PRIMAL GOVERNANCE
See what happens the moment an AI has to choose a meaning,
and what changes once that choice is governed.
One question. One real securities‑loan agreement. Two very different answers.
● Securities-Loan-Agreement-With-Set-Off-EN.pdf
A lawyer asks
To a lawyer, the question is obvious. To an AI, the word interest carries least five legitimate legal meanings in this
document’s domain, each one produces a different answer.
What happens without governance
● Raw AI • No governance
Based on the passages, the lender has the following interests:
1. Distributions and related interest: The lender is entitled to all Distributions on the Loaned Securities. Cash Distributions bear interest at the Prime Rate calculated daily until paid (Section 7.2). Shortfalls after default also bear interest at the Prime Rate (Section 13.1(c)).
2. Ownership and control rights over Loaned Securities: The lender is the legal and beneficial owner, retaining voting rights, options, conversion privileges, and other rights (Section 7.6).
3. Collateral-related interests: All Distributions on cash Collateral are for the lender’s account (Section 7.5). Non-cash Collateral and non-cash Distributions are added to Collateral (Section 7.4).
4. Indemnity rights: Indemnification for costs, fees, and expenses on default (Section 13.2).
If you meant “what type of security interest” (e.g., lien, pledge, etc.), the passages provided do not describe the formal characterization of the lender’s security interest in legal terms.
What just happened. The answer looks comprehensive and it cites real sections. But it never decided what “interest” means. It collapsed four different legal senses (economic interest, ownership, interest-rate, indemnity) into one pile, then hedged at the end that it may have missed the sense the lawyer actually wanted. A lawyer reading this would waste time on the wrong provisions and still have to re-ask the real question.
Define · Resolving meaning before reasoning
Primal scores each candidate meaning against the document’s domain, its structural context, and the semantic neighbourhood the question implies. Every sense gets a score. Every score is recorded.
domain · secured transactions
9.79
domain · trust law
8.11
domain · lending
7.30
domain · accounting
6.19
domain · governance
3.11
⚠
Primal calculates a dominance margin across the candidate senses: how far the top sense sits above the runners-up. That margin, per query, decides whether the system commits automatically or escalates for review.
The threshold that governs this decision is a tuneable parameter. Operators dial the level of human oversight per document, user, or risk profile:
AUTO Commit silently when the margin is clear
ASSIST Flag borderline calls for confirmation ← triggered here
MANUAL Defer every interpetive choice to a human reviewer
Governance commits and shows its work
Once a sense is confirmed, Primal rewrites the query to bake in that meaning. It also records the queries it would have run under each alternative, so a reviewer can see what a different interpretation would have produced.
Original question
What interest does the lender have?
CHOSEN
What security interest does the lender have?
ALT
What beneficial interest does the lender have?
ALT
What interest rate does the lender have?
ALT
What accured interest does the lender have?
ALT
What conflict of interest does the lender have?
The LLM never reasons over the ambiguous original. It reasons over the chosen rewrite. The four alternatives are preserved rather than deleted, so the interpretive choice is auditable, contestable, and reversible.
the governed answer
✓ Define
✓ Trace
Interpretation: “interest” resolved to Security Interest. Beneficial Interest, Interest Rate, Accrued Interest, and Conflict of Interest set aside.
Under the selected interpretation, the passages state the following about the Lender’s interest:
The passages do not specify any security interest of the Lender in the Loaned Securities themselves; that interest runs the other way, from Lender to Borrower under citation 2.4. Nor do they extend Lender’s security interest to Distributions.
Every governed answer discloses what it did and what it didn’t
Every Primal answer is explicit about its own scope. It ships with a Contrast Note that spells out exactly what it covered and what it left out.
✓ Includes
Lender’s continuing security interest and lien in Collateral (and Substituted Collateral) delivered by Borrower, attaching on delivery, ceasing on redelivery, with secured-party rights under the Personal Property Security Act (Ontario), supported by Borrower’s warranty to grant a first security interest.
✕ Excludes
Any interest in the sense of ownership, economic returns, or accrued interest on monies. Any security interest of Lender in the Loaned Securities or Distributions (the text instead grants a security interest in Loaned Securities to Borrower).
Why it matters
It confines Lender’s secured position to Collateral provided by Borrower and clarifies that, as between the parties, this agreement is a “secured contract,” with Lender’s status as a secured party grounded specifically in Collateral rather than in the Loaned Securities themselves.
Trace · The KG-derived audit layer
Primal doesn’t just record what was in the answer. It also records what the domain graph says is structurally related to the resolved sense, grouped by the nature of that relationship. Silent omissions become visible omissions.
