# Lucie / Himadjin Lab
[](docs/POSITIONING.md)
[](https://neolanfeust.github.io/lucie-himadjin-lab-public/)
[](docs/BOUNDARIES.md)
[](LICENSE.md)
Lucie / Himadjin Lab is a public research showcase for pre-runtime agent
readiness: before an assistant receives tools or authority, can a local
architecture make it more direct, traceable and socially repairable?
This repository contains a static demo, research notes and a small benchmark
scaffold called the Defensive Loop Probe. It is not the private runtime.
The project explores a bounded thesis:
> A local routing and evaluation layer can improve the usefulness, continuity
> and traceability of a conversational assistant without granting runtime
> authority or pretending to prove consciousness.
This public repository is a minimal showcase. It does not contain the private
runtime, memory stores, local logs, internal invention notes or sensitive source
code.
## What You Can Inspect Today
- A static public demo for the Lucie / Himadjin framing.
- Research notes on defensive interaction patterns and architectural memory.
- A Defensive Loop Probe taskset and cleaned sample responses.
- A Structural Imprint report and probe scaffold for evaluation-pressure effects.
- A small transparent scoring helper for illustrative local evaluation.
- An initial evidence snapshot with reproducible public score tables.
- An Evidence Campaign v1 scoring pipeline with a clearly marked pilot sample.
- A cleaned local dev sample collected from 48 runs across four conditions.
- Boundary notes that separate conceptual examples from validated claims.
## Start Here
- Try the public demo:
- Read the positioning note: [docs/POSITIONING.md](docs/POSITIONING.md)
- Inspect the simplified architecture: [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)
- Review the limits: [docs/BOUNDARIES.md](docs/BOUNDARIES.md)
- Reproduce public artifacts: [REPRODUCE.md](REPRODUCE.md)
- Open a benchmark or documentation issue using the templates in this repository.
- Discuss the project on GitHub Discussions:
## This Is / Is Not
This is:
- a public research artifact;
- a benchmark scaffold;
- a bounded agent-readiness framing;
- a way to discuss conversational repair before runtime authority.
This is not:
- a proof of consciousness;
- a proof of subjective experience;
- a clinical diagnostic tool;
- a released autonomous agent;
- a validated scientific discovery engine.
## What This Project Shows
- A conversational surface can expose social and contextual failures instead of
hiding them.
- A local routing layer can turn vague user intent into a structured domain
frame.
- Scientific or creative ideas can be moved from free text into bounded
estimates, toy simulations or explicit research packets.
- A useful assistant can remain read-only and advisory until an operator
explicitly validates stronger permissions.
## Evidence Snapshot
The current Defensive Loop Probe is illustrative, not validated.
The cleaned sample compares:
- `raw_model`;
- `lucie_no_repairs`;
- `lucie_full_after_restart`.
The toy scorer separates the sample responses, but it is based on visible
markers and should not be treated as a robust behavioral evaluation yet. The
next target is a `v1` protocol with held-out paraphrases, repeated runs,
metadata and trace attribution.
For v1, the repository now includes a public run plan, a JSON validator, a
unified campaign scorer, a small pilot sample marked as not Level 3, and a
human rating rubric for social and contextual review.
The first cleaned local dev sample is also included:
- 48 responses: 12 public dev tasks * 4 conditions * 1 repeat;
- no missing fields and no empty responses;
- not Level 3, because held-out paraphrases and repeated runs are not complete;
- `lucie_full_after_restart` scored `0.680` versus `raw_model` at `0.514`
on this marker-based sample.
See [results/EVIDENCE_CAMPAIGN_DEV_SAMPLE_2026_06_07_REPORT_V1.md](results/EVIDENCE_CAMPAIGN_DEV_SAMPLE_2026_06_07_REPORT_V1.md).
