Defensible AI research — built for universities that have to stand behind the result.
🌐 axiovexsystems.com · Theory sibling: AXIOVEX/epistemic-ledger · Record layer: AXIOVEX/axiomere
Tekmerion (τεκμήριον — conclusive evidence by which a thing is proved; renamed from the Axiovex Research Engine, 2026-10-11, DN-006) is Axiovex's defensible-research framework. Its records live in Axiomere, a truth ledger where records are revised by new snapshots and never deleted.
Most AI research tools sell speed and summaries. This one sells something different: an audit trail a university can rely on. Every claim the engine makes carries the evidence it rests on and the test that would falsify it. Verdicts are registered before outcomes are seen. Records are append-only — a correction is a new entry, never a quiet rewrite of the old one. When the engine cannot support a claim, it says so, in a verdict designed for exactly that: inconclusive is an answer, not a failure.
AI-assisted research is fast, fluent — and hard to defend. Citations are generated as easily as they are retrieved, and some are fabricated outright. Methods shift mid-run without a record. A summary arrives with no way to ask what does this rest on, what would change it, and who checked? For a student paper that is an inconvenience. For a research integrity office, a systematic-review team, or a lab whose work will be replicated, challenged, or published, it is disqualifying.
The missing piece is not a better summary. It is a research process whose every step leaves evidence.
- Register the question before answering it. Research questions, inclusion criteria, and verdict rules are frozen and timestamped before evidence is gathered — the same discipline as a pre-registered study.
- Retrieve from scholarly infrastructure, with provenance. Sources are real scholarly systems (Crossref, OpenAlex, CORE — the three connectors that have shipped; see the source registry), every query is logged, and every retrieved item carries its origin. Nothing enters the record without a source that can be re-checked. (This list earlier named Semantic Scholar as well; no Semantic Scholar connector has shipped, and the list was corrected to the shipped three on 2026-10-13.)
- Build a claim register, not a prose answer. Findings are decomposed into claims with stable IDs, each carrying its evidence, its dependencies, and an explicit falsifier — the observation that would overturn it.
- Count evidence honestly. Derivative sources that repeat one origin are counted once, by origin group. Ten echoes of a single study are one piece of evidence, not ten.
- Render verdicts that can fail. Each claim resolves to a verdict from a fixed taxonomy — Supported, Contradicted, Inconclusive, or Invalid (the run itself was defective). Controls are designed so they can fail: a check that cannot differentially fail marks the run invalid rather than passing it.
- Keep the record append-only and hash-verified. The full history of what was believed, when, and why stays queryable. Exports are dossiers and audit bundles a university can keep, deposit (OSF, Zenodo), and hand to a reviewer or an integrity board.
- Frame. State the question, the claims it implies, and what would falsify each one.
- Freeze. Register the design — sources, criteria, verdict rules — before outcomes are seen.
- Gather. Retrieve from scholarly sources; log every query; deduplicate by DOI and identity; attach provenance to every item.
- Weigh. Group evidence by origin, test it against the falsifiers, run the controls.
- Decide. Issue verdicts claim by claim, with intervals and margins — never a single blended score that hides a weak link.
- Deposit. Export the registration, the claim register, and the audit bundle. The record stands on its own, without the engine.
The method is not new decoration. It is the epistemic engineering framework Axiovex has developed and measured in the open — claim registers with falsifiers, origin-group evidence counting, discriminating controls, freeze-and-replay integrity, and an append-only ledger whose scoring machinery is published, with its negative results, in the theory sibling repository, AXIOVEX/epistemic-ledger. This repository is the tool built on that theory.
- Research integrity offices that must inspect how an AI-assisted conclusion was reached.
- Systematic-review and meta-analysis teams that need reproducible searches and auditable screening.
- Graduate research programs teaching — and enforcing — defensible AI use.
- Libraries and labs that need research records they can keep, deposit, and re-run.
Built (October 2026). What follows is the state of the work as the records support it; the specifications under specs/ carry the claim-by-claim evidence. Release history: CHANGELOG.md.
