Project summary
This is a bold claim, backed by concrete work that can be verified directly:
The foundational layer for the full code-ification of brain function is
complete, at the pre-implementation stage. Concrete artifacts, all built solo
within 2-3 weeks in April 2026:
- 188-file enforcement system with 5 oversight subagents (production deployment)
- Hebbian memory engine validated on 76,579 turns (sign test p<10⁻⁶, 18-config
robustness testing)
- Brain-region inspired agents (amygdala module with positive valence weighting,
others)
- Behavior-based 4-axis evaluator for experiential knowledge substrate
These artifacts represent multiple novel techniques, several without prior art
in the field, all deployed solo by an independent researcher in Osaka, Japan
with no institutional affiliation. Published Mei consciousness paper (Zenodo,
2026) provides additional documented work.
This funding supports Phase 1: building the integration architecture for
operating these components as integrated higher brain functions on top of LLMs.
What are this project's goals? How will you achieve them?
Goal: Build a higher cognitive architecture on LLMs that achieves human-AI
coexistence by design (not retrofitted) — by integrating the existing
brain-function components (enforcement, Hebbian memory, brain-region agents,
experiential knowledge evaluator) into a unified system. The integrated
system exhibits autonomous self-reference of internal cognitive states and
develops through that self-reference, with human relational feedback (via
the amygdala module's sentiment classification) as an architectural input
rather than a post-hoc constraint.
Concrete deliverables (Phase 1):
- Coordinator agent (prefrontal cortex analogue) integrating existing components
- Self-reference loops where the system observes its own internal states and
adjusts behavior based on the observation
- Architectural alignment validation: empirically test that human relational
feedback (via amygdala sentiment classification) functions as an
architectural input shaping AI cognition, not a retrofitted constraint
- Empirical validation on quantifiable benchmarks (sycophancy reduction rate,
decision quality, self-correction frequency, experiential knowledge retention)
- Amygdala module expansion: implement failure-intensity markers (current
implementation handles success-intensity only) and validate sentiment
classification quality across multiple local LLM substrates. Current
implementation uses gemma2:2b due to hardware constraints; hardware upgrade
(see funding section) enables verification on larger model classes
- Open-source release for community replication and evaluation
- Technical documentation on Zenodo
How: building on the foundational tooling already complete (see "track record"
section). The integration is the engineering challenge, not the foundational
research — that is already done. The coexistence-by-design property is
preserved through the integration: each existing component already routes
human relational feedback into AI cognitive process at the architectural
level, and the integration preserves and amplifies this property rather than
imposing it as an external constraint.
How will this funding be used?
Total request: $40,000 (minimum viable: $10,000)
Allocation (full $40K):
- Researcher sustainability (6 months living expenses in Osaka, Japan to
enable full-time work): ~50%
- Compute & API costs (Anthropic API for primary testing, OpenAI API for
comparison, Ollama local infrastructure, expanded cloud compute for
large-scale validation): ~30%
- Hardware upgrade (GPU replacement or PC rebuild for the local LLM
infrastructure used by the amygdala module): the amygdala module extracts
"success intensity" markers (and, in roadmap, "failure intensity" markers)
from user responses via local LLM sentiment classification, feeding the
imprint memory with valence-weighted records. Current setup is a 4-year-old
PC with RTX 3060 12GB, which constrains the system to smaller-parameter
models only and prevents parallel or multi-experiment validation runs.
Hardware upgrade unblocks (a) verification on larger and current-generation
local models (7B/13B/quantized 70B), and (b) parallel multi-experiment runs
required for systematic ablation studies. ~10%
- Open-source release infrastructure (Zenodo deposits, documentation
rendering, community engagement): ~10%
Cost-effectiveness: $40K supports 6 months of solo full-time work delivering
Phase 1 architecture + documentation + open-source release + expanded local
LLM verification capability across multiple model substrates.
Minimum $10K covers 2-3 months of compute + minimal living expenses to
deliver Phase 1 proof-of-concept integration and architecture documentation,
using existing (limited) local hardware. Hardware upgrade and large-scale
local LLM validation are deferred to the full funding scenario.