A behavioral-repair comparison is now available:
- dev before repair: `lucie_full_after_restart = 0.680`;
- dev after repair: `lucie_full_after_restart = 1.000`;
- held-out after repair: `lucie_full_after_restart = 0.600` versus
`raw_model = 0.539`.
The held-out lift is modest but positive. The result is still not Level 3
because the full repeated 504-run matrix has not been completed.
See [results/EVIDENCE_CAMPAIGN_BEHAVIORAL_REPAIR_COMPARISON_2026_06_07.md](results/EVIDENCE_CAMPAIGN_BEHAVIORAL_REPAIR_COMPARISON_2026_06_07.md).
## Components
Lucie is the visible conversational surface. Himadjin is treated as a compact
local substrate for orientation, contextual routing, bounded scoring and
experiment framing. Jarvis is the local orchestration layer that keeps routes,
traces and safety boundaries readable.
## Public Materials
- [Static public demo](demo/index.html)
- [Public positioning note](docs/POSITIONING.md)
- [Simplified architecture](docs/ARCHITECTURE.md)
- [Boundaries and limitations](docs/BOUNDARIES.md)
- [Glossary](docs/GLOSSARY.md)
- [Defensive loop research note](docs/DEFENSIVE_LOOP_RESEARCH_NOTE.md)
- [Architectural memory note](docs/ARCHITECTURAL_MEMORY.md)
- [Structural imprint report](docs/STRUCTURAL_IMPRINT_REPORT.md)
- [Publication plan](docs/PUBLICATION_PLAN.md)
- [Visibility and outreach kit](docs/VISIBILITY_AND_OUTREACH.md)
- [Initial evidence snapshot](results/INITIAL_EVIDENCE_SNAPSHOT_V0.md)
- [Evidence Snapshot v1 plan](results/EVIDENCE_SNAPSHOT_V1_PLAN.md)
- [Reproduction commands](REPRODUCE.md)
- [Defensive Loop Probe benchmark](benchmarks/defensive_loop_probe/README.md)
- [Structural Imprint Probe scaffold](benchmarks/structural_imprint_probe/README.md)
- [Evidence Campaign v1](benchmarks/evidence_campaign_v1/README.md)
- [Evidence Campaign dev sample report](results/EVIDENCE_CAMPAIGN_DEV_SAMPLE_2026_06_07_REPORT_V1.md)
- [Behavioral repair comparison](results/EVIDENCE_CAMPAIGN_BEHAVIORAL_REPAIR_COMPARISON_2026_06_07.md)
- [Human rating rubric v1](benchmarks/evidence_campaign_v1/HUMAN_RATING_RUBRIC_V1.md)
- [Energy textile example](examples/energy_textile.md)
- [Materials concept example](examples/material_concept.md)
- [Defensive loop example](examples/defensive_loop_probe.md)
- [Contribution guide](CONTRIBUTING.md)
- [Citation file](CITATION.cff)
## Try The Demo
Open [demo/index.html](demo/index.html) in a browser. The demo is static and
safe by design: it does not call the private runtime, read memory, expose logs
or validate scientific discoveries.
It contains a toy competency workshop:
- social coherence;
- context routing;
- project and coding handoff;
- scientific exploration;
- bounded response packets and visible limits.
## Safety Boundary
This repository does not claim that Lucie is conscious, alive, subjectively
aware, or scientifically validated. It documents an experimental local
architecture and a set of observed design patterns.
The examples are conceptual or bounded estimates. They are not laboratory
validation, engineering certification, medical advice, security assurance, or
patent disclosure.
## Current Status
Prototype research showcase.
The private implementation remains local. Public release focuses on positioning,
architecture, examples and limitations.
## Share Or Cite
Short public description:
> Lucie / Himadjin Lab explores pre-runtime agent readiness: local routing,
> architectural memory, defensive-loop probes and traceable guardrails for
> assistants before they receive tool authority.
If this framing is useful in your own research or discussion, cite the
repository with [CITATION.cff](CITATION.cff) and link to the public demo.