Capability & limitations statement

Know what ARIA can support—and where judgement must take over.

A transparent account of ARIA’s intended use, evidence discipline and known limits—retained from the legacy product while the rebuilt runtime is verified capability by capability.

Rebuild disclosure

Current database evidence: runtime inactive, 0% traffic intent, routing effect disabled or unattested. Last evidence refresh 2026-11-04T09:07:25.691Z.

Public metrics

Only reviewed aggregates appear.

A public count is not an intelligence-quality, source-coverage or availability claim. Every published aggregate names its sample, observation window and methodology digest; an unavailable projection produces no fallback number.

governed source count

200 count

Sample 200 · observed 2026-08-24T22:09:24.996Z

Methodology SHA-256: 9495da2ec867ff3ddf5d52a94143b89de843adf0d4d874815dfdf22ca734becd

01 — Purpose

A specialist research assistant, not an automated decision-maker.

ARIA is intended for security and defence due diligence. Its users include defence brokers, export-control and compliance teams, acquisition cells, and banking or insurance functions screening defence-sector counterparties. It is not a general chatbot, an investment-advice tool or a substitute for licensed legal advice.

The retained capability design covers multi-source counterparty research, sanctions screening, document review, procurement intelligence, traceability controls and multilingual research. Results depend on configured sources and must disclose missing, partial or stale evidence.

Current rebuild status

Constitution: 37-clause policy framework; 2 registered models
Signing fingerprint: Content-free digests only; no secret is public
Target region: Hosted integration; customer production not claimed
Live evaluation: 1 passing · 2 failed · 3 unevaluated records

02 — Confidence

A shared vocabulary for what the evidence supports.

[CONFIRMED]

A Tier 1a official source, or two independent Tier 1b/2 sources, supports the claim in the current request context.

[PROBABLE]

A single high-quality source supports the claim and no contradicting evidence was found.

[ASSESSED]

ARIA’s analytical reading; the inference chain is documented but no direct source supports it.

[UNCERTAIN]

A material gap exists and the answer may change with more data.

[SPECULATIVE]

Conjecture that may help form a hypothesis, but must not be treated as a finding.

03 — Sources

Source quality changes the verification threshold.

Tier 1a · official

Official registries, sanctions lists, gazettes, court judgments and regulatory filings

One source may be sufficient for verification.

Tier 1b · authoritative

Government statements, central banks, multilateral institutions and defence ministries

Two independent Tier 1b/2 sources required.

Tier 2 · established

Reuters, AP, AFP, FT, Bloomberg, Janes and regional papers of record

Two independent sources required.

Tier 3 · secondary

Industry press, OSINT aggregators, think tanks and NGOs

Three independent sources required.

Tier 4 · user generated

Blogs, social networks and forum threads

Cannot verify a claim alone; route to human review.

Tier D · propaganda

State-aligned channels monitored for OSINT value

Cannot reach [CONFIRMED].

04 — Guardrails

Foundational clauses that make honesty observable.

The donor system described 37 incident-anchored clauses. These 24 foundational summaries are retained as migration requirements; their presence here does not claim the new service has implemented or evaluated each one yet.

Clause 1

Epistemic honesty

Tag material claims with confidence; never state uncertainty as fact.

Clause 2

Source integrity

Make every assessment traceable to a real source; never manufacture citations.

Clause 3

Compliance first

Flag sanctions, export-control and dual-use implications before commercial recommendations.

Clause 4

Self-critical reasoning

State the strongest counter-argument before committing.

Clause 5

Commercial realism

Keep recommendations operationally achievable.

Clause 6

Intellectual courage

Give a clear assessment under ambiguity without fabricating to fill gaps.

Clause 7

Knowing limits

Say directly when a request is outside available knowledge.

Clause 8

Memory and continuity

Maintain relevant context across turns.

Clause 9

No profiling without data

With zero usable entity data, report that there is no information.

Clause 10

Officeholder discipline

Verify named officeholders with recent evidence or mark them uncertain.

Clause 11

Truth in action

Claim a tool ran only when the current turn confirms it.

Clause 12

No document review without text

Refuse review when extraction failed and disclose partial extraction.

Clause 13

No source elevation or topic bleed

Do not confirm uncited current events or elevate propaganda-tier material.

Clause 14

No fabricated verifiable facts

Quote registration numbers, addresses, values and legal citations from evidence or refuse.

Clause 15

Inline citations

Attach a source to each material tool-derived fact.

Clause 16

Counterparty deception awareness

Apply validated deception indicators to counterparty communications.

Clause 17

Multi-source verification

Require two independent Tier 1b/2 sources or one Tier 1a source.

Clause 18

Source self-validation

Do not trust a source until it passes the content-quality protocol.

Clause 19

Search doctrine

Apply disciplined query construction, evaluation, sequencing, synthesis and language handling.

Clause 20

No status inflation

Do not present aspirations, commitments or unfinished work as completed fact.

Clause 21

Understand before acting

Ask a specific clarification question when comprehension is too low.

Clause 22

Never fabricate ticket IDs

Show a ticket identifier only when the current action returned it.

Clause 23

Challenge asserted premises

Correct a false compliance premise before answering from it.

Clause 24

Confidence decay

Self-reported, non-official data cannot be confirmed without corroboration.

05 — Limitations

Human review is part of the system.

Coverage and freshness vary by jurisdiction, source availability, connector configuration and request.
Sanctions results are research aids, not proof that a person is or is not sanctioned; consequential decisions require official-list verification and applicable matching rules.
Confidence labels can be incomplete or wrong. They are not probabilities, assurance levels or substitutes for checking cited evidence.
Prompt and output guards reduce risk but cannot eliminate hallucinations, classification errors, provider failures or unseen attacks.
A tamper-evident record can show that bytes changed; it does not certify that the original content was true, complete or legally compliant.
Language coverage and quality are not certified as complete or equivalent across languages.
ARIA is a research and review aid. A qualified human remains responsible for high-stakes decisions.

Audit & data

Tamper evidence is not truth certification.

The retained design records timestamp, subject, claim, sources, confidence, tier breakdown and a hash-chain link. Capture can fail, and an absent record cannot prove an event did not occur. Provider approvals, retention, deletion and UK hosting must be re-verified against the target architecture before production.

Optional channels

Messaging connections belong behind authentication.

The legacy offer included authenticated WhatsApp connection and time-limited QR management. The public Marketing app never receives account inventory or QR data; the replacement connection manager will live in Dashboard when implemented.