Project Summary

$25,000 to run a 4-month independent-user acquisition and evidence-documentation sprint for BDC Bridge — a live bounded-evaluation service for AI workflows running at bridge.bdc-hive.com.

The goal is to recruit AI teams with real workflows, facilitate their submission of Bridge packets, document their deployment decisions and subsequent outcomes, and recompute calibration based on this third-party evidence.

This is not "user growth" for its own sake. These users represent the critical missing evidence required to test the H1 Tier 3 calibration hypothesis: whether Bridge’s evaluator remains statistically calibrated when transitioning from owner-operated systems to independent, external teams.

I am Danil Avakumov, operating solo from Kokand, Uzbekistan, with no legal entity. I have spent 18 months building BDC Bridge and the underlying research line.

The scientific spine so far:

• 34 measured gates: 32 PASS / 2 preserved FAIL.

• Six bounded mechanisms cooperatively assembled.

• External chaotic validation: 6 of 12 architectural laws confirmed across two LLM providers and two languages on real-world data chunks.

• Current calibration baseline: Tier 2 PASS, Brier = 0.03178, ECE = 0.17 (n=5 admissible cases).

The Gap: Those 5 cases are owner-operated. Bridge is currently calibrated on systems I control. It has yet to demonstrate calibration on systems run by teams I do not control. This grant is designed to close that empirical gap.

Manifund is the ideal fit because this is "uncomfortable" evidence work. The most valuable outcome may be negative: Bridge may over-abstain or miscalibrate on independent teams. Per BDC discipline, such a failure will be published as a preserved-failure result rather than reframed as success.

What are this project’s goals? How will you achieve them?

Goals (4 months from grant start)

1. Recruit 5 independent teams with production-grade AI workflows.

2. Document 5 complete cycles: input packet, Bridge recommendation, team’s deployment decision, and the observed outcome (deployed / refused / deferred / failed).

3. Recompute H1 calibration on a combined pool of 10+ admissible cases to assess statistical significance.

4. Publish a Proof Pack: a claim-by-claim atlas showing "before/after" calibration metrics.

5. Preserve Failure: If calibration degrades, publish the findings as a canonical FAIL entry. No silent rewrites.

How

• Months 1-2 (Outreach): Direct outreach and warm-intro brokering targeting solo founders, AI agencies, and safety-adjacent labs.

• Incentivization: Teams receive an honorarium ($1,500) for the documentation cycle, not for positive results. Contradictory outcomes (where Bridge was wrong) are explicitly valuable.

• Onboarding: Utilizing the live companion repo github.com/malishomen/bdc-bridge-passport which contains templates and FAQs to minimize friction.

• Month 4 (Analysis): Final calibration recompute and publishing the "Integrity Review" report.

How will this funding be used?

Budget Breakdown ($25,000 total):

• Participation Honoraria ($7,500): 5 teams × $1,500. Paid for packet preparation and outcome reporting.

• Outreach & Brokering ($4,000): Warm-intro brokering, qualification, and follow-up operations.

• Compute & Infra ($4,000): Evaluation workload, storage, and monitoring during the sprint.

• Operator Living Expense ($8,000): 4 months full-time at Kokand cost-of-living.

• Public Proof Packaging ($1,500): Calibration delta report, preserved-failure registry, and claim atlas update.

Why $25K? A smaller grant (e.g., $5K) would fund only 1-2 cases, which is statistically insufficient to close the Tier 3 calibration question. $25K is the floor for a meaningful recompute (n=10+).

Team & Track Record

Team: Solo (Danil Avakumov). No entity, no co-founders. Track Record: 18 months of independent execution.

• Scientific: 34 gates, 32 PASS/2 FAIL. Calibration baseline (Tier 2) is already live and documented.

• Product: BDC Bridge is a functional service with a distroless worker, hash-locked dependencies, and a public API.

• Discipline: I adhere to a preserved-failure rule. My record includes canonical FAILs (R10, R15) that were never "pivoted" away.

Verification:

• Live Service: bridge.bdc-hive.com

• Health: .../v1/health

• Repo: github.com/malishomen/bdc-bridge-passport

Failure Modes & Outcomes

1. Low Recruitment: If only 3-4 teams sign on, the calibration surface still improves, but I will report "insufficient case count for Tier 3 closure."

2. Calibration Decay: If Bridge fails on independent teams, the result is published as a "Generalization Failure." This is a highly valuable finding for the AI safety community regarding bounded-evaluation limits.

3. Statistical Inconclusivity: If the delta is too small to call, the result is published as "signal indeterminate."

Commitment: No scope creep, no silent rewrites, no narrative-fitting. Failure is documented with the same rigor as success.

Funding History & Disclosures

• Last 12 Months: $0 (100% self-funded).

• Parallel Applications: LTFF (submitted 2026-05-02, $40K), Emergent Ventures (planned), Foresight (planned).

• Transparency: If multiple grants convert, I will accept the smallest viable non-overlapping combination and disclose status to all parties.