Governing the Determinism Spectrum in the Collector's Court

Skill selection belongs to the model; a clerk cannot compel a skill to fire. Commands invert that. A command is operator-invoked — a named, logged, auditable entry point into a procedure that runs the same way for every clerk, every bench assistant, every officer in the chain. You cannot make the model choose the right path; you can hand the office a path the model cannot decline.

• /register-appeal — registers a new appeal under MLRC § 247AppealsSection 247 provides the appeal against orders of subordinate revenue officers — the Tahsildar, Naib Tahsildar and SDO — to the appellate authority; alongside it sit revision (§ 257) and review. The jurisdictional doorway to every file in this pilot.Maharashtra Land Revenue Code, 1966 ↗ with structured metadata — parties, village, survey number, impugned order, the officer who passed it — and computes the limitation position at the door rather than at the first hearing. A registration that “usually captures everything” is a defective file in waiting; as a command it is a guaranteed gate with a record that it fired. registration
• /scrutiny — runs the case-type checklist deterministically: the 7/12 extractसातबारा · village forms VII & XIIThe record-of-rights extract for a land parcel — occupants, tenure, crops — maintained under the MLRC and the Record of Rights rules. The single most-demanded document in any revenue appeal, and the first thing found missing., the mutation extractMLRC § 150The register of mutations records changes in rights over land; disputed entries go to the register of disputed cases. Appeals against mutation entries are a staple of the Additional Collector's docket.MLRC, 1966, § 150 ↗, the map, the certified copy of the impugned order, the lower record — and issues a deficiency memo the same day, not at the hearing three months later. deficiency
• /cause-list — assembles the day's cause-list from the case database: party readiness, pending notices, prior order-sheets — one invocation replacing the manual re-assembly before each sitting. hearings
• /hearing-record — the structured capture of a hearing: recording or upload, type (Roznamaरोजनामा · order sheetThe running diary of a revenue case — each hearing's date, appearances, what transpired, and the interim directions. The procedural backbone an appellate or revisional authority reads first. / party statement / officer note), participants, hearing date — routed to the right slot in the file every time, in the same shape, by every member of the hearing staff. roznama
• /case-brief — generates the officer-ready brief before a hearing from the indexed record — so the third officer to touch the file inherits a structured brief instead of re-reading 687 pages from scratch. The prose is the model's; that a brief exists, in the office's format, before every hearing, is the command's. preparation

whether registration, scrutiny and the hearing record happened at all.

A workflow phase written in natural language leaves both interpretation and code generation to the model at runtime — the right default for novel work, the wrong one for a calculation that must come out the same for every appellant. Once a phase has been generated, reviewed and approved, the code is frozen as a versioned script and the phase reduced to “invoke it.” The non-determinism of re-derivation collapses into a reviewed, auditable artifact.

• Limitation arithmetic. The appeal window computed deterministically from the date of the impugned order, with the time spent obtaining the certified copy excluded the same way every time. Whether an appeal is in time decides jurisdiction; it is precisely the arithmetic that must never be improvised afresh per file. limitation
• Deficiency checklist per case type. Which records are mandatory for a mutation appeal versus a partition or demarcation matter is a frozen, versioned table — not a judgment call the model re-makes per file. The AI reads documents; the list of what must exist is code. deficiency
• Digitisation & indexing pipeline. Scanning, OCR and indexing of Marathi, Hindi and English records as a reproducible pipeline: the same 687-page file yields the same indexed record on every run, with every extracted event tied to its page. Corrections are new versions, not silent drift. record
• Pendency & dashboard metrics. Cases this month, pending review, long-pending, reserved orders — computed by query, not narrated by a model. A dashboard the Collector reads must never be a hallucination surface. pendency
• Bilingual order templates. The skeleton of a final order — cause title, parties, procedural history from the Roznama, operative-part scaffolding — populated deterministically from case metadata in Marathi and English, removing transcription error from the very documents that get appealed. orders

limitation dates, mandatory-record lists, and the contents of the indexed record.

