01
The judgmentThe decision guidance itself, in plain language a person could act on without translation.
Katacylizm · Architecture
The missing intelligence infrastructure
Four planes. One governed path from judgment to outcome.
The industry stops at agents. Katacylizm continues into memory, governance, and provenance. Reasoning is captured once, validated, governed, retrieved with context, improved through evidence, and preserved with contributor lineage — so the reasoning tax is paid once and amortised, not re-billed on every call.
The four planes
Each plane owns one question the others are not allowed to answer.
Separation is the point. What may be reused is not decided by the thing that wants to use it, and what is allowed to ship is not decided by the thing that generated it.
PLANE 01
Reasoning Asset Plane
What is allowed to be remembered.
HOLDSValidated, versioned decision assetsEach asset carries the judgment itself plus its context of validity, its boundaries, its supporting evidence, its version history, and its contributor lineage.
ENFORCESAdmission by standard, not by generationOutput does not become an asset because a model produced it. It becomes an asset because a human-governed standard admitted it.
REFUSESSilent overwriteReasoning is superseded, never quietly replaced. The prior version stays addressable, so a decision made last quarter can still be explained this quarter.
PLANE 02
Runtime Decision Plane
What to do about the situation in front of you right now.
RESOLVESituation before retrievalThe live context — who, which environment, what stage, what has already been tried — is resolved first. Retrieval without resolution is how systems give confident, irrelevant answers.
RETRIEVEMinimum sufficient intelligenceThe smallest set of assets that can responsibly answer the situation. Not the largest pile the context window can hold.
ACT OR ESCALATEA real third optionAdapt and act, or stop and hand to a human. Escalation is a first-class outcome, not an error state — which is what makes the layer safe to point at customers.
PLANE 03
Trust + Governance Plane
What must never happen, regardless of what the model wants to say.
AUTHORIZATIONScope per actorHumans and agents operate inside an explicit scope. An agent cannot reach reasoning it was never authorised to use.
ISOLATIONTenant boundariesOne organization’s reasoning never leaks into another’s runtime. Isolation is structural, not a filter applied late.
QUALITY GATEThe admission standardGeneric, unsafe, or out-of-bounds output is rejected before it can become strategy — and before it can enter memory.
STOP LOGICWhere persuasion endsPressure, hostility, and unsafe persuasion are detected and halted. Safety is a property of the layer, not a setting someone can switch off.
AUDIT + ROLLBACKAnswerability after the factEvery governed decision leaves a record, and every asset version can be retired or reverted without retraining anything.
OVERSIGHTHumans at the checkpointsReview is a required step in the path, not an optional queue nobody opens.
PLANE 04
Learning + Provenance Plane
How the system gets better without quietly getting worse.
CANDIDATESOutcomes become proposalsReal results generate governed candidates for new or revised reasoning. Candidates are proposals, not promotions — they still have to pass the gate.
LINEAGEContributor, evidence, versionWho contributed the judgment, what evidence supported it, and which version was in force are preserved through deployment and reuse.
MEASUREMENTReuse and outcome signalsWhich assets are being used, in what situations, with what result — the input to both improvement and attribution.
One rule spans all four: nothing becomes reusable intelligence until a human-validated gate says it may.
The unit of the system
Anatomy of a reasoning asset.
A document is a blob. A prompt is an instruction. A reasoning asset is a governed object that knows where it applies, what it is based on, who produced it, and when it stops being true.
02
Context of validityThe situations where this applies: environment, stage, relationship, risk posture. Specific enough to be retrievable, honest enough to be limited.
03
BoundariesWhere it must not be used, and what it must never be used to do — encoded with the asset rather than left to the operator’s discretion.
04
EvidenceWhat supports it: observed outcomes, prior use, expert review. Confidence is claimed against evidence, not asserted.
05
Version + review stateWhich revision this is, what it superseded, who reviewed it, and whether it is currently in force.
06
Contributor lineageThe authenticated record of whose judgment produced it — the hook the Expert Value Framework hangs on.
The governed lifecycle
Every durable insight travels the same seven steps.
Memory is not automatic. It is earned, reviewed, versioned, and tied to the human and the evidence that produced it.
1
CaptureReal language, real situations — not a form somebody has to remember to fill in.2
ValidateThe Quality Gate admits or rejects. Rejection is the common case, by design.3
AttributeContributor and evidence are bound to the asset before it can be used.4
DeployMade retrievable to authorised humans and agents, inside its stated boundaries.5
MeasureUse and outcome are observed — including where it was wrong.6
ImproveEvidence produces a governed candidate for the next version.7
RewardContribution becomes visible, measurable, and — by contract — participable.↻ Reward feeds back into capture. Each validated use strengthens the corpus and the proof loop, which is why the asset appreciates while a static playbook decays.
“The next unlock isn’t models that think harder. It’s systems that stop paying to rediscover expertise the organization already owns.”The design premise behind all four planes
At runtime
What happens between a signal and an action.
1
Signal arrivesA human pastes it, or an agent encounters it mid-conversation.2
Situation resolvedEnvironment, stage, and history established before anything is retrieved.3
Assets retrievedMinimum sufficient intelligence, inside authorisation scope.4
Gate appliedResponse checked against standard, boundary, and stop logic.5
Act or escalatePlain-language guidance out — or a hand-off to a human, recorded either way.
·No integration required to evaluate
·Escalation is a valid outcome
·Every decision leaves an audit record
·Same rules for humans and agents
Proof, not diagram
Eleven product layers map to these four planes, with live, protected, expanding, and planned status disclosed.
Architecture that has never met a customer is a slide. Katalyst is the running instance: the Objection Intelligence engine, the Quality Gate, Persuasion Memory, three workspaces, asset generation, CRM-ready export, and live practice — each sitting on the plane that governs it.
Is this a database, a model, or an application?
None of the three on its own. It is the layer between them: a governed store of reasoning assets, a runtime that resolves and retrieves, a gate that decides what may ship, and a provenance graph that survives both. Models are engines it uses; applications are consumers that sit on top.
Where does our data live and who can reach it?
Tenant isolation is structural, and every actor — human or agent — operates inside an explicit authorisation scope. Specific deployment topology and security controls are covered under NDA rather than published here.
How is this different from retrieval over our call recordings?
Retrieval over transcripts returns what was said. This returns validated reasoning about what to do, bounded to the situation it applies to, with lineage attached. The gate and the versioning are the difference between a search result and a decision you can defend.
What happens when a validated asset turns out to be wrong?
It is versioned or retired, and because lineage is preserved you can find every place it was in force. That is the point of rollback: correcting reasoning should not require retraining a model or re-onboarding a team.
Do agents get the same access as people?
They get the same rules and narrower scope. Agents operate under authorisation, boundaries, and stop logic identical in kind to a human’s — with less latitude, because they cannot yet be held accountable.
What is described on this page is the observable architecture — the structure, the ordered steps, and the guarantees. Implementation values, scoring standards, corpus contents, and security specifics are deliberately not published and are held as trade secrets. Patent pending, U.S. Provisional Application No. 64/099,949.
Next step
The architecture is the asset. The wedge is the proof.
Technical and investor briefs cover the same ground in more depth: plane responsibilities, the asset model, the role of the gate, and the twelve-month proof plan.
Patent pending · The Thesis · Product