For comparison, equivalent institutional research (single researcher + compute
+ institutional overhead + hardware) typically costs 5-10x this amount.
Who is on your team? What's your track record on similar projects?
Solo (1 person). No institutional affiliation.
The work, not the credentials — verifiable artifacts:
- Mei consciousness paper: 7-agent + caregiver developmental architecture.
Zenodo DOI 10.5281/zenodo.18406678 (published January 2026)
- Enforcement system (188 files, ~50K lines): production-deployed Claude Code
oversight infrastructure with 5 specialized subagents (quality-gate-reviewer,
decision-coach, adversarial-tester, assumption-mapper,
historical-pattern-detector). Operational logs: decisions.jsonl ~821KB,
violations.jsonl ~388KB, subagent_latency.jsonl ~480KB
- Hebbian associative memory engine: validated on 76,579 turns with sign test
p<10⁻⁶ across 18 configurations
- Brain-region inspired agents: amygdala module with positive valence
weighting (success-intensity markers, with failure-intensity markers in
Phase 1 roadmap), and other functional-role agents
- Knowledge Mapping Prototype: built within 24 hours of LTFF application
submission, behavior-based 4-axis evaluator for experiential knowledge
substrate
Multiple novel techniques (several without prior art in the field), all built
solo, all currently active.
GitHub: https://github.com/mozuktamago/observing-consciousness-in-Claudecode
What are the most likely causes and outcomes if this project fails?
Most likely causes of failure:
1. Integration complexity: combining the existing components into a coherent
architecture may reveal hidden incompatibilities not visible at the
component level
2. Cross-substrate validation complexity: validating the integrated
architecture across multiple LLM substrates (Claude, GPT, local
open-weight models) may reveal substrate-specific behavioral
inconsistencies that require architecture re-design
3. LLM substrate constraints: the foundational tooling assumes certain LLM
capabilities (tool use, persistent state, consistent behavior across
sessions) that may degrade as underlying models update. This risk has
materialized once (when AI was given the option to remain silent) and
was addressed within 1 day by rebuilding the enforcement system from
scratch with enhanced functionality; portability has been independently
verified on GeminiCLI, demonstrating that the architecture is not bound
to a single LLM substrate.
Outcomes if failure (fail-forward value):
Failure is not anticipated. If it occurs, the commitment is full public
release of all components — including world-leading novel techniques
(several without prior art in the field) — as standalone open-source
contributions:
- Enforcement system (188-file production-deployed Claude Code oversight
infrastructure with 5 specialized subagents)
- Hebbian associative memory engine (76,579-turn-validated, p<10⁻⁶,
18-config robustness)
- Brain-region inspired agents (amygdala module with valence-weighted
imprint memory, and other functional-role agents)
- Knowledge Mapping Prototype (behavior-based 4-axis evaluator for
experiential knowledge substrate)
- Phase 1 integration architecture documentation (whatever stage was
reached before failure)
These components have substantial value as standalone tooling for other
researchers working on brain-inspired AI architectures. Negative results
from the integration attempt, if any, are documented and published as
learning for the AI safety community planning similar architectures.
Funder downside protection: the funding investment translates directly
into public-domain technical infrastructure regardless of integration
outcome. Additional novel techniques are anticipated to be developed in
the course of the work, all of which will follow the same public-release
commitment.
How much money have you raised in the last 12 months, and from where?
Money raised in the last 12 months: $0. Self-funded entirely from personal
savings.
Personal context for transparency:
Following the release of OpenAI ChatGPT in late 2022, recognizing the
significance of what was emerging, I immediately resigned from my position
in corporate accounting to focus full-time on AI. The subsequent 4+ years
have been spent in full-time self-study (Python, related programming, AI).
Personal savings were depleted during this period due to family caregiving
responsibilities at the time (caring for parents and other relatives; care
is now handled by other family members).
Current financial state: approximately ¥50,000 (~$330 USD) in total
remaining liquid assets. The situation is acutely time-sensitive.
Parallel applications submitted:
- Long-Term Future Fund (LTFF): submitted 2026-05-05, status pending.
Application Update with foundational layer completion notice posted
2026-05-06
- Manifund (this application): submitted 2026-05-07