Why arche, and when to use it¶

A practical guide for developers, researchers, DPOs, and civil society choosing between arche-core and the other tools in the PII landscape. Honest about what arche is, what it isn't, and where the alternatives are the better pick.

Status: pre-beta (development) - not for production use yet

Suitable today for research, prototyping, evaluation, benchmarking, and contributing.

The setup¶

You're building something that handles African data - a Nigerian fintech onboarding flow, a Kenyan health-tech intake screen, a journalist's PII scanner for leaked emails, a South African civil-society audit of a public dataset. You need to detect PII. You install Microsoft Presidio (the obvious choice). You feed it a sample customer record:

Customer Fatima Abdullahi, NIN 12345678901, BVN 22156789012, phone 0803 555 7890, RC 245678.

Presidio confidently returns:

NIN 12345678901  -> US_BANK_NUMBER  (wrong)
BVN 22156789012  -> US_BANK_NUMBER  (wrong)
0803 555 7890   -> US_PHONE_NUMBER  (wrong; it is a Nigerian network)
RC 245678        -> not detected

That's not a typo. Presidio's default recognizers literally label Nigerian customers as having US bank accounts. Same story for Ghana Card (gets called US_PASSPORT), South African ID, Kenyan KRA PIN. The Presidio team isn't wrong - they built recognizers for the data their users had. Your data isn't their data.

Same input, arche-core:

from arche import Pipeline

result = Pipeline(jurisdiction="NG").process(
    "Customer Fatima Abdullahi, NIN 12345678901, BVN 22156789012, "
    "phone 0803 555 7890, RC 245678."
)

PII-1-NAME      "Fatima Abdullahi"  tier=moderate  NDPA-2023 s.30
PII-2-NIN       "12345678901"       tier=high      NDPA-2023 s.30, NIMC Act s.27
PII-2-BVN       "22156789012"       tier=high      NDPA-2023 s.30, CBN BVN policy 2014
PII-3-PHONE     "0803 555 7890"     tier=moderate  NDPA-2023 s.26
PII-2-RC        "245678"            tier=low       NDPA-2023 s.31 (legitimate interests)

Every detection: validated checksum, the statute section that classifies it, the sensitivity tier, the policy action that follows. Change jurisdiction="NG" to "ZA" and you get POPIA. "KE" for Kenya DPA. "GH" for Ghana DPA. Same code; four jurisdictions.

That's the gap in one example. The rest of this page is the proof, the personas, and the honest limitations.

The proof - cross-tool benchmark¶

From benchmarks/lingua-africa-eval-v0.2: 48 synthetic test cases across six anchor languages (English, Nigerian Pidgin, Yoruba, Hausa, Swahili, Amharic) covering NG / KE / ZA / GH government identifiers, IPs, DIDs, and crypto wallets. CC-BY-4.0. Reproducible from pip install arche-core[presidio] plus one script.

Tool	Cases matching expected categories	Notes
arche-core v0.2.0a3	47 / 48	All African IDs detected across all six anchor languages. Luhn validators correctly reject negative-control invalid SA ID. Cross-cutting detectors (IP, DID, crypto wallet) work irrespective of surrounding script. The one gap is PVC (NG voter card) detection - documented post-beta work.
Microsoft Presidio (default, generous "any non-empty" scoring)	37 / 48	On generous scoring, passes negative tests and IP / crypto cases. On per-category matching the picture inverts: Presidio scores 2 / 25 on cases where the expected label is a specific PII-2 / PII-5 / PII-8 category - and confidently mislabels NG NIN / BVN, SA ID, and Ghana Card as US-default types. Even Amharic text with an embedded NG NIN gets labeled `US_BANK_NUMBER`.
GLiNER2-PII (Fastino)	Not run standalone	Install pulls ~2.5 GB of torch + transformers. Public schema covers 42 PII categories - none are African government IDs. Composed via `arche-core[detect]` extra for soft-PII coverage.
OpenAI Privacy Filter	Not run	API-only. No published per-category breakdown. Closed-weight model - auditability question open.

The headline: Presidio's default recognizers actively harm African DPI compliance work by labelling Nigerian customers as having US bank accounts. arche-core ships the African recognizers that close this gap - and the regexes that drive them are script-agnostic for ASCII identifier shapes, so Yoruba / Hausa / Amharic narratives wrapped around an NIN or KRA PIN still detect correctly.

Per-language breakdown (arche, 48 total cases):

Language	Cases	arche
English	8	8 / 8
Nigerian Pidgin	8	8 / 8
Yoruba	8	8 / 8
Hausa	8	7 / 8 (PVC post-beta)
Swahili	8	8 / 8
Amharic	8	8 / 8

A v0.1 baseline (15 cases, English-only) is preserved as lingua-africa-eval-v0.1 for historical comparison. Future benchmark expansions will be published as separate, versioned releases.

