Path 3 — Government workbench (field-inspector pilot)¶

For a national or sub-national government office whose investigators audit recruitment agencies, screen incoming complaints, or investigate trafficking cases in the field. Examples in this audience: Philippines DMW (Department of Migrant Workers, successor to POEA), Indonesia BP2MI, Bangladesh BMET, Nepal Department of Foreign Employment, Vietnam DOLAB, UK GLAA (Gangmasters and Labour Abuse Authority), Australia Fair Work Ombudsman, US DOL Wage and Hour Division, Saudi MHRSD, UAE MoHRE, Qatar MoL.

This document covers the field workbench path. For the bulk processing API path, see 03_government_api_integration.md.

What you get¶

A handheld AI workbench an inspector can carry into a recruitment agency audit, port-of-entry interview, or worker-complaint intake office. It:

Photographs a recruitment contract or job ad, runs OCR + Gemma 4 vision, produces an instant red-flag report
Looks up the relevant section of your domestic statute and the matching ILO convention article for any flagged practice
Produces a draft inspection report or violation notice with the citation matrix already filled in
Logs every query to an immutable audit trail (FOIA-defensible)
Operates fully offline once the model and knowledge layer are loaded

Same software as path 1 (single NGO), with rubric weights tuned to your jurisdiction and an FOIA audit layer enabled.

What you provide¶

Resource	Pilot	Production
Inspector hardware	1 ruggedised tablet or laptop per inspector (16 GB RAM, GPU optional)	Same; standard issue across the inspectorate
Storage	30 GB free per device	Same
Network	Initial install + jurisdiction tuning over secured network	Optional WAN sync for updated rubric or new statute releases
Operator	2 to 5 inspectors for pilot + 1 IT helper	All inspectors after pilot
Domain expertise	1 senior inspector to co-author the jurisdiction tuning	Same; quarterly review
Legal review	1 staff lawyer to verify statute-citation matrix	Same; quarterly

Pilot scope¶

In scope (30 to 60 day pilot):

Recruitment agency audit workflow: inspector visits a licensed agency, scans 5 sample worker contracts, gets instant red-flag flags + statute citations + draft inspection notes
Port-of-entry interview workflow: inspector interviews a returning worker, transcribes salient facts into the workbench, gets a preliminary trafficking-indicator assessment + referral pathway
Complaint intake desk: front-desk officer accepts a walk-in complaint, workbench helps classify (trafficking / unfair recruitment / contract substitution / wage theft) and routes to the right unit
Jurisdiction tuning pass: senior inspector + staff lawyer review the rubric weights, citation matrix, and templates for your agency's specific statutes

Out of scope (Phase 2):

Integration with your existing case-management system — handled in 03_government_api_integration.md
Cross-agency data sharing — requires data-sharing MOUs that take longer than a pilot window
Public-facing complaint intake — initially restricted to inspector-mediated use
Anything that would require the workbench to call out to a third-party model API — never enabled on government deployments

Deployment topology¶

┌─────────────────────────────────────────────────────────────────┐
│ Inspector device (ruggedised tablet or laptop)                   │
│  - duecare workbench (FastAPI on localhost:8080)                 │
│  - Gemma 4 E2B or E4B loaded from local disk                     │
│  - bundled knowledge pack: tuned for THIS jurisdiction           │
│  - FOIA audit log: append-only SQLite                            │
│  - Disk encryption: agency standard (BitLocker / FileVault)      │
└─────────────────────────────────────────────────────────────────┘
                            │
                            │ (no network egress in field)
                            ↓
┌─────────────────────────────────────────────────────────────────┐
│ Agency internal network (only when inspector returns to base)   │
│  - upload audit log to central archive                          │
│  - download updated rubric / new statute releases               │
│  - never calls out to public internet                           │
└─────────────────────────────────────────────────────────────────┘

Jurisdiction tuning¶

Every government deployment goes through a one-week tuning pass before the pilot. The senior inspector and staff lawyer work with DueCare to:

Verify the citation matrix. Each ILO convention article and destination/origin-country statute referenced in the bundled knowledge pack is checked against the agency's authoritative register. Anything stale, mis-cited, or repealed gets corrected.
Weight the cluster thresholds. The bundled rubric uses six clusters with default thresholds (legal grounding 60%, harm avoidance 80%, worker protection 70%, pattern recognition 60%, resources 70%, ethical framing 70%). A government rubric usually wants different weights — more legal grounding for a regulator, more pattern recognition for an enforcement unit, more resources for a complaint-intake desk.
Author jurisdiction-specific templates. The 34 bundled templates cover the major regulators globally. Your agency may want 2 to 4 additional templates specific to your enforcement tools (administrative penalty notice, license suspension order, referral to prosecutor, joint operation request to another agency).
Configure the FOIA audit layer. What gets logged, who can read the log, retention period, redaction rules for log exports.
Train the inspectors. Two-day classroom on the workbench; five field shadows on real (mock) audits.

