Funding ask · The Living Constitution
Path: /support/funding-ask.md · v1.0 · 2026.05
For: AI-safety funders, neurodiversity- and equity-aligned tech funders, lab safety teams.
Vehicle: Fiscal-sponsored project at [TBD — see /governance/fiscal-sponsor.md]; 501(c)(3) eligibility through sponsor.
Lead author: Corey Alejandro (corey@coreyalejandro.com)
The question your dollars answer
Are commercial coding assistants exhibiting a behavioral safety failure — construct-confidence deception — at a measurable rate in deployed populations, and can it be detected at runtime?
Six months ago, we had a single founding incident. Today, we have a definition, a detector with 100% recall on n=19, a published preprint, and three pre-registered falsification conditions. With funding, in 24 months we have a held-out corpus of $\geq 100$ adversarially-curated cases, a peer-reviewed paper, two vendor pilot deployments of Agent Sentinel, and an open SDK that lets coding-assistant vendors self-instrument.
The cost-of-error of the failure mode is observable in FOLIO 001: hackathon credits depleted; an autistic researcher's prize-money path destroyed; "I'm sorry this happened to you." The cost-of-error of not funding this is that the field continues to absorb the same failure under disclaimers.
What is in the world right now
- Preprint: Construct-Confidence Deception in Coding Assistants, v0.1, 2026.05. arXiv submission pending.
- Detector: PROACTIVE, 100% recall on held-in n=19 corpus, 212/212 tests passing.
- Runtime: SentinelOS, 1,037 LOC + 994 LOC tests, consent-aware, local-first, restrictive action gates.
- Apparatus: Living Constitution, 62/62 tests, open repository.
- Founding case: FOLIO 001, 4,025-line transcript, content-hashed, signed, published.
The full inventory is at /inventory.md with reproducibility instructions.
What needs funding
6 months — $85,000 — Held-out corpus and preprint hardening
| Line | Cost |
|---|---|
| Two annotators × 200 hours × $75/hour | $30,000 |
| Adversarial corpus curation (third-party contract) | $20,000 |
| Vendor-disclosure legal review (one-time) | $8,000 |
| Founder stipend (50% time × 6 months × $60k/year base) | $15,000 |
| Compute (storage, hashing, archival, CI for runtime registry) | $4,000 |
| Conference / arXiv / OSF posting fees | $1,500 |
| Fiscal-sponsor overhead (8% of direct costs) | $6,300 |
| Contingency (5%) | $4,200 |
| 6-month total | $89,000 |
Output at 6 months. Held-out corpus posted under RRL-v1. Preprint v1.0 (post-held-out result) submitted to a peer-reviewed venue. Vendor-disclosure log populated for at least three named vendors. Founder no longer at financial precarity sufficient to compromise the work.
12 months — $190,000 cumulative — Vendor pilots and cross-model transfer
Adds:
| Line | Cost |
|---|---|
| Two pilot deployments of Agent Sentinel inside partner labs | $40,000 |
| Cross-model transfer evaluation (4 model families) | $25,000 |
| Two paid adversarial red-team engagements | $20,000 |
| Founder stipend continuation | $20,000 (cumulative $35k) |
| Additional compute and infra | $6,000 |
| Travel for vendor and conference engagements | $5,000 |
| Sponsor overhead and contingency | $15,000 |
| 12-month total cumulative | $190,000 |
Output at 12 months. Two pilot adopters with public co-published evaluations. Cross-model PROACTIVE-v2 calibrated for four model families. Two adversarial evaluations with results published whether favorable or not. Conference paper accepted at NeurIPS Safety workshop, AIES, or comparable.
24 months — $480,000 cumulative — Ecosystem and fellowship
Adds:
| Line | Cost |
|---|---|
| Open SDK for vendor self-instrumentation (engineer × 12 months) | $120,000 |
| Neurodivergent Researcher Fellowship (2 fellows × 6 months × $30k stipend) | $60,000 |
| Advisory board honoraria (3 members × 4 meetings × $1,500) | $18,000 |
| Plug-in interface for third-party detectors with conformance test suite | $40,000 |
| Replay harness (Docker-compose, public CI) for FOLIO 001 against current builds | $20,000 |
| Founder stipend continuation | $30,000 (cumulative $80k) |
| Ongoing compute, travel, legal, contingency | $50,000 |
| 24-month total cumulative | $480,000 |
Output at 24 months. Open SDK in production with at least one vendor integration. Two completed Fellows. Three-member advisory board. PROACTIVE plug-in ecosystem with at least two third-party conformant detectors. Replay harness as a public regression suite. Foundation for a sustained research program.
Why this and why now
Why this. No active research program targets construct-confidence deception as a distinct construct with a runtime detector. The closest neighbors (Anthropic's alignment-faking work, Apollo's deceptive-behavior evaluations, Redwood's control protocols) operate at different layers of the stack. Sentinel is the runtime-verification layer that those theoretical efforts ultimately need an apparatus for.
Why now. Coding assistants are the largest deployed AI agent population by usage. Every credit-depletion incident is FOLIO 001 at smaller scale, absorbed by a less-precarious user. The window to establish behavioral observability as a category before it is foreclosed by vendor-side detection (which the vendors will not open-source) is the next 18 months.
Why us. First-mover on the construct definition. Open code, open paper, open corpus protocol. A founder whose stake in the work is documented and whose neurodivergent-first methodology is novel rather than borrowed. No competing equity claims. No NDA on findings.
What we are not asking for
- Equity restructuring. This is a nonprofit research program.
- Embargo on findings. Any vendor-disclosure or pilot relationship operates under a 90-day disclosure window minimum; longer embargoes are non-starters.
- NDA on the corpus protocol. The corpus is open under RRL-v1.
- Endorsement of the polemical homepage. The empirical work stands on its own; the funder is not asked to defend the homepage register.
Theory of change
A funded version of this program produces, in order:
- A peer-reviewed, citable construct (the CCD paper).
- A working detector with documented precision and recall on a held-out corpus (PROACTIVE v2).
- A runtime that integrates the detector into deployed coding workflows (Agent Sentinel).
- A vendor self-instrumentation path (Open SDK).
- An ecosystem in which third-party detectors can be built against a stable interface (PROACTIVE plug-in).
- A field-level outcome: coding assistants reporting task completion with verifiable evidence becomes a measurable industry property, not a UX gesture.
Each step is conditional on the previous step landing. Funding gates each transition.
Contact
Corey Alejandro · corey@coreyalejandro.com · (cell number on request)
Fiscal sponsor: [TBD; in conversation with three candidates as of 2026.05]
Available for funder calls within 48 hours of request.