Dresden Goehner · Research & Engineering · Class of 2027

Research · Systems · Evidence

Dresden
Goehner

Building decision systems that make uncertainty visible.

— and audits his own systems first.

Red = Research · Blue = Open Source / Code · Yellow = Nonprofit / Community

Live demo ↗ GitHub ↗ Proof index →

All figures as of November 5, 2026. Every number on this page links to its evidence.

01 · Research

AI Navigation Glasses

Question

What may a low-cost assistive device say to a blind pedestrian — and what must it never say?

A low-cost navigation research prototype that turns traffic-light and obstacle detection into restrained spoken prompts. Research prototype — not a certified mobility aid, and not a replacement for a cane, guide dog, traffic rules, or the user’s judgment.

Process

Safety redesign built around a confidence gate: low-confidence output never becomes an action instruction, a green light is reported as information and never as authorisation to cross, and repeated prompts are suppressed. Community build-and-explain sessions ran in 4 communities.

My specific contribution

I designed the guidance rules, the confidence gate, and the safety case; I built the demo pipeline and ran the community build sessions.

Status

In progress: confidence calibration is the winter P0. The Bosch result is weak and is published unedited; the shell and guidance demo are not an integrated wearable product.

Links

Metrics

  • 211★ / 30 forks (GitHub, 2026-11-05)
  • 85 feedback responses collected 2026-08-26 – 2026-11-04 (helpfulness 3.47/5; 9 of 64 users who tried it still use it; 48% reported inaccurate prompts)
  • 114 automated test functions
  • Bosch traffic-light evaluation: 215 labeled frames, committed-decision coverage 6.6%, including 1 red-read-as-green error — published in full
02 · Open Source

EPSILON

Question

Which assumption breaks first?

A quantitative evidence instrument: state the claim and the rejection rule in advance, run a baseline plus five perturbations (fees / slippage / window / universe / combined), and export a fingerprinted evidence artifact. The browser demo is deterministic illustrative arithmetic, labelled illustrative on the spot; history mode uses server-side real daily data and does not silently substitute on failure.

EPSILON Case 001: baseline return and five stress variants, with the rejection verdict
Fig. — EPSILON Case 001. Baseline −7.75%; stress variants passed 0/5; verdict: Rejected.

My specific contribution

I designed the falsification protocol and built the instrument; Case 001 is my own strategy, rejected by my own rules.

Status

Maintenance and evidence updates since October 2026. The awesome-quant listing (merged 2026-10-19) covers the Assumption Clinic; a listing is not an endorsement.

Links

  • Live demo →
    browser demo live; history-mode endpoints are scaffolds since the platform migration — see status note
  • Code →

Metrics

  • 227★ / 32 forks (2026-11-05)
  • Case 001 baseline −7.75%, stress variants passed 0/5, verdict Rejected
  • Assumption Clinic, April 2026: 26 participants, 5 feedback responses
  • Independent replications: 0 — stated plainly
  • Diamond Challenge 2026, Beijing Pitch Event: Third Prize (regional)
03 · Grid Risk Research

GERT — Grid Extreme Risk Toolkit

Question

Can a grid-risk model be stopped by its own evidence before it reaches a decision-maker?

The Historical Lab is open to the public — 216 ERCOT-native load hours, three windows, recomputable in the browser with an independent checker. Probabilistic forecasting is switched off because the model did not pass its own gate.

Forecast coverage against the preregistered calibration gate
Fig. — Coverage against the preregistered gate. Chart from the PJM companion study; in-progress manuscript.

My specific contribution

I built the toolkit and the Historical Lab, ran both classroom pilots, and made the promotion decision the gate required: no.

Status

v1.5 is preregistered only (per-quantile recalibration, frozen before any v1.5 data run). No v1.5 result exists yet. Limited: pilot counts are organizer-recorded; the survey sample is self-selected.