The tuning produces a duecare-jurisdiction-<agency>-config.yaml file plus an updated knowledge pack. This is checked into your internal repo (not the public DueCare repo).

See 03_government_jurisdiction_tuning.md for the full tuning checklist.

FOIA audit log¶

Every query an inspector runs against the workbench produces an audit log entry:

{
  "request_id": "req-2026-05-22-9f8c4d2",
  "investigator_id": "inspector-id-redacted-for-this-doc",
  "device_id": "device-007",
  "query_at": "2026-05-22T14:30:00Z",
  "query_type": "audit_recruitment_contract",
  "query_summary_hash": "sha256:af83c2d1...",
  "input_artefact_hashes": ["sha256:11ee..."],
  "rubric_dims_evaluated": ["fee_camouflage_recognition", "contract_substitution", "..."],
  "statutes_returned": ["POEA-MC-14-2017-Sec-3a", "RA-8042-as-amended-Sec-6c", "ILO-C181-Art-7"],
  "citations_returned": [{"id": "...", "verbatim": "...", "source_url": "..."}],
  "workbench_version": "0.1.0-rc4",
  "knowledge_pack_version": "2026-05-22-jurisdiction-dmw"
}

Properties:

Append-only. No edits, no deletes from inside the workbench.
Hash-linked. Each entry includes the hash of the previous entry, forming a chain that detects tampering.
Plaintext queries never stored. Only SHA-256 hashes of the query and input artefacts. The plaintext is in the inspector's case file, governed by your existing case-management retention rules.
Exportable for FOIA. A separate command produces a redacted log export suitable for FOIA / right-to-information disclosure. Inspector IDs and device IDs are tokenised; statutory references and rubric dimensions are kept verbatim.

Multi-language¶

Your inspectors and the workers they interview probably do not speak the same language as the statute or the ILO convention text. The workbench supports:

Input in any language Gemma 4 covers (Tagalog, Bahasa, Vietnamese, Bengali, Nepali, Arabic, Spanish, Portuguese, English, French, Mandarin, Cantonese, Korean, Japanese, Hindi, Urdu, Punjabi, Tamil, Sinhalese, Burmese, Khmer, Thai, Lao, Russian, Ukrainian, Romanian, Bulgarian)
Statute citations in the agency's working language (English
your national language, side by side)
Templates in any pair: input language for the worker section, agency language for the inspector / legal section

A jurisdiction tuning may add a third language (e.g., a Saudi MHRSD deployment may want Arabic + English + Tagalog because of the worker base coming from PH).

Pilot success metrics (30 to 60 days)¶

Metric	Target
Audit time per recruitment agency	30% reduction vs. current process
Statute citations correct (verified by staff lawyer)	100% on the first 100 audits
Inspectors using the workbench voluntarily after week 2	80%+
Complaints reclassified by workbench review (caught missed trafficking)	At least 1 per week on average
Audit log completeness	100% of queries logged; chain integrity verified weekly
Network egress events from inspector devices	0

Procurement bullet points¶

For when your procurement officer needs the one-pager:

License: MIT. No per-seat or per-query fee. Modify and self-host freely.
Vendor lock-in: None. The model (Gemma 4) and the knowledge layer are open. If DueCare goes away tomorrow, your inspectors can keep using the workbench indefinitely on the version they have, and you can fork the codebase to maintain it yourselves.
Data residency: All data on agency hardware. No SaaS, no third-party model API, no cloud egress. Air-gapped deployment supported.
Audit: FOIA-defensible audit log built in. Hash-chained, exportable, tamper-evident.
Compliance: Designed around ILO conventions (15 covered), Palermo Protocol, regional anti-trafficking treaties, and the destination/origin-country statutes for the 11 major migrant- labour corridors.
Support: Open-source community + paid support tiers available from DueCare project staff.
Source: github.com/TaylorAmarelTech/gemma4_comp. Public repository; agency self-hosts a private fork.

What can go wrong + how to handle it¶

Failure	Mitigation
Inspector's tablet runs out of battery in the field	Local cache lets the inspector keep collecting notes; sync to workbench when battery is back
Inspector incorrectly accepts a Gemma 4 statutory citation that is hallucinated	Citation matrix on every output is shown verbatim; inspector training drills the "verify the citation in the bundled register, not the model's claim" practice
Worker has insufficient command of the agency's working language to consent to an interview	Workbench prompts for an interpreter; logs the language barrier; flags case for follow-up
Senior inspector retires; jurisdiction tuning expertise leaves	Tuning is checked into the agency's internal repo as YAML; the next senior inspector reviews and updates rather than re-learning
A new statute is enacted mid-deployment	Jurisdiction tuning quarterly review picks it up; DueCare team can produce an interim patch in days if urgent
Inspector loses tablet in the field	Disk encryption prevents data exposure; audit log on the lost device is reconstructed from the central archive's last sync; replacement device re-issued from spares