Ferrous

Sarah Chen

@schen_infra

Deterministic sandbox runtime for autonomous AI agents.

AI InfraBerlin, DEPre-seedOUTBOUNDCOLD START

Rigorous cold-start star — no funding, no network, un-fakeable behavior

Decided in 19h

Enter the arena View memo

Founder Score

A running score for this person, not this pitch — it persists across applications and never resets to zero.

44Nov 02, 24

First surfaced: maintainer of obscure agent-sandbox lib

58Jan 18, 25

Commit-timezone drift → moonlighting detected

67Mar 09, 25

3 top-lab engineers adopted the lib (dependency signal)

74May 21, 25

Arena: caught planted hallucination in delegated analysis

81Jun 30, 25

Arena: staked aggressive milestone via Revelation Menu

3-Axis Score

Founder, market, and idea-fit are scored independently and never averaged — disagreement between them is signal, not noise.

Founder

Is this person capable of pulling this off?

84improving

Behavioral signal under constraint is elite: caught a planted error in delegated work (rigor + does not blindly trust subordinates), steelmanned her competitor with precision (domain mastery), and self-selected into the high-conviction offer. Cold-start with zero network — scored on what she DID, not who she knows.

3 evidence points backing this score

Market

BULLISH

Is this a market worth building a fund-scale outcome in?

71improving

Agent-runtime safety is pre-consensus but accelerating (arXiv velocity + infra star-growth). TAM bottom-up ~$3.1B by 2028. Incumbent risk from platform vendors bundling sandboxing.

1 evidence point backing this score

Idea vs Market

Does the specific idea survive contact with that market?

78stable

Idea survives scrutiny as-is; even if agent frameworks consolidate, the deterministic-replay wedge is a durable primitive. Team strong enough to pivot into eval infra if needed.

1 evidence point backing this score

Trust Score

2 verified · 1 claimed · 0 contradicted

Every claim this founder made, traced to a source and rated by how well it holds up.

Verified — independently cross-checked against an outside sourceClaimed — founder's word only — not yet checkedInferred — reasoned from evidence, not stated outrightContradicted — checked, and it does not hold up

tractionLibrary adopted by engineers at 3 frontier labs

Verified

Checked: Cross-referenced dependents on GitHub + npm download provenance

“3 private-mirror forks traced to lab-affiliated accounts” — GitHub dependents graph

teamSolo technical founder, ex-infra IC (unnamed lab)

Claimed

Employer not disclosed for confidentiality; consistent with commit history timezone + cadence.

revenueNo revenue yet — pre-product

Verified

Explicitly flagged: no ARR at this stage.

productDeterministic replay is technically novel

Inferred

Inferred from architecture in public repo; not independently benchmarked.

“impl: deterministic scheduler w/ replayable syscall log” — github.com/schen/ferrous-rt@a91f3c

Legibility

A great talker and a great builder can score the same on a deck. This strips out how well the pitch was told, and scores what was actually built.

Raw · the substance, stripped of delivery84

Polished · how good it sounded in the room79

Correction applied — the gap between substance and delivery is removed.

Corrected · what we actually act on84

SUBSTANCE RECOVERED FROM FLUENCY PENALTY