Fitchhardt

Throttiq exists because when an engine is physically modified, a calibration of the engine’s electronics must take place. This electronic information relied on by the engine to operate is stored in the Electronic Control Unit (ECU) of the motor vehicle. Calibrating the engine involves moving the motor vehicle onto a dynamometer to then iteratively adjust values for fuel, ignition timing, boost, and cam timing in pursuit of individual priorities. The cells outnumber the hours anyone pays for, so, traditionally, no set is finished.

Throttiq is fundamentally an experiment in applying virtual harness architecture with real-world solutions. Throttiq v0 focuses precisely on understanding the challenges and limitations faced when compiling data from physical environments. While variable corrections are currently out of scope, Throttiq v0 is a recording product, designed to store raw, unaltered frames gathered from the car’s Electronic Control Unit via a CAN bus connection. Once the data is recorded, engine behavior, trust per sensor channel, and the delta between two sessions are compiled to provide the engine calibrator with a post-calibration report on integrity. Throttiq v1 will include the option for the calibrator to receive active analytical feedback during calibration.

As it stands, 19 October 2026

• All five stages run on a virtual CAN bus: record; bundle, integrity-checked; report; the calibrator’s page; comparison of two sessions.
• We lose no frame up to about 3,000 frames per second; the ECU’s own rate is about 895.
• First code: 7 August 2026. First complete run, live: 29 August 2026.
• Fault drill, 12 to 14 September 2026: 17 failure promises in the runbook, 11 tested.
• Hardware: a chain of 18 links; bill of materials verified; specified September 2026. Ordered: no.
• Pilot: UMS Tuning, Mesa, Arizona; a Haltech Elite in a Nissan B14 Sentra. The date is not set.
• Prediction 1, written 15 September 2026: at least 1 flag on session 1. Result: —none yet. The predictions.
• Customers: 1 shop, the pilot. Others: 0.
• Absent: variable correction; any ECU other than the Haltech Elite; a price; revenue; a business model, deferred on purpose until the wedge is tested.
• v0 exists: the report after a session. v1, feedback during a session, is planned and not built; date unset.
• Capital: the two founders’ own. Outside capital: 0. Loans: 0.

Erik’s part

Ventures executed by Fitchhardt are never solo operations; while Erik Fitch builds the product, Manuel Eberhardt is responsible for the company’s legal affairs. Their latest start-up is called Throttiq, and it provides engine calibrators with the helping hands to avoid lying sensor data and save time. The company has been an early experimentalist in AI, and it actively pursues new opportunities in emerging harness research. By orchestrating context management among three participating forces, a recording agent, a simulation agent for bench testing, and a safety agent, Erik builds the recording harness used by engine calibrators. As of 17 October 2026, this work has resulted in 82,592 lines of test-driven development and a 94.8% kill rate on 7,024 mutants in testing, which keeps an eye on 25,531 lines of Python product code. The transmit path in rule 10 is blocked on real hardware; its one exception is a verbose override on the bench, never in the car.

The setback

Often in test-driven development, the bugs that crawl in, to die silently, are the most tedious. Often, finding them is so sufficiently challenging that they remain entirely undetected by purely traditional testing. Thus, it was clear that mutation testing needed to be a founding standard for Throttiq. Mutation testing means deliberately breaking a codebase to examine if its tests ring when something actually breaks.

First introduced at the beginning of development in July 2026, the mutation framework had a rocky start. The reason the first three weeks looked and sounded good, but ran badly under the hood, was because the mutation gate was an unreviewed nightly run executing a job that couldn’t fail. The problem was that its 9 hour run had a 60-min cap, and when it exited prematurely, the logs remained, but the final record went up in smoke.

Since it’s correct implementation, the mutation framework has proven itself as highly effective. Now the mutants are sharded; only the most relevant are executed during a pull request, and any new surviving mutant fails the merge. The first honest sample had a fifth of the mutants surviving. Three weeks of sweeps later, mutation testing ran with a 94.8% kill rate.

What changed

Before Fitchhardt was officially inaugurated, Erik and Manuel were already pursuing the thought that he who desired to invent had to let go of what function existing systems were believed to have and, through thought, combine them to create entirely new systems. Indeed, to become inventors was one of the earliest things they agreed upon during their first conversations in 2018. Thus, as the internet had already been commonplace by their time, artificial intelligence was the major change in perspective. This societal, though individually directable, force symbolized the beginning of a runaway reaction, characterized by what are eight years later, two founders gaining firsthand experience in what it means to manage a professional team. Artificial intelligence was the major change for Fitchhardt, because what used to be only possible by going to Stanford and being picked out by venture capitalists was now possible from a village in the south of Germany.

Method

In building systems professionals rely on, it is vital to focus first on the systems of which the emerging systems are progenitors. Fitchhardt pursues in Throttiq local optimization by compartmentalizing the larger force into three subsystems: the simulation agent, the safety agent, and the recording agent. While Erik periodically reviews and adjusts their trajectory, their harnesses run locally on a rented server in the cloud, providing them with a nearly all-round uptime. For reviewing and adjusting trajectories, he has established a ticket system that guides the agents in their work and, additionally, report systems too that give them a format to ask clarifying questions and pause before making major changes.

