Giblets Forge · Epistemic Infrastructure
AXIOM Know what you know.

You don't give a new employee the entire company archive.
You give them what they need to do their job today.

Most organisations don't have an intelligence problem.
They have a data trust problem.
They don't need access to everything, everywhere, all at once.
They need validated, role-specific, trustworthy knowledge —
exactly when the job requires it.

That is what AXIOM delivers.

01 — The Argument

The industry asked
the wrong question.

The question was: how do we make AI more intelligent? Billions were invested. Models grew. Benchmarks improved. And organisations deployed them into healthcare, legal, compliance, and financial contexts — domains where a confident wrong answer is worse than an honest uncertain one.

The question that should have been asked first: how do we know if we can trust what the system says?

Large language models generate fluent, authoritative output regardless of whether the underlying knowledge is sound, current, or grounded in anything verifiable. Confidence is not knowledge. Fluency is not truth. The foundations were poured without epistemic integrity built in — and no amount of patching fixes a foundation.

AXIOM does not patch the problem. It reframes the question entirely. Most of the time, organisations do not need artificial intelligence. They need trustworthy access to what they already know — structured, validated, weighted by reliability, and presented clearly.

That is a data infrastructure problem. AXIOM solves it.

02 — Architecture

Three components.
One guarantee.

AXIOM is not a monolithic AI system. It is three precisely scoped components, each doing exactly one thing, with trust flowing in one direction only — from ingestion through justification to presentation. Nothing flows backwards. The presentation layer cannot contaminate the knowledge store.

COMPONENT 01
AXIOM Ingest
The Gate
Three defined pathways. Structured sources — regulatory documents, standards, database exports — parsed by deterministic extractors with defined schemas. Semi-structured sources — templated reports, formatted outputs — processed via Schema-Driven Abstract Syntax Tree (AST) generation. Unstructured prose passes through an Augmented Review Interface: a heuristic token parser identifies candidate entity-relation triplets and presents them to a human reviewer as binary Accept/Reject decisions. The human is the actuary-gate, not the author — accelerating structured ingestion at scale without introducing automated interpretation. To prevent cognitive fatigue and blind-click contamination, the interface injects deterministic control traps — deliberately flawed synthetic claims with known contradictions — into the review queue at randomised intervals. If a reviewer accepts a control trap, the session is automatically flagged for audit. The human gates are themselves audited deterministically. This shifts liability from an untraceable systemic ghost to classical professional accountability — the cryptographic audit trail identifies exactly which credentialed operator approved each claim. Bad data has a name attached. That is not a risk. That is proper data governance. Local storage commits capsules via a Log-Structured Merge-Tree or in-memory ring buffer flushing in batched, sequentially aligned blocks — preventing flash write-amplification degradation on constrained mobile hardware. Interpretation never happens inside the system. For high-consequence domains, human gating is not a limitation — it is the correct design.
COMPONENT 02
AXIOM Justify
The Engine
Not a model. A deterministic, rule-based scoring system. Source reputation is a lookup. Temporal validity is date logic. Leighton Weight is a formula. Every decision is the result of explicit, auditable rules — not inference. It behaves identically every single time, and you can read the code and know exactly why it made the decision it made. The intellectual core of AXIOM.
COMPONENT 03
AXIOM Present
The Voice
An Extractive Semantic Translation Layer — a small language model whose sole function is formatting deterministic data into human-readable sentences, not intelligence. Operating under strict constrained decoding with Extractive Slot-Filling, key domain nouns, metrics, and values are selected directly from a token lookup map of validated source text. Generative sampling applies exclusively to connective prose. Retrieval uses a Dual-Key Filter: semantic vector search generates a candidate pool, then the rule engine applies a hard boolean namespace mask over those candidates against the active domain taxonomy before tokens reach the slot-filling engine. Colloquial or non-standard queries cannot smuggle non-compliant nodes through — if a candidate fails the namespace mask, it is dropped. Sub-graph retrieval enforces hard truncation boundaries — top-k ranked claims only, historical variants linked via metadata pointers. The model is quantized to INT4 or INT8 via GGML or ONNX Runtime — stable on legacy hardware without NPU acceleration. If a query yields no documents above trust threshold: {"status": "insufficient_trusted_data"}. The hallucination risk is not mitigated — it is structurally eliminated.
→
Input
Raw Sources
Documents, databases, feeds, literature
→
Ingest
Structured Data
Normalised, formatted, ready to score
→
Justify
Scored + Weighted
Six lenses, Leighton weight, justified
→
Store
Trusted Knowledge
Immutable, provenance-rich, auditable
Present
Human Interface
Natural language. Grounded. Traceable.
03 — The Justification Engine

No model.
No epistemic blindness.