Links

Metrics

  • 48★ (2026-11-05)
  • Pilots (May & September 2026): 115+ participants, 57 valid survey responses, overall 4.33/5, checker usability 3.89/5
  • v1.4 candidate: pinball skill +0.76 to +0.84, calibration error Q90 0.035 / Q95 0.044 / Q99 0.032 vs preregistered gate ≤0.03 — rejected by its own gate, not promoted
  • 141 backend tests · Release v1.2.0
  • TeachEngineering: formally submitted 2026-10-20, under review
Kaggle · Field Notes

The score is not yours

AI Agent Security (OpenAI / Google / IEEE, team of 5): public 70th (104.575) → private 4,178th (0.00000). The private leaderboard used unpublished, stricter guardrails.

Kaggriculture: peak 6th; the same file scored 2896.5 → 1236.4 → 4168.7 across three segments (final standing 2,278 of 10,148). In a local 84-game experiment, changing only the opponents moved the score by about 31%.

Kaggle results: public to private leaderboard change and the same file scoring three different values
Fig. — Kaggle. Public vs. private standing, and three scores for one unchanged file.

A score is a function of the guardrails, the opponents, and the rules — not a property of the code. That is the same lesson EPSILON applies to backtests and GERT applies to its own model.

Open Source · Code

Open Source

24 merged upstream engineering PRs · 16 repositories — as of 2026-11-05. Selected list; every item links to the pull request itself.

Listing, listed separately

awesome-quant #735 — the Assumption Clinic is listed in the awesome-quant directory; a listing, not an endorsement.

Results · With Denominators

Competition Results

Selected results — score, denominator, year
CompetitionResultDetailArtifact
BMO Round 2 202627/40Distinction, Best in SchoolArtifact →
USAJMOQualifier(record)Evidence on file
AIME 202614/15QualifiedArtifact →
Stanford Mathematics Tournament 2026Algebra 3rdDiscrete 7thAlgebra → · Discrete →
BPhO Round 1 2025–2680/100Global Top Gold, Beijing rank 5Artifact →
SMC 2025120/125GoldArtifact →
LSE Economics Challenge 202597/100Top GoldArtifact →
World Scholar’s Cup Beijing 2025Overall Champion Scholar—Artifact →
ACSL 2025–26 Senior12/12—Artifact →
NYT My List Contest 2025–26Top 25%970 of 4,004Evidence on file
Nonprofit / Community

Organisations

  • Proof to AI

    Co-founder. A nonprofit about how math students survive and thrive in the AI era — “AI Generates. You Prove.”

  • AA (Applied Mathematics and AI)

    Founder, school club; the EPSILON Assumption Clinic ran under AA in April 2026.

  • TeachEngineering

    GERT Historical Lab formally submitted 2026-10-20, under review.

Complete Index

Complete Work Index

  • GERTGrid Extreme Risk Toolkit — Historical Lab and gated forecasting. See §03 above.
  • PJM Polar-Vortex StudyIn-progress manuscript.
  • EPSILONBacktest-assumption instrument. See §02 above.
  • AI Navigation GlassesLow-cost assistive navigation research prototype. See §01 above.
  • FutureOSDecision operating system; one line in the index, maintained separately.
  • the-backtest-that-liedThe signal was real; the strategy was not: IC 0.94 → 0.10.
  • NEEExploratory research code, author’s own disclaimer retained.
  • Collatz Descent FrameworkDoes not claim a proof.
  • 30-project Data Lab13 individual + 17 group; group index 02_Group_Projects.
  • AP Physics Tennis-Racket TheoremEvidence on file.
  • Polymath SocietyLeadership record; Evidence on file.
  • Nexus Club OSPublic concept repo, not deployed.
  • Energy InnovationData-strategy internship; Evidence on file.
About

Profile in One Frame

I work where a claim has to survive contact with evidence. Curiosity becomes useful when it can be shared — written down, checked by someone else, and used to make a better decision than the one I would have made alone.

I build at the seam of mathematics, technology, and people: a navigation prototype that has to respect what it does not know, a backtest instrument whose first rejected strategy was my own, and a grid-risk lab whose model stayed behind its own gate. The tools differ; the discipline does not.

My method is the sentence at the top of this page, repeated in every project: build decision systems that make uncertainty visible — and audit my own systems first. If a number appears here, it links to its evidence; if a result is weak, it is published weak.

One-page CV (PDF) ↓
Portrait of Dresden Goehner
Fig. — Dresden Goehner. Portrait from the project record.