Hooks fire deterministic code at fixed points in the loop. Their headline job here is at the exit — inspecting a draft before it is allowed to be “done” — but the same mechanism also gates each tool call: a PreToolUse hook can allow, deny or rewrite a call before it runs, and a PostToolUse hook inspects the result once it returns. Any of them can gate the work, and a stop-hook can re-drive it: a failed check feeds its reason back to the model as the next instruction, so the failure is worked off rather than landing on the officer's desk. One guardrail is mandatory at the exit: the harness exposes a flag that is true when the model is already in a forced continuation, and the stop-hook must honour it or it will loop forever.

• Provision & citation verifier. Every MLRC section, rule and case-law citation in a draft order is checked against the office's statute store — the Code as amended through the configured cut-off date — and against authoritative law reports before the draft can leave the harness. This is the mechanical answer to the run of Indian fabricated-citation incidents, to the Supreme Court's misconduct warning, and to the Kerala High Court's AI policyFirst binding judicial AI policy · July 2025The Kerala High Court's policy on AI tools in the district judiciary (19 July 2025): AI is never a substitute for decision-making or legal reasoning, only approved tools may be used, and all AI-produced citations and translations must be meticulously verified.Kerala HC AI Guidelines (PDF) ↗ requirement that AI citations be meticulously verified — enforced as a gate, not a reminder. citations
• Source-link gate. No paragraph of a draft order or case brief may survive without a page-level citation into the digitised record (p.3 · p.239). An unsourced finding cannot be the final turn — the “every line cites its source” promise of the pilot, made unskippable. provenance
• Natural-justice gate. Before an order finalises, a deterministic check that audi alteram partemNatural justice“Hear the other side” — the rule that no one is condemned unheard. For a quasi-judicial authority, an order passed without notice and opportunity of hearing is liable to be set aside regardless of its merits. is satisfied on the record: notices issued and served on every respondent, appearances or ex-parte procedure recorded in the Roznama. An order that skipped a hearing is void on arrival; the gate makes it unreachable instead. natural justice
• Reasoned-order gate. Every operative conclusion must be supported by recorded reasons tied to the record — the duty the Supreme Court fixed on quasi-judicial authorities in Kranti Associates(2010) 9 SCC 496Kranti Associates Pvt. Ltd. v. Masood Ahmed Khan: the Supreme Court's synthesis of why quasi-judicial orders must record reasons — reasons restrain arbitrariness, enable appeal and revision, and are as indispensable as natural justice itself.Kranti Associates v. Masood Ahmed Khan ↗. A bare, unreasoned draft is blocked with the gaps named. reasons
• Egress & data-protection gate. Blocks any output that would route case records or citizens' personal data to an un-approved destination — the concrete control behind “no public AI model training on government data,” on-premise deployment, and the DPDP ActDigital Personal Data Protection Act, 2023India's personal-data statute. Land-dispute files are dense with citizens' personal data — names, holdings, family partitions — and a government platform processing them needs purpose-bound, auditable data flows. obligations that come with a file full of citizens' holdings and family disputes. confidentiality
• Record-access gate. A PreToolUse hook confines each retrieval to the cases the requesting role is permissioned for; a PostToolUse hook stamps provenance on every document returned and writes the access to the audit trail — case-wise permissioning enforced at the tool boundary, not in a policy document. access

Where the property needs judgment rather than a pass/fail rule — “does this order actually deal with every ground the appellant pressed?” — a prompt- or agent-type hook runs a cheap evaluator in the same slot. The shape is identical; only the oracle changes.

#!/usr/bin/env python3
"""Stop-hook gate: a draft order cannot leave the harness until every cited
provision verifies against the statute store, every paragraph carries a
page-level source link, and the natural-justice checks pass on the record.