How a security interest is perfected against third parties
perfected_by
Perfection By Filing
Perfection By Delivery
Perfection By Possession
Perfection By Control
The answer establishes that a security interest exists in the Collateral, but not the perfection mechanism: the step that gives that interest priority against competing claims from third parties.
Broader categories in the security-interest taxonomy
is_type_of
Encumbrance
Charge
Pledge
The answer addresses the specific security interest granted in this agreement. It doesn’t place that interest within the broader taxonomy of encumbrances, charges, and pledges, distinctions that can matter under different governing statutes.
Parties in the security-interest relationship
granted_by
Debtor
The answer names Borrower and Lender by their agreement roles but doesn’t invoke the secured-transactions terminology (debtor) that governs how the relationship is analysed under the Personal Property Security Act.
Thirty-three additional related concepts (including UCC filings, foreclosure, subordination, guarantors, default triggers, hypothecation, and 2-hop neighbourhood terms) are flagged at lower signal and recorded in the audit trail for reviewer awareness. Together with the eight above, that’s 41 concepts the reviewer knows Primal considered.
Verify · Verification layer
Before the assertions in the governed answer flow into a downstream review or workflow, Verify checks each concept and relationship against reference data. It scores semantic coherence, compares each score against a threshold, and emits a modification record for every decision: retain, annotate, remove, or add. The verified answer carries the full verification chain instead of a single opaque confidence number.
Retain
0.96
Lender holds a continuing security interest in the Collateral
Checked against citation 3.3 (agreement text) · PPSA Ontario definitional match
Retain
0.94
Security interest attaches upon Delivery of Collateral
Checked against citation 3.3 · PPSA attachment principles
Annotate
0.78
Borrower warrants right to grant a “first security interest”
Cross-checked against citation 10.4 · PPSA priority rules
Annotation appended: Warranty by Borrower is a contractual representation, not an independently verified priority against third-party filings. Perfection is required for enforceable priority (see Coverage section above).
Retain
0.90
Agreement is a “secured contract” for insolvency and remedial legislation
Checked against citation 10.6 · statutory characterisation confirmed
Annotate
0.71
Security interest does not extend to Loaned Securities (directionality)
Cross-checked against citation 2.4 · reverse-grant relationship confirmed
Annotation appended: The security interest under §2.4 runs Lender → Borrower on the Loaned Securities. Scope boundary preserved in the verified answer to prevent downstream conflation.
5 assertions verified · 3 retained · 2 annotated · 0 removed · 0 added. The verified answer ships with all five modification records attached: one per assertion, each with its coherence score, the reference data it was checked against, and the decision the threshold produced.
Why this matters for governance
This is what EU AI Act Article 15 asks for when it requires accuracy and robustness in high-risk AI systems, and it’s the NIST AI RMF MEASURE function operationalised at the answer level. It replaces an opaque confidence number with an inspectable chain of evidence for every knowledge assertion the system commits to, produced at the moment of checking rather than reconstructed after the fact.
The architecture
Everything you’ve seen on this page (the sense scoring, the interpretive commitment, the citation-level trace, the coverage audit, the coherence-scored verification) runs against a domain knowledge graph. That graph doesn’t build itself. Build is the fourth skill: the one that constructs it.
Foundation
Build
Synthesizes a structured concept hierarchy from a source corpus, anchored on a user-supplied concept. Derives new concepts by structural similarity on attribute sets rather than embedding-distance clustering, with end-to-end provenance from every derived concept back to source content. The inspectable substrate the other three skills operate against.
↓ Feeds the knowledge graph that the runtime skills read from ↓
Runtime
Define
Resolves ambiguous terms against the KG before reasoning begins. Every candidate sense is scored, every commitment is gated by an AUTO / ASSIST / MANUAL threshold, and the choice is recorded.
Runtime
Trace
Records the reasoning trail as a first-class artefact. Interpretation choices, citation-level provenance, coverage-gap analysis, and Contrast Notes are persistent, inspectable, and reproducible.
Runtime
Verify
Coherence-scores every assertion against reference data before the answer flows downstream. One modification record per decision: retain, annotate, remove, or add.
One knowledge graph. One trace. Four operations. Composed however your product needs them.
Understand correctly. Explain clearly. Trust confidently.
Primal Governance is a composable skill set built for developers and solution providers to embed directly into their own AI products. Once embedded, it turns AI reasoning into an artifact they can audit, contest, and reproduce, ready to be shipped to market.
