A governance layer for interchangeable language models.
02The problem
Same question, different answer
Today, LLM answers cannot be verified — no provenance, no repeatability, no trace.
No provenance
No trace
No control
Question · Which database fits our project?
Go with PostgreSQL. Relational, mature — and your data is clearly structured.
Question · Which database fits our project?
Go with MongoDB. Schemaless, quick to start — your data model is still changing.
Both plausible. Neither says where it came from.
03Core thesis
Not the prompt. The layer.
Emergence is not a chatbot with a better prompt, but an auditable governance layer over interchangeable models — every answer has a traceable provenance.
Model interchangeable
Behavior configured
Provenance traceable
Application
GovernanceOrchestration · Rules · Logging
Model A
Model B
Model C
Model D
swapped
The model changes. The layer stays.
04The pipeline
Four phases around every answer
Every request passes through the same four phases — understand, reason from perspectives, generate, check — and the actual run can be inspected as a graph.
PRE_GEN
classify, select knowledge
LENS_GEN
perspectives, optional
GEN
generate with tools
POST_GEN
inspect, regenerate if needed
Classifier
Rules · Notes
Lens
Lens
Synthesis
Answer
Tools
Inspection
Gate
The trace shows a real turn as a graph — from classification to the gate.
05Rules I · Effect
One rule, visibly in effect
Behavior is configured, not talked into the prompt: same request, one rule added, a visibly different answer.
Trigger + action
Effective immediately
Configuration, not prompt
without rule
rule added
with ruleThe rule consists of a trigger and an action.
06Rules II · Mechanism
Not in the prompt. Chosen at runtime.
The rule from the previous slide was in no prompt: for every request the system decides anew which rules apply — by trigger, by semantic similarity and along references between entries, again after every tool call.
Check triggers
Search similarity
Follow references
Knowledge base
Rule #886 from 05
Note
Note · similar
Module · via reference
Rule · trigger met
Rule
Context of this turn
Request arrives
Tool call · new state
In the trace the rule appears as a selected entity, trigger and action readable.
Off · Auto · On switches: what can be switched on can be switched off.
07Lenses
Perspectives run separately, then together
Each analytical perspective is its own run with only that view; a synthesis merges them and names contradictions instead of smoothing them over.
One run per lens
Step check per lens
Synthesis names conflicts
In the trace, two lenses run as parallel branches into the synthesis.
08Deep Mode
Plan, sub-runs, meta-synthesis
For hard questions the system first plans several distinct sub-investigations, runs them in parallel and merges them in a meta-synthesis.
Plan
Work in parallel
Merge
09Evidence required
Evidence required, not trust assumed
In research mode the turn keeps a register of claims and source hashes; the release checks that the answer is not empty, that every cited source resolves and that every claim appears verbatim in the answer.
Claim + source registered
Verbatim anchor
Release or plain-text reason
AnswerEvery request passes four fixed phases.C1A human decides on proposals.C2Changes can be reverted one by one.C3
Claim
Source
Source hash
verbatim anchor
C1
S1
sha256 ····
yes
C2
S2
sha256 ····
yes
C3
S3
sha256 ····
yesno
checking
blocked
released
BLOCKED
C3 is not anchored verbatim in the answer.
RELEASED
Answer not empty · sources resolve · claims anchored.
LimitationThe check confirms that a source is cited and resolves — not that it supports the claim. No factual truth check runs, and this is reported as not performed.
10Discarded drafts
The discarded draft stays visible
An always-on check pass flags issues only with a verbatim quote from its own draft; if that leads to regeneration, the rejected draft stays viewable with its reason.
Flag only with a quote
Regenerated, not deleted
Reason shown alongside
▸ Discarded answerReason: promise without basis — “guaranteed within one hour”
DraftThe migration is guaranteed within one hour. After that everything runs as usual.
Inspection · quote: “guaranteed within one hour” — promise without basis in the context.
RegeneratedThe duration depends on the data volume. A trial migration gives a reliable estimate.
11Audit
Every behavior change can be reverted
Every change to a rule, module or note leaves a before/after entry and can be reverted individually.
Before/after
Who, when, why
Revert individually
Rule
Note
Module
Rule
Note
reverted
The entry shows before and after and can be reverted on its own.
12Gate · Behavior
Nothing is learned automatically
The system proposes behavior changes itself — but every proposal passes an automatic critic and always ends with a human who decides.
System proposes
Critic annotates
Human decides
System
Proposal
Change + rationale
Automatic
Critic
Note on the proposal
Human
Yes / No · path ends here
A proposal awaits a decision, with critic note and rationale.
13Compliance report
The evidence is one click away
For any period the system produces a content-free report covering record-keeping, transparency, human oversight and robustness — with a checksum over the frozen data snapshot.
ScopeA structured self-report from operational data — not a legal assessment, not proof of conformity. Article references are orientation (“cf.”), not certification.
The issued report carries its period and a valid seal.
14Routines and connectors
Runs on schedule, not on discretion
A routine fetches its inputs on schedule, checks them against a stored goal and reports only what matters through fixed actions — and it starts as a draft that only runs after review, an action-free dry run and explicit activation.
Schedule, not ad hoc
Dry run without effect
Activation stays human
DRAFT
→
DRY RUN
→
ACTIVE
created. not running.working. no effect.a human switches it on.runs on schedule. reports once.
Spreetropol Systems