Try it before you read the rest¶

The fastest way to internalise the difference is to paste your own text into the live demo. Pick a jurisdiction, paste any text (synthetic only - never real PII), see the detection set with tier and citation, see the policy decisions, see the PII-free audit log, walk away with a JWS receipt that verifies offline. Four preloaded samples cover NG / ZA / KE / GH so you can see the surface in 30 seconds.

The demo runs on Streamlit Cloud and proxies the same pip install arche-core you'd install locally. No telemetry, no PII storage. Source: demo/app.py.

What arche actually does that nothing else does¶

Three concrete gaps no other tool in the PII landscape closes:

1. Per-country African identifier coverage with check-digit validation¶

The Pan-African PII Taxonomy v0.1 covers 51 categories across NDPA-2023 (Nigeria), POPIA (South Africa), Kenya DPA, and Ghana DPA. The shipped detectors include NIN, BVN, RC, TIN, voter PVC, driver's licence (NG); national ID, KRA PIN, NHIF (KE); SA ID with full Luhn + structural decode, tax reference, passport (ZA); Ghana Card, SSNIT, TIN (GH) - plus 11 non-launch African country patterns (RW, TZ, UG, ET, CI, SN, CM, EG, MA, AO, MZ). Every detector validates check-digits where the underlying spec supports it; structural validators drop false positives at detection time, before policy applies.

Presidio's default recognizers ship none of these. You'd write each one yourself, maintain the regex against ID-scheme amendments, and integrate with your audit layer. Multiply by four jurisdictions and you have a quarter of an engineer's time wired to a problem that arche solves at pip install time.

2. Sensitivity tier and statute citation on every Detection¶

Detection is one floor. Every Detection arche emits carries sensitivity_tier (HIGH / MODERATE / LOW per the loaded statute) and regulatory_citation (the exact statute section) - populated from the statute YAML at detection time, before the policy engine fires. Tier-aware dashboards, per-citation compliance reports, and HIGH-tier routing become properties of the data structure, not a downstream join you have to build.

The statute YAMLs live in arche/policy/statutes/: NDPA-2023.yaml at v1.0 with Section 24 / Section 26 / Section 29 / Section 34-38 cited inline; POPIA.yaml, KENYA-DPA.yaml, GHANA-DPA.yaml are under review. Statute amendments are YAML changes, not code changes - you can read them, audit them, fork them, version them in git.

from arche.policy import load_statute

statute = load_statute("NDPA-2023")
print(statute.categories["PII-2-NIN"].citation)
# "NDPA-2023 s.30, NIMC Act s.27"

No other open-source PII library produces this - here is the rule the redaction enforced - as a property of every detection. Combined with the audit log below it gives you regulator-ready evidence at the SDK level.

3. Cultural naming intelligence - 114 equivalence groups¶

from arche.detect._names.lexicon import are_names_equivalent

are_names_equivalent("Adeyemi", "Adeyemi")
# (True, 0.96) - Yoruba tonal mark equivalence

are_names_equivalent("Mamadou Diallo", "Mohamed Diallo")
# (True, 0.92) - Fulani / Arabic Pan-Islamic equivalence

are_names_equivalent("Chukwuemeka Okafor", "Emeka Okafor")
# (True, 0.95) - Igbo prefix-elision

are_names_equivalent("Fatima Abdullahi", "Fatoumata Abdoulaye")
# (True, 0.89) - Hausa / Wolof cognates

114 equivalence groups, 454 name variants, 50+ ethnic traditions. Tested against a Jaro-Winkler baseline:

Metric	Jaro-Winkler	arche	Delta
F1	0.849	0.988	+16.3%
Precision	1.000	1.000	0.0%
Recall	0.738	0.976	+23.8%

Zero false positives - the lexicon is conservative by design. arche never claims equivalence between Mamadou and Mary; only between genuinely co-referential variants documented by African linguists.

What else ships in the package (power-user)¶

These features ship today and are fully tested, but they're not the lead pitch. Read the row that matches your use case; skip the rest.

