You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
A quick note on AI detection: I am a non-native English speaker from Poland. I wrote the 152,000-line codebase, the 82,000-line specification, and the original technical dossier entirely myself. However, I used an AI assistant to translate and structure the text of this specific grant proposal into readable English. The prose is AI-assisted; the verifiable facts, the architecture, and the live system are 100% human. Please evaluate the engineering, not the translation.
Three distinct tools tackle three different problems, but they all run on one absolute rule: the machine is forbidden to claim anything it cannot justify.
AK-MASTER (170 live routes): This module forces constraints before any writing happens. It audits narrative integrity. Did the actor actually pay a real price for a benefit? Is a fictional literary device trying to pass as a factual event?
Mr. Helix (11 live routes): The investigative engine. It checks if sources are truly independent and tests competing explanations against each other. It then assigns a hard evidence rating to show exactly how well supported a claim is.
HELIX (7 live routes): Built to catch LLM hallucinations. It does not rely on a model's "opinion" to validate text. Instead, it uses strict mathematical formulas to determine if an output is fabricated.
There are zero imports between these three tools in either direction. This is a deliberate architectural decision.
Why this isn't just another AI wrapper: The standard industry fix for AI hallucinations is fundamentally broken. Developers usually just prompt a second model to evaluate the first one using prose. That inherits the exact same flaw it's supposed to fix.
My architecture does the opposite. I trust the model only to read and label. Every consequence of those labels is calculated by explicitly tested Python functions. The model is severely restricted—it is only allowed to return four specific values on a 0-3 scale. If it tries to sneak in its own calculated opinion, the engine throws an exception.
I enforce this rule against my own work, too. If a numeric value exists in the code but doesn't trace back to a specific line in the specification, it is a defect. In a recent codebase audit, I found four confidence thresholds that weren't properly specified. I deleted them immediately.
The goal is to fund the compute (Vertex API) and infrastructure needed to keep this massive system alive for an independent technical audit. A document ages; a command does not. Every metric I claim can be verified by a reviewer running a single command against the project (measured today, August 21, 2026, commit 1feb0dd):
• AK-MASTER (Canon blocks / enforced by code / blocking gaps): 14 / 13 / 0 Command: python3 -m engine.coverage_validator
• Mr. Helix (Capsules / states declared / unreachable): 99 / 583 / 0 Command: python3 -m mr_helix.engine.capsule_contract_audit
• HELIX (States declared / unreachable): 922 / 0 Command: python3 -m helix_pipeline.engine.helix_contract_audit
• Total System Tests (zero failures): 6,535 Commands: pytest -q and pytest helix_pipeline/tests -q
What "zero unreachable states" means: every state explicitly named in the 82,801 lines of specification can be returned by the code, verified programmatically. What it does not mean: that the state is returned under the right conditions, and it is not a certification under load.
The codebase is actively growing every single day—just today, I built and integrated an automated form-filling agent with 21 new tests, bringing the total architecture to 152,826 lines of Python across 740 files. However, the system runs on exhausted personal funds and cloud credits that expire on September 21, 2026.
The funding will be used strictly for:
API costs (Vertex API / Google Gemini) to run real grounded investigations instead of throttled simulations.
Cloud hosting to keep the production site live (web-production-b540f.up.railway.app).
Buying the necessary runway to open the repository for technical due diligence by external researchers, transforming this from a solo endeavor into an audited public protocol.
I am a single developer. While development is conducted with the help of AI tools, the code, the 82,801 lines of specifications, and all architectural decisions are strictly mine. No team, no external funding, no revenue.
My background involves writing 27 documentary and investigative books, which is where my habit of systematizing massive bodies of evidence comes from. I have spent the last three years building this framework alone in complete isolation. It ships: it is deployed, running, and has 189 live routes, proving my capacity to execute complex engineering.
Investors demand honesty about risks. Here is exactly what is missing from my project right now—the true failure modes:
No independent verification. Every proof comes from one repository owned by one person. This is the conclusion of my own audit, not a confession under pressure.
Zero external users in three years.
Zero revenue. The subscription logic is built and connected, but there are no paying users.
No live Red Team against a real model. Mr. Helix has 54 of 54 PASSes, but that is a simulation, not a production run. Thus, there is no production certification.
The 4th HELIX vector (MARKET) is intentionally unbuilt because its spec is not stable.
Furthermore, the system is fragile because it is honest. It has already caught itself failing five times:
Invented thresholds: 4 confidence bands existed in the code but not the spec. Deleted.
Temporal hallucination: The classifier declared real events as fiction by substituting its training cutoff for "today"—the exact error it was built to catch.
Contract debt: 179 and 127 states were unreachable while the first audit reported full coverage. Now closed to zero.
Fabricated codes passed the canon gate.
4 of 9 bans failed to block actions protecting the boundary between investigator and literary engine.
Zero. No grants, no investments, no revenue. The work has been 100% bootstrapped using personal retirement savings. The only external support was two small promotional cloud credits ($300 Google Cloud and a Gemini trial credit), which are nearly exhausted and expire on September 21, 2026. Neither was a grant; they were standard sign-up credits. This is my first application for funding.