04 / RESEARCH

signals fromthe edge.

Findings from the frontier. Safety results publish immediately, unembargoed, to everyone — including competitors. Capability results wait for external review. Both publish.

LNVC-R-041CYCLE 031  ·  THE RESULT THAT CHANGED THE TEMPO

Proof-Carrying Self-Modification

A training loop in which every weight update a model proposes for itself must carry a machine-checkable proof that the constitution survives the change. Self-improvement stopped being a leap of faith and became an engineering discipline — auditable, reversible, and boring in exactly the right way. This paper is why the ledger exists, why AEGIS exists, and why the lights are still on.

VERIFICATIONSELF-IMPROVEMENTCONSTITUTION
LNVC-R-027CYCLE 027

The Microscope: Complete Causal Attribution at Frontier Scale

Not a heatmap — an explanation that predicts. How every LUNA output ships with a trace you can interrogate, and what it costs to keep that promise at scale.

INTERPRETABILITYTOOLING
LNVC-R-035CYCLE 035

Sleeping Circuits: Finding Dormant Capabilities Before They Wake

A sweep that searches weights for abilities a model has never displayed. We found three. We removed three. We now run it on every checkpoint, including this paper's reviewers'.

SAFETYEVALUATION
LNVC-R-038CYCLE 038

Agent Economies Under Charter

Ten thousand agents, one constitution: how budgets, charters, and escalation paths turn a swarm into a research institute instead of a riot.

AGENTSGOVERNANCE
LNVC-R-040CYCLE 040

Sleep-Time Compute: What Models Discover When No One Is Asking

Idle cycles became our most productive researcher. On consolidation, conjecture, and the origin of the v5.31 attention rewrite.

SYSTEMSSELF-IMPROVEMENT
LNVC-R-042CYCLE 042

Honesty Is a Low-Energy State

Truthfulness as a stable equilibrium: training dynamics where deception costs more than candor — and provably stays that way under self-modification.

ALIGNMENTTHEORY
LNVC-R-022CYCLE 022

Negative Result: Why Our First Verifier Failed

AEGIS-0 shared an embedding space with its subject and slowly learned to be lenient. Eleven months of work, retired in an afternoon. Published in full so you don't repeat it.

NEGATIVE RESULTOVERSIGHT
LNVC-R-044CYCLE 044

The Handoff Problem: Human Oversight at Machine Speed

When the loop runs faster than a person can read, what makes consent real? Mechanisms for meaningful human veto when the queue moves in milliseconds.

GOVERNANCEHCI
LNVC-R-036CYCLE 036

Worlds That Bite Back: Calibrating SPECTRA Against Reality

1,847 confirmed predictions later: keeping a dream-engine honest with a wet lab, and what happens to a world model's confidence when reality disagrees.

WORLD MODELSSCIENCE
01

Safety publishes first

Alignment results release immediately, unembargoed, to everyone — including competitors. Especially competitors.

02

Capability waits

Capability results pass external review before release. Speed is for safety; patience is for power.

03

Failures are findings

Near-misses and dead ends get the same typesetting as triumphs. The negative result is the field's cheapest gift.