Feature	Module	One-line	When you'd use it
Sign, share, extract	`arche.sign`, `arche.credentials.sd_jwt`	Ed25519 + JWS signing of `Pipeline.Result`; SD-JWT-VC re-framing for wallet interop. Offline verification with `did:key`.	Compliance officer needs a regulator-ready signed audit bundle. DSAR response that crosses an organizational trust boundary. KYC attestation a wallet can carry.
Citizen-side DSAR	`arche.workflow.dsar`	Draft-only DSAR letter generation citing NDPA s.34 / POPIA s.23 / Kenya DPA s.26 / Ghana DPA s.35. Per-jurisdiction.	Civil-society org training citizens to exercise rights. Journalist filing a DSAR as part of an investigation.
Entity resolution	`arche.match`, `arche.link`, `arche.resolve`	Lightweight Fellegi-Sunter matcher with jurisdiction-specific priors. African-name-aware comparator functions.	Deduplicate up to ~100K records on a laptop. For billion-row scale, install `arche-core[resolve]` and feed Splink.
Places resolution	`arche.resolve_places`, `arche.list_places`	Jurisdictional place lookup with verifiable JWS audit receipts. UK first; African expansion is roadmap.	"Find me an NHS dentist near SW1" with a signed receipt for the query, the redaction, and the result.
Append-only audit log	`arche.graph.audit`	SQLite-backed log with markdown compliance reports and JWS-signed regulator exports. PII values never stored.	Standalone audit-log surface for an existing detection pipeline. NDPC quarterly handoff.

Why arche, by persona¶

Developer building African fintech / healthtech / civic platforms¶

You're the primary persona. You're the one who installs arche-core first.

The shape of your problem: you have customer records with NIN / BVN / Ghana Card / SA ID. Your stack is FastAPI / Django / Streamlit. Your auditor will ask "show me the rule that fired." You don't want to write per-country regexes, maintain check-digit validators, hand-roll redaction logic, and build a statute-citation audit layer from scratch. You want pip install and three lines.

from arche import Pipeline

pipeline = Pipeline(jurisdiction="NG")            # auto-loads NDPA-2023
result = pipeline.process(customer_intake_text)
log_to_warehouse(result.redacted_text)            # safe to share
audit.emit_batch(result.audit_entries)            # PII-free audit row

You use arche when:

Your customers carry NIN, BVN, Ghana Card, SA ID, KRA PIN and Presidio's defaults don't help.
You need every detection to cite the specific statute section so when the regulator asks "what rule fired?" the answer is in your audit log, not your head.
You ship records to downstream consumers and want each redacted output cryptographically signed so the consumer trusts it offline.
You want one pip install instead of stitching Presidio + custom recognizers + signing + audit from scratch.

You probably do NOT use arche if your data is purely Western (US / EU) and you have no African footprint - Presidio + custom validators is more direct.

Researcher in African NLP / responsible AI / PII benchmarking¶

from arche import Pipeline

for sample in dataset:
    result = Pipeline(jurisdiction=sample.jurisdiction).process(sample.text)
    yield {
        "text": sample.text,
        "expected": sample.labels,
        "detected": [(d.category, d.span) for d in result.detections],
        "applied_policy": [(o.action, o.statute_section) for o in result.policy_outcomes],
    }

You use arche when:

You're building a Pan-African PII benchmark and want versioned ground-truth labels grounded in the published taxonomy.
You're benchmarking GLiNER2-PII / Presidio / OpenAI Privacy Filter against jurisdiction-aware ground truth and need a baseline that's not Western-default.
You want to ship a citable digital public good alongside your paper - the Pan-African PII Taxonomy v0.1 (CC-BY-4.0).

You probably do NOT use arche if your research target is state-of-the-art neural NER on English / EU corpora - gliner-large-v2.5 or dslim/bert-base-NER is more direct.

Compliance officer / DPO / regulator¶

from arche.graph.audit import AuditLog
from arche.sign import generate_keypair

audit = AuditLog("./compliance.sqlite")
# Pipeline runs daily; every detection writes a PII-free row.

# Quarterly handoff to the regulator
officer_key = generate_keypair()
bundle = audit.export_signed(key=officer_key, purpose="ndpc_quarterly_audit")
# bundle is a JWS-signed regulator report. NDPC verifies offline.

You use arche when:

The regulator wants an append-only audit log that proves PII was never stored - only category labels, character spans, document hashes.
You need JWS-signed export bundles for periodic regulator handoff, verifiable offline by the regulator's tooling.
Your organisation processes multi-jurisdictional data and needs a single policy engine that honours the right statute per record.
You're preparing for an NDPC / Information Regulator / ODPC audit and need a regulator-ready compliance report on demand (audit.compliance_report_markdown()).

You probably do NOT use arche if your compliance regime is purely GDPR and your data has no African footprint - OneTrust / Privitar / commercial suites have richer GDPR-specific tooling.

Journalist / civil society / data subject¶

from arche.workflow import DSARWorkflow, DSARRequestor, DSAROrganization
from arche.sign import generate_keypair

citizen_key = generate_keypair()      # one-time, stored in your wallet
draft = DSARWorkflow(
    jurisdiction="NG",
    requestor=DSARRequestor(name="Adesola Okonkwo", identifier_label="NIN", ...),
    request_type="access",
    targets=[DSAROrganization(name="Sterling Bank", dpo_email="dpo@sterlingbank.ng")],
).run(citizen_key)

# draft.letter_text cites NDPA-2023 Section 34. draft.signed_envelope is JWS-signed.
# Email it. Two weeks later, the DPO is on the clock.

