Documenting Why the Agent Was Overruled

The section officer rejected fourteen scholarship drafts in a week. The interface offered a thumbs-down. She used it. The vendor dashboard reported 14 overrides, engagement healthy. When CAG picked file 4418, nobody could say whether she had seen a better certificate, disagreed with the circular, or simply did not trust the machine on Fridays. The human was in the loop. The loop had no language.

Override logging is the eighth field in the action schema made visible. It protects the officer (she can show she thought), protects the citizen (the real reason is on the file), and protects the programme (you can see whether the corpus is wrong or the desk is rubber-stamping). Silence helps no one except a dashboard.

Guide for CIOs who own the desk UI. 16 August 2026. Not legal advice.

Codes, then a sentence

Start with a short list. Officers will not complete an essay at 17:50.

Code	Meaning	Sentence you still want
WRONG_RETRIEVAL	The chunks were the wrong law or the wrong year	Which document should have been used
HALLUCINATED_CITE	Invented section or paragraph	What the officer used instead
FACTS_CHANGED	Citizen produced a new paper	What arrived and when
POLICY_DISAGREE	The draft applied a rule the officer reads differently	The reading, with a circular cite
TONE_OR_LANGUAGE	Wrong script or an unfit tone	Which language the citizen actually used
OTHER	Escape hatch	A real sentence — OTHER without text is a fail

Put the code and the sentence into the system of record, not only into the agent UI. If eOffice is where the speaking order lives, the override lives there or is linked there. A vendor dashboard that dies with the tenant is not the file.

Accept is also a reason — sometimes

You do not need an essay on every accept. You do need a signal if accepts are happening in two seconds on money-class files. That is not a gate. That is a stamp. Internal audit should sample fast accepts. The override article is also the rubber-stamp article.

Modified — the officer kept a paragraph and rewrote another — should store a diff or the final text. Otherwise you cannot say what the machine contributed. CAG will ask. So will a court.

Do not train on the widow

Overrides contain personal data and official disagreement. Mine them as patterns: 'circular 2024 is being preferred over 2023 in district X' is a corpus ticket. Shipping the raw rows to a foundation-model trainer is how your internal legal debate becomes someone else's weights. Purpose-limit the override store. Retention-map it. The DPO signs the mining job.

Two weeks of thumbs

Objections

Officers say they have no time. Answer: one code and one line. If that is impossible, the workload is the finding, not the log.

Unions say reasons will be used against them. Answer: silence is used against them first, when a bad draft goes out under their name. The sentence is a shield.

The vendor wants free text only, for 'sentiment'. Answer: free text without a code cannot be sampled. Codes first.

A scientist wants every override as RLHF. Answer: not on our citizens, not in their cloud.

A four-week playbook

• Week 1: add the six codes and a mandatory sentence on reject and modify. Disable naked thumbs.
• Week 2: land the override on the official packet, not only the vendor UI.
• Week 3: sample fast accepts and OTHER-without-text. Fix the UI or the workload.
• Week 4: open a redacted pattern review with the DPO. Turn patterns into corpus or prompt tickets, not training jobs.

File note you can paste

Subject: Recording reasons when an officer overrules or modifies an agent draft.

An accept, reject or modify without an official identity is not a human gate. Reject and modify shall carry a reason code and one sentence in the system of record. Fast unexamined accepts on citizen-facing or money-class workflows will be sampled by internal audit.

Override rows will not be shipped to a vendor for training. Pattern mining, if any, will be redacted and purpose-limited. This note is not legal advice.

Using overrides without punishing the desk

If officers believe override codes will appear on their APAR, they will accept bad drafts. Say, in writing, that codes are for corpus and prompt hygiene, and that rubber-stamping is the conduct issue, not disagreement. Then keep the promise. A dashboard of 'officers who overrule most' on a secretary's wall is how you kill the gate.

The useful dashboard is codes by workflow: WRONG_RETRIEVAL climbing on one circular is a content ticket; HALLUCINATED_CITE climbing after a prompt change is a version ticket; POLICY_DISAGREE climbing in one district may be a training need or a local circular you never indexed. Those are programme objects. They are not names.

When the officer is wrong and the agent was right, the packet still shows both. A later appeal can see that the human chose. That is the constitutional order. If the same officer is systematically wrong on the same rule, that is supervision, not a model update. Do not ask the machine to outvote a post. Ask the competent authority to train or to transfer.

Modified drafts should keep the agent's text as an object, not only the final. Courts will ask what the machine contributed. If you only store the final, you have hidden the clerk. Hidden clerks are how departments later claim the human did everything, until the log says otherwise.

• Disable OTHER without a sentence at the UI layer, not in a style guide.
• Sample ten rejects a month for empty or sarcastic sentences.
• Never put override text into a vendor 'improvement' tenant.

Informational field guidance. Confirm office-procedure rules on noting and DPDP duties on the override store.

How this survives CAG, RTI or the Board

“Documenting Why the Agent Was Overruled” is not a workshop slide. A P1 CIO/CTO will have to reconstruct a decision after the officer who clicked approve has been transferred. Write the artefact that lets a stranger replay the case: the log fields, the approval, the override, the register row.

A human gate that leaves no reason is a rumour. When the officer overrules the agent, capture a short code and a line of official text. That row is how you prove the loop and how you improve the machine without training on the citizen. India AI Governance Guidelines (November 2025) are guidelines, not a statute. DPDP still allocates fiduciary duty. Delegation of Financial Powers still allocates who may spend. Do not hide those instruments behind the word governance.

If you cannot show who acted, on which purpose, with which data class, and who could have refused, you do not have accountability. You have a chatbot with a charter PDF.

• Name the owner of “override logging AI” inside the department, not the vendor.
• Keep CERT-In-relevant logs in India for the required period.
• Store overrides with a reason an auditor can read.
• Put the workflow on the AI register before it touches a citizen.

Close this loop before the next CAB

Put “Documenting Why the Agent Was Overruled” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P1 CIO/CTO, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “override logging AI” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

What the next noting must contain

“Documenting Why the Agent Was Overruled” belongs in a file, not only in a search result. A P1 CIO/CTO should be able to point at one artefact that proves “override logging AI”: a packet capture, a processing schedule, a scored evaluation row, a dated notice, or a refusal rule. If the only evidence is a slide, you have a heading.

A human gate that leaves no reason is a rumour. When the officer overrules the agent, capture a short code and a line of official text. That row is how you prove the loop and how you improve the machine without training on the citizen. DPDP 2023 does not define sovereign AI and does not write a blanket localisation rule for every model hop. CERT-In’s 28 April 2022 directions still set specified incident clocks and 180-day log retention in India for in-scope events. The November 2025 AI governance text is guidance, not a statute. A Proprietary Article Certificate, when it is lawful, lives in GFR Rule 166 — not Rule 161.

Write three dated sentences under C6 Governance & Audit: what was decided, which designation owns it after the next posting order, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the designation that owns “override logging AI”, plus a deputy.
• Attach one artefact a stranger can open next year.
• Name the instrument you are actually using — Act, direction, GFR clause, GeM term, or guideline paragraph.
• Leave unsourced percentages, GMV slides and house forecasts out of the noting.
• Revisit when the model, the SI, the notice, the region or the posting changes.