In creating Throttiq, Fitchhardt doesn’t pursue an innate trust in the code written by its agents. Thus, through continuous integration and mutation testing, the rules that gate pushed code have tightened over time. The rachet in the mutation tally, for example, only allows a kill rate equal to or higher than the previous best. Or even more so exemplary, the rule exists that when an incident occurs, the agent must themselves write a new one-line-long rule to, in the future, circumvent another faux pas. Additionally, only one agent may edit an individual file at a time to prevent two from overwriting each other. As a result, the codebase has grown to 25,531 lines in product and 82,592 lines in test code as of 17 October 2026, and the mutation tally of September 2026 stands at 7,024 mutants, of which 94.8% are killed.

The handbook.

Handbook

11 operational rules, written after each incident, never before one.

1. Every agent output is replied to. No reply means the agent has halted.
2. A decision binds and carries a machine-checkable pass or fail. Advice is not a decision.
3. A report contains tool output and no interpretation.
4. The prediction is written first; the measurement comes second.
5. A file has one owner at a time: one agent.
6. A refactor leaves the test files frozen. A behavior change ships with its test diff and the reason.
7. Hand-transcribed data is provisional until a live measurement promotes it; a document never does.
8. Tool versions are exact and locked.
9. The advisor’s mistakes are logged; the catch rate is the metric.
10. The transmit path is blocked on real hardware. The exception is a verbose override, on the bench only; in the car, never.
11. A flag of not clean with the list of frozen or stale channels is success.

The catch list

• A mutant turned a continue into a break inside recover_frames and survived a green suite at 99.95% coverage, because every test built its input from the same encoder and that encoder never produced an interior blank line; on a real one, every frame after it would have been dropped with the whole-file check still true and no error raised. It was documented on 28 August 2026 on issue 45.
• A driver call corrupted memory and the process died later, after a normal return.
• The test suite passed on every machine we tried, and the pass was not valid.
• After a new version of the mutation tool, not ours, every change failed CI; rule 8 came of it.
• The drill of 12 to 14 September 2026 found the runbook and the code disagreeing once; we fixed it.
• On 14 August 2026 a mutation harness that edited the live tree was killed mid-sweep and left one character changed in src/: fsum had become sum. Every static gate stayed green; 4 tests caught the distinction. The cause, in the agent’s words: two subagents colliding over a shared scratchpad path. We verified it by diff and hash, not by the report.
• The gate that could not fail, told above under the setback: from 30 July 2026 a nightly job whose exit code was discarded; its only green run on 14 August 2026; 10 nights canceled at the cap from 15 to 24 August 2026; issue 45 filed 24 August 2026; the one readable run, 4,502 mutants generated, 395 executed, 86 survivors, 21.8% of the sample; the fix in PR 50 on 25 August 2026.
• Rule 9’s log holds more than one of the advisor’s errors.

Predictions

A prediction goes here before its test; the result stays blank until the measurement exists, and a miss stays on the list.

Since Erik and Manuel have begun thinking, they have been collecting solutions to inconveniences and major system flaws alike. As of 2026, the database spans a measurable 50/50 split of over 100 software and hardware solutions.

While they had small but meaningful projects beforehand, in January of 2025 their project-driven personalities drove them to pick an idea from the list, and so they created BeamIdea — the platform that understood employees’ possession of judgment in companies and could route it when needed.

The MVP was hand-coded at the time and able to fulfill its function by making use of AI text embedding and vector databases. It worked by collecting and compiling as much information about an employee’s professional experience as possible to then embed all text into a vector and use cosine to find the smallest angle to the vector of a case needing judgment. Once this highly efficient process was completed, an LLM could then reason the match to the employee and point them towards how they could help.

Reaching out to companies with an easy-to-experience demo of the MVP, BeamIdea was to be sold as a B2B SaaS solution, distributed over the Apple Business Manager. Outreach was at the time primarily focused on companies with about 1,000 employees, because for the product to be effective, enough judgment needed to be available in the company for a case needing judgment to have a high chance of receiving a great match.

BeamIdea retired in June 2026 after hundreds of individually written emails, because no iteration made in outreach led to a pilot. Lesson: how difficult it is for two founders, short of a network, to sell a high-touch software product to a large company with little evidence.

App Store release: 2025 · emails: hundreds · pilots: 0 · customers: 0 · revenue: 0

Also

• 2 Jul 2026. Erik’s talk ended a school’s AI seminar: 90 minutes.
• 26 May – 6 Jun 2026. Erik pitched BeamIdea to investors and founders in the Bay Area and was invited into one Stanford Continuing Studies class.
• 9 Oct 2025. Erik’s part in The Human Movement began on X; his earliest dated post is from this day: x.com/erikfitch_.
• Apr 2023 – Feb 2025. Erik wrote a summaries app for self-improvement in Flutter, every line of it; LLMs ended it.
• Oct 2022 – Apr 2023. PodHead AI: a camera on a mast for a solo sailor, aimed at nets and floating debris; we wrote a simple recognizer in Python.
• 2022. Venture 1: a website that taught riding on push bikes by the German method, for the US market; it lived for months.

Ledger