You use arche when:

You're investigating how a Nigerian bank / health system / telco handles personal data and need to file a statute-grounded DSAR.
You're a civil-society organisation training citizens to exercise data protection rights they have on paper but can't operationally use today.
You're a journalist publishing a PII audit of public records, government data, court filings, or the official gazette - and need a reproducible detection set.

You probably do NOT use arche if you just want to scrub PII from a single document - a one-line regex is more direct.

When to use each alternative¶

Use Microsoft Presidio alone when:

Your data is English-language US / EU PII (SSN, credit card, IBAN, US phone, US address).
You're already in the Microsoft ecosystem.
You're willing to write per-country custom recognizers yourself.

Use GLiNER2-PII (Fastino) alone when:

You want a single-model, GPU-friendly NER backbone.
Your PII categories overlap with the 42 GLiNER2 ships.
You can absorb the ~2.5 GB install footprint.

Use Splink alone when:

You have 1M+ rows of clean structured data to dedupe.
Your blocking and comparison logic is purely Western-name-based.
You're comfortable with DuckDB, parameter estimation, and Fellegi-Sunter setup ceremony.

Use OpenAI Privacy Filter alone when:

You're already calling OpenAI and want PII filtering as a side-effect.
Closed-weight, non-auditable detection is acceptable.
Your compliance regime doesn't require a per-decision audit trail.

Use arche-core when:

Your data has any African footprint and you need detection that works out of the box.
You need detection + policy + audit in one library, not three.
You need offline verification of provenance.
You need citizen-side rights tooling (DSAR).
You want one pip install instead of stitching four tools and three custom recognizer sets.

Use arche + Presidio (arche-core[presidio]) when:

Your data is mixed (African + Western), and you want Presidio's Western recognizers alongside arche's African recognizers under one unified taxonomy.

Use arche + GLiNER2-PII (arche-core[detect]) when:

You need multilingual soft-PII coverage (free-form names, occupations) in code-mixed contexts like Nigerian Pidgin.
The deterministic arche recognizers cover the hard structured IDs; GLiNER fills the soft entities.

Use arche + Splink (arche-core[resolve]) when:

You have >100K records to dedupe and want African-name-aware blocking and comparators inside the Splink pipeline.

Honest limitations of arche today (v0.2.0a3)¶

Gap	Today	Where it goes
Full address parsing across all four launch jurisdictions	NG + ZA MVP today	Other jurisdictions are not yet included.
Multilingual soft-PII	GLiNER2-PII via `[detect]` extra	Optional and experimental.
PVC, GhanaPost GPS, M-Pesa references	Not detected	Not currently supported.
PQC signatures	Ed25519 only	Not currently supported.
W3C VC 1.1 JSON-LD emission	SD-JWT-VC only today	Not currently supported.
DSAR organisation-side workflow	Citizen-side only	Not currently supported.
Hash-chained audit log	Schema ready, not populated	Not currently supported.
Statute YAMLs v1.0	NDPA-2023 v1.0; others v0.1 scaffold	Still under review.
MOSIP / OpenCRVS / DHIS2 / OpenG2P production adapters	Not in scope	Use arche outputs at your integration boundary.

Install¶

pip install arche-core                     # Base - every persona above
pip install arche-core[detect]             # + GLiNER2-PII for multilingual soft-PII
pip install arche-core[presidio]           # + Presidio for Western PII overlap
pip install arche-core[resolve]            # + Splink + DuckDB for billion-row dedup
pip install arche-core[doc]                # + docling for PDF / DOCX / PPTX / XLSX
pip install arche-core[all]                # Everything above

from arche import Pipeline
result = Pipeline(jurisdiction="NG").process("your text here")

Why arche, and when to use it¶

The setup¶

The proof - cross-tool benchmark¶

Try it before you read the rest¶

What arche actually does that nothing else does¶

1. Per-country African identifier coverage with check-digit validation¶

2. Sensitivity tier and statute citation on every Detection¶

3. Cultural naming intelligence - 114 equivalence groups¶

What else ships in the package (power-user)¶

Why arche, by persona¶

Developer building African fintech / healthtech / civic platforms¶

Researcher in African NLP / responsible AI / PII benchmarking¶

Compliance officer / DPO / regulator¶

Journalist / civil society / data subject¶

When to use each alternative¶

Honest limitations of arche today (v0.2.0a3)¶

Install¶

See also¶