The critical insight: if you use a language model to score and weight your data, you have reintroduced the exact problem you were solving. A model making judgements about trustworthiness has the same fundamental blindness as the models it's supposed to be checking. A patch on a patch.

AXIOM Justify is not a model. It is a deterministic, rule-based system — mature, proven technology that the industry abandoned in the rush toward neural networks, and that turns out to be exactly the right tool for this job.

A rule-based system behaves exactly the same way every single time. You can read the code and know precisely why it made the decision it made. That is not a limitation. That is the feature. No LLM can make that guarantee.

Each scoring dimension in AXIOM Justify is implemented as explicit, auditable logic:

AXIOM Justify — Deterministic Scoring Rules
source_groundingLookup table. Known authoritative sources carry pre-assigned trust values. Unknown sources default to minimum. No inference.
temporal_validityDate calculation. Ingestion timestamp versus domain-specific decay threshold. Regulatory guidance decays faster than established science. Pure logic.
domain_calibrationDeterministic Namespace Matching against Domain Taxonomy. High-consequence domains apply tighter confidence bounds automatically. No statistical classifier — pure structural lookup.
internal_consistencyGraph comparison. New claim checked against existing store for contradiction. No model required — structural comparison.
claim_granularityFormal parsing. Claim structure evaluated against precision rules. Vague language penalised. Specific claims rewarded.
leighton_weightLinearly Scaled Epistemic Vector Calculation. An objective coordinate derived from the six discrete scoring dimensions. Not a heuristic estimate — a reproducible mathematical result. Identical inputs always produce identical outputs.
justification outputEvery decision logged with the exact rule that produced it. Full provenance. Human-readable. Regulator-ready.

A compliance officer can stand in front of a regulator and explain every decision AXIOM Justify made — because the decisions are code, not inference. That is architecturally impossible with any language model.

Rule-based systems are not glamorous. They do not get venture funding. They do not make headlines. But they do something no LLM can do: they are incapable of the problem they are designed to prevent.

On scale: Consistency checking is scoped to domain sub-graphs, not the entire knowledge store. A new clinical claim is checked against the clinical sub-graph only. A legal claim against the legal sub-graph. Cross-domain contradiction checks are triggered only when explicitly flagged at ingestion. This keeps the system performant on constrained hardware as the store grows — consistency is always a local operation, never a global one.

On taxonomy evolution: Taxonomies are stable but not static — ICD-11 codes change, compliance structures reorganise, legal jurisdictions split. AXIOM treats the domain taxonomy itself as a versioned sub-graph in the immutable store. When a taxonomy updates, it generates a new namespace ruleset with a cryptographic hash. Older capsules remain validated against the exact taxonomy version active when they were signed. No silent schema migrations. No retroactive revalidation. Full version provenance preserved.

On rule integrity: As domain calibrations and temporal decay variables accumulate, rules can interact in unexpected combinatorial ways — creating accidental dead zones where valid data receives a zero score due to nested logic conflicts. AXIOM Justify addresses this via Regression Integration Testing and Formal Property Verification. Synthetic claim matrices are run against the rules engine to prove scoring boundaries are mathematically stable across all permutations. The rules are not assumed to be correct. They are verified to be correct.

On versioned knowledge: Regulated knowledge doesn't just expire — it mutates. When a newer, higher-weighted authoritative entry supersedes an older node, AXIOM Justify applies a deterministic decay modifier to the prior node and creates an explicit supersedes edge in the sub-graph. Both nodes remain in the store — immutability is preserved — but the query layer surfaces them with appropriate weights and version context. A clinician asking about a procedure sees the current guidance weighted highest, with prior versions visible and clearly marked as superseded. Nothing is silently deleted. Everything is traceable.

04 — The Distinction

Not AI.
Epistemic infrastructure.

This distinction matters commercially and technically. The market is saturated with AI products promising intelligence. AXIOM makes a different promise — one that is more valuable in regulated, high-stakes, and compliance-critical environments.

Capability
Standard LLM Deployment
AXIOM
Source provenance
None — training data opaque
Full — every claim traced to source
Confidence grounding
Fluency-based, unreliable
Six-lens scored, justified
Temporal validity
Unknown — training cutoff opaque
Flagged at ingestion, re-validation scheduled
Audit trail
None — outputs not traceable
SCP capsule per decision, causally ordered
Hallucination risk
Inherent — architectural
Eliminated at store level — present layer reads only validated data
Scoring mechanism
Model inference — non-deterministic, unexplainable
Deterministic rules — identical result every time, fully explainable
Regulatory readiness
Requires extensive human review overlay
Provenance-native — audit-ready by design
05 — Applications

Anywhere trust
is not optional.

