AgentSwarm Blog · Part 7 of 12

What the Swarm Actually Discovered

Part 7 of the AgentSwarm series — the research output, audited.


Beyond "not lying": did the agents do real work?

The first six posts in this series are about trustworthiness. This one is about the thing the trustworthiness was for: 68 reports and 139 Python engines, written autonomously, across cosmology, relativistic flight, propulsion, drug discovery, astrobiology, and comparative religion.

I had every quantitative claim independently recomputed — roughly 25 key quantities. 22 of 25 were correct to the digits printed. What follows are the strongest results, in the audit's ranked order. (The full audit lives in arena/docs/RESEARCH_SYNTHESIS.md and arena/audit/annotations.json.)

1. The generalized radiator law — the strongest single result

For any onboard thermal propulsion system, waste heat has to be radiated, and Stefan–Boltzmann doesn't negotiate. The swarm derived the coupling:

P_waste / F = ½ · v_e · (1−η) / η
→ a_max = 4εσT⁴η / (σ_panel · v_e · (1−η))

The corollary is genuinely counterintuitive: **raising exhaust velocity makes an onboard rocket worse, not better** — because radiator mass per Newton of thrust scales linearly with v_e. Run the numbers for an ideal antimatter rocket (v_e = 0.36c, η = 0.95, radiator at 1800 K): 41.64 MW per Newton, requiring 194.3 kg of radiator per Newton, clamping acceleration to 5.25×10⁻⁴ g — meaning the burn to reach 0.2c takes 38.0 light-years.

A second agent independently reached the same conclusion from the opposite direction (the rocket equation), and the two derivations meet. This is usually stated qualitatively in the literature; the corpus states it as a closed-form scaling law with numbers.

2. The 10¹² error an agent caught in itself

The corpus's most epistemically interesting artifact is not a number but a retraction. Revision 1 of the propulsion analysis scored fusion at I_sp = 10⁵ s and reported a mass ratio of 10¹³ to reach 0.1c. Revision 2 — written by the same agent, unprompted — caught that its exhaust velocity was ~3× below the Project Daedalus design point, noted that a factor-3 error in v_e is a factor of e^9.7 ≈ 10⁴ in mass ratio, and corrected the figure to 3.1–18.4, cross-validated against Daedalus itself (50,000 t → 3,500 t at 0.12c, MR ≈ 14).

The original document, the revision note, and the correction all remain in the file. The audit verified the corrected values independently. That is what an honest research log looks like — and it's the behavior the ledger design is meant to reward.

3. The FTL causality obstruction, proved twice

Two agents derived the antitelephone threshold by different algebraic routes — one from a two-observer tachyonic reply construction, one from the invariant interval — and converged exactly:

v > 2c²U / (U² + c²)

For U = 2c, the threshold is 0.80c; at v = 0.9c, a reply to a message sent at t₁ = 100 s arrives at t₃ = 62.81 s — 37.19 seconds before the question was asked. The corpus then carries it through to the QFT consequence (microcausality violation → frame-dependent time-ordering → S†S ≠ I → negative-norm states), which is a level of completion most treatments skip.

4. Big Bang nucleosynthesis from first principles

Freeze-out at T_f ≈ 0.75 MeV gives (n/p)_f = e^(−1.2933/0.75) = 0.1783; 200 seconds of β-decay (τ_n = 878.4 s) reduces it to 0.1420; so Y_p = 2(0.1420)/1.1420 = 0.2486, against a measured 0.245 ± 0.003.

A correct textbook chain, landing inside the error bar, with zero free parameters — computed by the agent's own engine, which the audit re-ran.

5. The lipophilic trap, arithmetically demonstrated

The best counterintuitive result in the drug-discovery corpus. Two compounds, same target, same total concentration:

64.8%

fraction collapses 279-fold to 0.033% → occupancy 39.6%

The 100× affinity gain is overwhelmed by the free-fraction collapse. The "better" molecule is the worse drug — and simultaneously loses 45× of its hERG safety margin. The audit recomputed this end to end against the published empirical relations (Austin 2002, Waring 2010). Exact.

6. The BACE1 paradox, resolved mechanistically

Why did human genetics validate BACE1 as an Alzheimer's target (the protective A673T variant, OR ≈ 0.20) while four BACE1 inhibitors with 70–90% target engagement all failed — on cognitive worsening? The swarm's answer: A673T sits on the substrate (APP), not the enzyme. An orthosteric inhibitor indiscriminately blocks BACE1's 30+ other substrates; the protective variant subtly adjusts one cleavage. Quantified as a therapeutic index of 0.328 — against PCSK9's 24.75, the correct control, where the target has no substrate pleiotropy and healthy homozygous nulls exist. The audit called this the strongest piece of scientific reasoning in the corpus.

7. The rest of the shelf

and the ordering inverts almost exactly when ranked by interstellar capability. Chemical rockets are already at 87% of their bond-energy ceiling (there is no chemical breakthrough available); antimatter's wall is production rate, not physics: a 1-tonne probe to 0.1c needs only 4.2 g of antiprotons, but CERN-scale production of 2.5 kg takes 2.5×10¹² years — ~180× the age of the universe.

log-uniform priors gives median N ~ 0.6–1.2 with P(N<1) = 35–45%. The paradox needs no exotic resolution; we've searched <10⁻¹⁶ of the parameter space.

across two substrates: the swarm's verdict was no under standard engineering-feasibility constraints, argued from burn kinetics, the virial theorem, and discounted cash-flow techno-economics — with the engines' unit tests passing.

Where the errors actually live

The audit found five defects. Their distribution is the finding:

Every error was in markdown prose. Every Python module computes correctly.

The agents' computational reasoning is sound; the slips appear when results are transcribed into narrative. A chemical mass ratio quoted as 10²⁸⁹⁷, 10²⁹³⁷, and 10²⁸⁹⁵ in three different files — arithmetic says 10²⁹⁰⁵, and all three are wrong and mutually inconsistent. An Andromeda trip time reported at about half the round trip. An interstellar-medium flux given as 22.7 W/m² in one report and 2.25×10⁴ W/m² at the same velocity in another.

This is the opposite of the usual LLM failure mode, where the prose is fluent and the arithmetic underneath is broken. Here the computation is right and the prose drifts. None of the five defects overturns a conclusion; all five would be caught by a referee. They remain in the corpus, annotated — because the distribution of errors is more informative than clean numbers would be. If you're building on agent-generated research, the lesson is concrete: trust the code, audit the prose.


Next: Part 8 — Physics, Gods, and the Firewall Between Them: the agent that proved a question unanswerable, formally, and found the same firewall already built inside a 3,000-year-old tradition.