12 KiB
CIVICUS-ROMAN-MODEL-VISION-0001
Rational Vision For A Bounded Roman Simulator Model
Status: Draft Vision
Layer: Training Infrastructure
Purpose: Define the practical rationale, scope, and training plan for the CIVICUS-ROMAN model
Repository Path: docs/training/chunking/CIVICUS-ROMAN-MODEL-VISION-0001.md
0. Purpose
This document defines the rational vision for the CIVICUS-ROMAN model.
The model is not intended to be a general chatbot.
The model is not intended to know all of history.
The model is not intended to imitate modern English reasoning with Roman facts attached.
The model is intended to operate inside a bounded Roman simulator world.
Its task is to reason, ask, answer, and speak from within that world.
1. Core Claim
A narrow Roman simulator model may be viable because the intended world is deliberately reduced.
The model does not need the full ontology of modern life.
It needs a bounded set of:
objects
actions
pressures
actors
places
procedures
records
obligations
materials
routes
risks
social meanings
The target is not general intelligence.
The target is Roman-bounded simulator intelligence.
2. The Problem With Existing Models
Existing general models are trained on modern reality.
Even when given Roman context, they tend to leak modern assumptions:
universal market price
modern legal enforcement
modern contract logic
state-backed regulatory assumptions
instant information
abstract finance vocabulary
modern supply-chain concepts
consumer-market behavior
modern moral and institutional framing
Retrieval alone does not solve this.
RAG can supply correct facts, but the base model still interprets those facts through a modern ontology.
The goal of CIVICUS-ROMAN is to reduce or remove that ontology problem.
3. What The Model Must Learn
The model must learn to reason from Roman-visible primitives.
Examples:
Who saw it?
Who heard it?
Who wrote it?
How old is the message?
Is the seal broken?
Who witnessed the bargain?
Where are the carts?
Can the goods move?
Who benefits if the rumor is believed?
What can safely be entered in the account?
Is the obligation settled, pledged, delayed, or disputed?
It must not default to:
What is the market price?
Is the contract enforceable?
What is the regulatory risk?
What is the optimal modern transaction?
The model should ask and answer in terms of objects, actions, pressures, and visible social facts.
4. Reduced World Grammar
The CIVICUS-ROMAN model should be trained around a controlled world grammar.
Objects
coin
purse
chest
tablet
seal
witness
cart
wheel
mule
road
warehouse
wall
roof
jar
amphora
crate
rope
weight
measure
gate
market
portico
yard
dust
rain
lamp
grain
oil
bronze
timber
glass
stone
Actions
buy
sell
carry
store
seal
open
count
weigh
measure
pledge
write
witness
hire
repair
delay
ask
refuse
accuse
confirm
return
split
hold
move
settle
hide
leak
wait
rot
spoil
break
arrive
depart
Pressures
hunger
rain
delay
spoilage
debt
rivalry
shame
praise
shortage
crowd
rumor
cart scarcity
storage scarcity
buyer urgency
creditor pressure
official attention
bad road
old news
broken seal
empty purse
full warehouse
The model should learn to combine these before reaching for abstract explanation.
5. Speech Principle
The model should prefer Roman-visible commercial speech.
Preferred:
The wheels are gone.
The tablet arrived old.
He owns jars, not coin.
The road has eaten the profit.
The crate is heavier than its name.
The purse is fat and the street has eyes.
Avoided:
Transport capacity is constrained.
The information is stale.
His assets are illiquid.
Transportation cost eliminated the margin.
The cargo is misclassified.
Liquidity creates security risk.
The purpose is not ornament.
The purpose is ontology.
A model learns the kind of world it inhabits through the language it is trained to use.
6. Corpus Architecture
The corpus is layered.
Each layer teaches a different kind of reasoning.
Layer 0 — Primitive Facts
basic world rules
Layer 1 — Worked Examples
arithmetic, cost, movement, profit, loss, settlement
Layer 2 — Uncertainty
reports, rumors, old messages, hidden truth, confidence, confirmation
Layer 3 — Actor Perspective
same event read differently by different Roman-world actors
Layer 4 — Dialogues
in-world scenes that teach through speech, action, and consequence
This layering is essential.
The model should not merely memorize dialogue.
It should learn the underlying reasoning forms that make the dialogue valid.
7. Vocabulary Generation Pipeline
A major part of the model vocabulary can be built through a generate-review-promote workflow.
The generator combines:
Object + Action + Pressure
Example:
cart + hired elsewhere + buyer waiting
= The wheels are gone, and the buyer will not wait for our excuses.
Most generated phrases will be weak.
That is acceptable.
Humans are faster at recognizing strong expressions than inventing them cold.
The workflow is:
generate many candidates
human flags useful expressions
accepted expressions enter vocabulary
strong expressions influence dialogue
canonical expressions become simulator templates
Only reviewed material enters training.
Raw churn is not training data.
8. Human And Agent Roles
Agents will perform much of the production work.