- The engine (Specs 001–002) is complete through a pre-registered evaluation (Phase 7): questions, verdict rules, and predictions were frozen before the run, all five verdicts matched the frozen predictions, and every threshold set in advance was met. The evaluation is deposited at Zenodo (doi:10.5281/zenodo.23319126). The earlier Phase 5 demonstration run was INVALID — its swap control could not fail — and is recorded as such, not quietly dropped (doi:10.5281/zenodo.23311479). Spec 001 was re-assessed on 2026-10-12; three of its seven claims still sit below the evidence bar, each for a stated reason in its record.
- The study interface (Phase 6) is built: a server-rendered web app over the same services, implemented from owner-approved wireframes.
- The mathematics layer (Spec 003) is complete, with a worked study of the 2026 AI-mathematics claims: five verdicts — one supported, two contested, two inconclusive — deposited at Zenodo (doi:10.5281/zenodo.23317259).
- Domain adapters (Specs 006–008) connect independent checkers — crystallography validation, reproducibility certification, and statistical-consistency checking — each with a registered statement of what its check cannot establish. Two of the three checkers could not be run live in this environment; the records say so plainly.
- The enterprise platform (Spec 005) is built: sign-in by OIDC and SAML 2.0, SCIM provisioning, multi-tenant isolation, break-glass emergency access, compliance evidence (including a CMMC responsibility matrix), and the admin console. Thirty of its 31 claims are supported; the one open claim is a regulatory question about the hosted shape (DFARS FedRAMP equivalency), left open deliberately pending a legal decision.
- Agent and registration surfaces (Specs 009–012) are built: an MCP server and agent skill that run studies end to end (29 tools), OSF registration of study records (public registration doi:10.17605/OSF.IO/84DK7), completed evaluation machinery, and a third scholarly connector (CORE, alongside Crossref and OpenAlex).
- A partner pilot is designed and its candidate shortlist approved; no partner has been contacted yet.
Under the project's stage model (DN-010), the engineering claims stand at checked — verified, not settled. Acceptance is recorded only where a named human made the decision (the owner's gate ratifications, and one mathematics claim that named humans also explained); nothing is presented as digested. The founding decisions are recorded in Design Note 001, the architecture in docs/architecture/, and the governing specifications in specs/ under the project constitution. Full documentation lives at tekmerion.readthedocs.io. A status in this repository is a claim, and claims here carry evidence. (This section's earlier text — "implemented through Phase 5… Phase 6 awaits owner approval" — went stale as the work completed; corrected 2026-10-12. The correction before it is recorded in the stage-model audit, docs/audits/2026-10-12-stage-model-audit.md.)
.
├── README.md # This page
├── LICENSE # All rights reserved, © 2026 Axiovex Systems, LLC
├── CHANGELOG.md # Release history (Semantic Versioning)
├── .specify/memory/ # Project constitution
├── specs/ # Governing specifications, each with spec, plan, tasks, AEE claims
│ ├── 001-research-engine/ # The engine: claims, falsifiers, verdicts
│ ├── 002-mvp-implementation/
│ ├── 003-math-claims/
│ ├── 004-lean-verification-service/
│ ├── 005-enterprise-platform/
│ ├── 006-checkcif-adapter/ # Domain adapters (006–008)
│ ├── 007-codecheck-adapter/
│ ├── 008-statcheck-grim-adapter/
│ ├── 009-mcp-server/ # MCP server + agent skill
│ ├── 010-osf-registration/
│ ├── 011-evaluation-machinery/
│ ├── 012-core-connector/
│ └── 013-red-team-remediation/
├── src/tekmerion/ # The engine, platform, console, and MCP server
├── tests/ # Test suite
├── skill/ # The agent skill for the MCP server
├── documentation/ # ReadTheDocs sources (tekmerion.readthedocs.io)
├── frozen-artifacts/ # Frozen deposited study bundles — records, do not edit
└── docs/
├── architecture/ # The six architecture workstreams
├── design-notes/ # Numbered, append-only design decisions
├── audits/ # Dated audit reports
├── pilot/ # Pilot design, shortlist, and readiness packet
├── surveys/ # Domain surveys
├── wireframes/ # Owner-approved UI wireframes
└── install.md # Install + MCP quickstart
All rights reserved. © 2026 Axiovex Systems, LLC. See LICENSE.
Claims with falsifiers. Verdicts with receipts.