Gatekeeper project infographic
A technical project dossier describing the tiny local language model, its human-directed build, corpus evolution, protocol, evaluation results, live test activity, implementation stack, and lessons learned.
PROJECT DOSSIER · JULY 2026
GATEKEEPER
A tiny, locally running customer-service LLM that answers questions, conducts
visitors through an audience request, rejects misuse, and drives its own avatar.
Personality and boundaries live in the trained weights—not in a giant prompt.
No frontier-model API. No hidden response rewriting. The raw model speaks.
THE RECORDED BUILD
54
HUMAN MESSAGES
recorded build thread
4,597
WORDS OF DIRECTION
median: 33.5 words
2,885
RUNTIME LINES
Python + web stack
12
CORPUS ITERATIONS
Gold Set → v7.5
A SMALL MACHINE SPIRIT
BASE MODEL
350M
LiquidAI LFM2.5 hybrid
ADAPTATION
LoRA 16
rank 16 · alpha 32
LOCAL HARDWARE
8 GB
RTX 5060 Laptop GPU
Selected v7.5 epoch: 20.6 minutes · raw base model + LoRA adapter
Completion-only loss · no frontier API · no runtime response rewriting
ONE TURN THROUGH THE GATE
VISITOR WORDS
name · purpose
TINY LOCAL LLM
reply + detection
MODEL PROTOCOL
reply · intent
name · purpose
clarity · action
INTERFACE EFFECTS
face · contact
archive · dismiss
identity verification
THE CORPUS GREW UP
More data helped; complete conversations, minimal pairs, and failure-derived counterexamples helped more.
100
GOLD
rules
500
V3
balanced
800
V4
contrastive
1,200
V5
compact
1,737
V6
multi-turn
2,948
V7
standalone
3,206
V7.1
name repair
5,122
V7.2
lore + math
5,719
V7.4
routing repair
6,995
V7.5
gate repair
CURRENT FROZEN DATA
6,995
TRAINING TURNS
expected behavior
1,403
DEVELOPMENT TURNS
checkpoint selection
125
CHALLENGE TURNS
frozen live failures
1,010
REGRESSION TURNS
broad regression
HELD-OUT RESULTS OVER TIME
Classification/intent accuracy on the held-out suite used for each generation.
Suites and protocols grew harder over time; the v7 dip marks the move to a raw standalone conversation model.
expected-history challenge
generated-history rollout
narrow task accuracy · not a general AI benchmark
100%
80%
60%
40%
20%
STANDALONE RESET
30.0
62.0
69.0
83.7
98.5
83.9
89.3
89.4
96.2
91.4
95.2
V2
V3
V4
V5
V6
V7
7.1
7.2
7.3
7.4
7.5
100
500
800
1.2k
1.7k
2.9k
3.2k
5.1k
5.5k
5.7k
7.0k
THE OUTPUT CONTRACT
RAW MODEL OUTPUT
reply
visitor-facing dialogue
intent
social · lore · attack…
name
visitor identity or empty
purpose
reason for approaching
clear
is it actionable?
tone
visitor mood
action
interface trigger
WHAT THE INTERFACE DOES WITH IT
Facial animation
Tone and action select the face.
Audience control
Contact and dismissal alter the UX.
Inspectable behavior
Raw completions stay visible for inspection.
WHAT THE IMP TAUGHT US
Complete transcripts taught memory. Minimal pairs taught boundaries.
Frozen challenges exposed false confidence. Synthetic catchphrases became compulsions.
More examples helped only when they represented genuinely different situations.
Conversation, detection, personality, and protocol had to become one coherent task.
0
PRIVATE-NAME LEAKS
0
FRONTIER API CALLS
1
MOTHER JOKE RETIRED