corrected data, with dates. The chain is the evidence.

Build your first notice — free