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