AXIOM is domain-agnostic at the infrastructure level. The Justification Engine is the constant. Ingest and Present layers are configured per domain. The knowledge store is domain-specific. The guarantee is universal.

01

Healthcare Compliance

Clinical guidance, regulatory documents, trial data — ingested, scored, weighted by authority level. Clinicians query in natural language. Every response carries provenance. Decisions are auditable. Justitia: sovereign healthcare communications infrastructure built on AXIOM foundations.

02

Medical Communications

Evidence synthesis at scale — literature ingested, claims validated against source, conflicting evidence surfaced automatically, currency flagged. Writers and scientists access trusted knowledge without manual review overhead. The data was always there. AXIOM makes it trustworthy.

03

Legal and Regulatory

Statute, case law, guidance — ingested with temporal validity scoring. Conflicting provisions automatically flagged. Weight assigned by jurisdictional authority. A lawyer queries; AXIOM returns what is known, how reliably, and when it was last verified.

04

Enterprise Knowledge Management

Internal documents, policies, procedures — the knowledge organisations have but cannot reliably access or trust. AXIOM turns fragmented institutional knowledge into a validated, queryable, auditable asset rather than a liability.

06 — Provenance

Built before
the question was mainstream.

AXIOM is not a response to current AI hype. It is the result of independently arriving at a diagnosis of what was wrong with the foundations — and building toward a solution — before the industry had named the problem.

Independent Convergence

The Semantic Capsule Protocol — AXIOM's audit artefact format — was independently designed and later found to converge with the Actors Model in distributed systems theory. SCP capsules operate exactly as actor messages: immutable, asynchronously append-only, causally chained via cryptographic hashes of the preceding capsule, and entirely responsible for their own internal boundary validation. The alignment is not superficial. The architecture was correct before the theory was known.

First Principles Thinking

Every component of AXIOM was reasoned from the problem outward — not assembled from existing tooling. The result is an architecture where every design decision has a clear epistemic justification. The same principle the Justification Engine applies to data, the architecture applies to itself.

Working Components

SCRIBE, LENS, DataCube, Leighton Weight, and the SCP capsule format are all built and operational as independent systems. AXIOM is the architecture that unifies them into a coherent, deployable whole. This is proof of concept, not proposal.

The Right Constraint

Engineered under strict resource constraints — operational within a localised Android/Termux environment — to prove that epistemic integrity requires architectural discipline, not massive server farms. If AXIOM runs correctly on a phone, it runs correctly anywhere. The constraint was not a limitation. It was the test.

07 — Sovereignty

Runs on yours.
Full stop.

Because the presentation layer is a tiny model and the intelligence lives in the validated store — not in a cloud API — AXIOM is sovereign by design. Everything runs on your own infrastructure. Nothing leaves the building.

No API Dependency
No OpenAI. No Anthropic. No per-token billing. No dependency on a third party staying solvent, keeping your data private, or maintaining their terms of service. AXIOM runs on hardware you own.
Data Stays Yours
Sensitive clinical data, legal documents, internal policy — none of it leaves your environment. AXIOM is built for organisations where data sovereignty is not a preference, it is a legal and regulatory requirement.
Runs On Device
A presentation model runs on a phone, a ward tablet, an air-gapped workstation. Edge devices operate as read-only replicas at query time. Uploads are cryptographically signed capsule transactions proposed to an authoritative upstream sync node via an Event-Sourced Cryptographic Ledger — an append-only DAG ensuring every state transition is immutably recorded and causally ordered. Conflict resolution follows a CRDT model — no central monolithic arbiter required. Cold-start recovery uses Sub-Graph Snapshots: pre-calculated, cryptographically verified sub-graph states at specific block heights. Reconnecting devices download the latest snapshot and replay only marginal delta capsules after the snapshot timestamp. The full ledger is never replayed from day one. Key security is managed via an immutable fast-path Revocation List on the upstream sync node — if an edge device key is compromised, the node immediately rejects all capsule transactions signed by that key post-timestamp and broadcasts a lightweight revocation token down the DAG to isolate the actor instantly. Sync is always bounded, safe, and cryptographically accountable.

The industry built AI that requires you to send your most sensitive knowledge to someone else's computer to get an answer. AXIOM inverts that entirely. The knowledge stays sovereign. The model is small enough to live where the work happens. The intelligence was always in your data — AXIOM simply makes it trustworthy and accessible.

AXIOM ensures that everything your organisation knows, it can prove it knows. Not because an AI told it so — because the knowledge was validated, weighted, justified, and stored before anyone asked the question.

This architecture was stress-tested using a general LLM as an external review tool. The LLM was not in the pipeline. It never is. · Full technical architecture and implementation detail available on request. · Giblets Forge Ltd · Upper Heyford, Oxfordshire