The harness invokes this when the model tries to end its turn. Emitting a
"block" decision returns control to the model with `reason` as its next
instruction; exiting 0 silently lets the turn finish.
"""
import json, sys
from collectorate.verify import (
    unverified_provisions,   # MLRC as amended through the cut-off
    unsourced_paragraphs,    # every finding needs a page citation
    natural_justice_breaches, # notice served, hearing on the Roznama
)

payload = json.load(sys.stdin)

# Honour the forced-continuation flag, or this gate loops forever.
if payload.get("stop_hook_active"):
    sys.exit(0)

draft = payload["last_output"]
problems = (unverified_provisions(draft)
            + unsourced_paragraphs(draft)
            + natural_justice_breaches(draft))
if problems:
    print(json.dumps({
        "decision": "block",
        "reason": "Resolve before finishing:\n- " + "\n- ".join(problems),
    }))
sys.exit(0)

cited provisions, source links, notice & hearing, reasons, and data egress.

A loop construct runs a task repeatedly until a satisfactory outcome is reached. It is not deterministic on its own — but it raises the floor of what the office can expect, and it composes naturally with hooks: the hook is the oracle, the loop is the driver. The honest caveat is that a loop is only as trustworthy as its verifier; a loop without a real test simply burns time.

• Draft order, until clean. Redraft until the provision verifier, the source-link gate and the reasoned-order gate all report clean — convergence on verifiable properties, not a single best-effort pass that the officer must then debug. orders
• Case timeline, until sourced. Refine the reconstructed history of the case — notices, hearings, mutations, prior orders — until every event links to its page in the digitised file; no orphan facts surviving into the brief the officer relies on. timeline
• Hearing record, until complete. Iterate transcription and structuring until every hearing on the cause-list has its Roznama entry, participant log and order sheet attached to the right case. roznama

that a draft is only “finished” once a named test is satisfied.

A custom wrapper drives the harness through its API or CLI — the most deterministic surface, because the control flow lives in code the office owns rather than in the model's discretion. A first headless call returns a session identifier; subsequent calls resume that session, preserving full context across turns, so an external program can hold a case open from registration to disposal, insert hard human gates between phases, and bound iteration. The model supplies capability; the wrapper supplies the guarantee that nothing issues from the office unsigned.

• Officer sign-off gate. “AI shall not issue orders” made structural: the issue step is unreachable in code without the Additional Collector's recorded approval. The AI drafts; the workflow makes it impossible for a draft to become an order any other way. sign-off
• Role-routed workflow. Record-room staff digitise, scrutiny staff run checklists, hearing staff capture Roznamas, the officer decides — the wrapper enforces who can trigger what, so case-wise, role-based permissioning is workflow fact rather than access-policy aspiration. roles
• Objective docket record. The wrapper emits a structured, real-time view of every appeal — what ran, what is pending, what is deficient, what is overdue — so the Collector supervises on an objective record rather than on what a monthly return chooses to report. supervision
• Tamper-evident custody. The wrapper owns the thread that drives every document pull and every edit, writing each to an append-only audit trail — the chain of custody that lets a digitised record stand behind the certificate an electronic record needs in evidence. audit

"""Deterministic envelope around the harness for an appeal file.

The office — not the model — owns the control flow: it opens a session,
holds it open by session id across turns, and makes the "issue" step
structurally unreachable without the Additional Collector's sign-off.
"""
import json, subprocess

def run(prompt, session=None):
    """Run one turn of the harness; return its parsed JSON result.

    :param prompt: the operator instruction for this turn.
    :param session: a session id to resume, or None to open a new case thread.
    :returns: the result dict, including `session_id` for continuation.
    """
    cmd = ["claude", "-p", prompt, "--output-format", "json", "--max-turns", "8"]
    if session:
        cmd += ["--resume", session]
    return json.loads(subprocess.run(cmd, capture_output=True, text=True).stdout)

draft = run("Draft the order for appeal 118/A-247/2026. Do not issue anything.")
sid = draft["session_id"]

# The order cannot issue without the competent authority's approval.
if additional_collector_signs_off(draft["result"]):
    run("Finalise the approved order and enter it in the case record.", session=sid)

what leaves the office as an order, and who signed it.