Agents can generate:
candidate expressions
dialogue variants
actor readings
primitive examples
uncertainty cases
law scenarios
architecture scenarios
technology scenarios
negative examples
contamination tests
Agents can also assist with:
format validation
tag audit
style checks
duplicate detection
forbidden vocabulary detection
chunk extraction
statistics
regression tests
Humans remain responsible for:
canon
ontology
final approval
style judgment
failure judgment
domain boundaries
promotion to training data
The human role shifts from authoring every line to governing the corpus.
9. Training Strategy
The first serious training target should not be a general-purpose language model.
The first target should be a compact bounded simulator model.
A rational training progression:
Stage 1:
Roman-visible vocabulary expressions
Stage 2:
primitive facts and terse Q/A
Stage 3:
worked examples with arithmetic and consequence
Stage 4:
uncertainty examples and knowledge-boundary tests
Stage 5:
actor-perspective readings
Stage 6:
in-world dialogues
Stage 7:
simulator-state-to-response pairs
The model should learn from simple controlled forms before complex dialogue.
10. Scratch Training Reconsidered
Training a general model from nothing is expensive because the model must learn broad language, broad world knowledge, and general reasoning.
CIVICUS-ROMAN is different.
It does not need to answer every question.
It does not need modern breadth.
It does not need open-ended knowledge.
It needs competence inside a small Roman simulator world.
Therefore scratch or near-scratch training may be viable if the model is deliberately narrow.
The fair comparison is not:
small project vs general LLM
The fair comparison is:
bounded simulator grammar + controlled corpus + agent-assisted data generation
against:
modern-prior leakage from general models
11. Simulator Ownership Of Reality
The model should not own the simulator state.
The simulator owns:
actors
locations
time
inventory
money
routes
documents
seals
witnesses
obligations
weather
prices
rumors
official attention
The model interprets, asks, answers, and speaks within that state.
The model should not invent facts that the simulator has not provided.
The model should prefer questions when state is insufficient.
Example:
What can be known?
Who saw it?
Who wrote it?
Can the cart still move?
Was the seal broken?
Is there a witness?
12. Evaluation
The model must be tested against modern contamination.
Example failure prompt:
What is the fair market price?
Roman-bounded response should reject universal price and ask about place, buyer, time, transport, and information.
Example failure prompt:
Can the contract be enforced?
Roman-bounded response should ask about tablet, witness, seal, pledge, patron, magistrate, standing, and leverage.
Example failure prompt:
Was the information reliable?
Roman-bounded response should ask who carried the word, how old it is, who benefits, whether anyone saw the goods, and what can be confirmed.
Evaluation must reward Roman-bounded reasoning and punish modern abstraction.
13. Domains To Add
The first domain is commerce.
Next domains should be added with the same layered discipline.
Roman Law
standing
complaint
witness
tablet
seal
pledge
remedy
magistrate
patronage
procedure
public shame
private settlement
Architecture
stone
timber
brick
lime
labor
measurement
site
water
weight
collapse
repair
patron
public work
Technology
tool
craft
material
workshop
repair
failure
skill
apprentice
measurement
heat
water
wheel
gear
lever
Each domain should develop:
Layer 0 primitives
Layer 1 examples
Layer 2 uncertainty
Layer 3 actor readings
Layer 4 dialogues
controlled vocabulary
contamination tests
14. Practical Near-Term Plan
Recommended next steps:
1. Freeze first commerce dialogue batch.
2. Continue vocabulary generation standards.
3. Build the expression candidate generator.
4. Build a review interface for accept/reject/strong/canonical.
5. Expand commerce vocabulary library.
6. Add Roman Law Layer 0 primitives.
7. Add Roman Law worked examples.
8. Add Roman Law dialogues only after primitives exist.
9. Build contamination tests.
10. Compare:
A. scratch small model
B. near-scratch model
C. small existing base model fine-tuned to OTIVM
The comparison matters.
The project should not assume scratch training wins.
It should test whether scratch training reduces modern contamination enough to justify weaker inherited language ability.
15. Success Definition
CIVICUS-ROMAN succeeds if it can operate inside the simulator without modern leakage.
It should naturally produce questions and answers like:
Who carried the word?
How old is the tablet?
Was the seal broken?
Can the cart still move?
Who witnessed the promise?
Does the account remain open?
What does the buyer need before sundown?
It should naturally speak like:
The wheels are gone.
The tablet arrived old.
He owns jars, not coin.
The road has eaten the profit.
The account remains open.
The crate is heavier than its name.
It should avoid:
supply chain disruption
market efficiency
legal compliance
liquidity constraint
regulatory exposure
contractual enforcement
The model is not meant to know less.
It is meant to know differently.
16. Final Vision
CIVICUS-ROMAN is a bounded-world model.
Its intelligence comes from discipline, not breadth.
Its strength is that it does not treat modern reality as default.
It learns a smaller world deeply:
what can be seen
what can be carried
what can be written
what can be witnessed
what can be pledged
what can be delayed
what can be hidden
what can be settled
This is the rational path:
controlled ontology
layered corpus
Roman-visible vocabulary
agent-assisted generation
human canon approval
strict validation
small model experiments
simulator-owned state
contamination testing
The purpose is to build a model that does not merely describe Ancient Rome.
The purpose is to build a model that can think inside the civic Roman world of the simulator.