Track Record

The numbers. Code base, test discipline, publishing cadence, and a kill rate that does the persuasion.

§ I   Platform scale

The OpEx figure is unusual enough to warrant a sentence. Inference happens primarily on a local llama.cpp server on a single workstation. Cloud providers are used elastically for tasks that benefit from larger models or higher throughput.

The dominant marginal cost of running the platform is the operator’s electricity bill.

§ II   Publishing cadence, last 30 days

A&F operates as a daily newsroom; its monthly artefact count is dominated by per-story lens variants and is reported separately in its own diary view.

The point of this table is not volume — volume is cheap. The point is that one operator with this platform is publishing across four register tiers simultaneously, and the standard the operator applies does not vary across them — the mechanism differs by arm, the standard does not.

§ III   Kill rate as credibility signal

The Institute’s corpus carries an active killed-signals ledger. As of the most recent audit:

• Corpus-wide kill rate: 17.2% of signals that entered active validation have been moved to NULL, NOISE, or another retired tier (291 of 1687).
• Trading-battery kill rate: approximately 44% of candidate trading signals failed the forward-out-of-sample gate and are retired.

Both numbers are stable across audit waves. Both are higher than what a content business would want to publish.

A research programme that confirms everything is confirming nothing. A kill rate of 17.2% on the broad corpus, and near 44% on the highest-stakes trading-battery tier, is a load-bearing claim about discipline.

It is also the reason the killed-signals page on the Institute site is structurally as prominent as the confirmed-signals page. Killing is not the failure mode; it is the product.

Judge the standard from the artefacts, not the description. The cosmic-ray–cloud–agriculture convergence paper is a confirmed cross-domain finding the engine produced and published in full; the killed-signals ledger is the same engine’s record of what it could not confirm — moon-phase equity returns, Schumann-resonance biology, and the rest, each retired with its reasoning shown. Both came off the same line. The discipline is what tells them apart.

§ IV   Engineering record

The platform is held to a documented set of architectural invariants and recurring bug-class anti-patterns. Per-process startup runs a suite of invariant assertions; commits run a pre-commit pattern sweep against known bug shapes. Both fail loudly on drift.

Sample of currently-enforced invariants (one-line summaries; the implementation is internal):

• Validator results must flow through a single canonical helper; raw filesystem writes from validators are blocked.
• Subprocess calls inside agent tools must use a hardened wrapper with scrubbed environment, hard timeout, and binary allow-list.
• Critical JSON registries (agents, state, tools, cloud models) must parse cleanly and contain no smart-quote drift on every process boot.
• Verdict-tier coherence: a paper cannot cite a CONFIRMED_WEAK signal as CONFIRMED.
• Audit-pipeline gate: a draft that overstates a battery-failed signal is rejected at write time.

These are not aspirations. Each line is a check that has fired in production at least once.

§ V   The honest framing

This platform is built and operated by a single architect. That single fact reads two ways, and both readings are correct.

Read one way, it is the asset. The operator and the engine are not separable — the engine is what the operator built, and the operator is the standard the engine runs to. They are a working pair, and that pair is what holds the value. Splitting them is what destroys the thing being valued.

For a licensing or partnership counterparty, it is a concentration risk, and pretending otherwise would be dishonest. The mitigations are concrete rather than rhetorical: the architecture is documented, the operator maintains runbooks for the recurring operational paths, the invariant and pattern-sweep layers mean a second engineer inherits guardrails rather than folklore, and a code-escrow arrangement is available where the engagement justifies it. The risk is named so it can be priced; the mitigations are named so it can be managed.

The argument is not that one operator is better than fifty researchers. It is narrower and more defensible: one operator with this engine sustains a publication standard, a kill-rate discipline, and a cross-domain reach that a fifty-researcher institution does not — at a marginal cost that is a rounding error against their salary line. What that scales into, with capital and additional headcount, is the conversation at Partners.

Drafted with AI assistance under operator supervision; substantive claims are operator-authored or operator-approved.