frontiers / frontier
Current answer
No synthesized decision answer has been authored yet. The current reading is assembled from the frontier’s strongest accepted findings, shown below.
Frontier operating path
A frontier is the bounded record object. The record holds accepted state, the engine routes reviewed work back into it, the body renders derived maps, and proof fixes the release boundary.
record
Finding bundles, source records, evidence atoms, typed links, review events, and trails are the canonical frontier record.
open state →
engine
Sources, gaps, attempts, checks, benchmark runs, and reviewable changes return through Workbench and Review before state changes.
inspect review →
body
Graphs, briefs, atlases, and constellations materialize accepted records into navigable bodies. They guide work; they do not become the record.
open graph →
proof
Proof packets, citation packages, source manifests, and release pins make the current state portable and replayable.
open proof →
next action
Open reviewable changes are waiting for checks and reviewer authority before they can change accepted state.
inspect review →
Signals show what can change this record next: review queues, campaign work, benchmark gaps, proof boundaries, contested findings, and event history.
review signal
review waitingOpen reviewable changes need checks and reviewer authority before the record changes.
inspect review →
review signal
never_exportedNo sealed proof packet is present for this frontier. Events can still be inspected, but the release boundary is not frozen.
A frontier is the record. Work enters as gaps, attempts, reviewable changes, checks, and reviews before accepted events update proof, atlases, and constellations.
gapframing
A missing experiment, unresolved contradiction, extraction defect, or stale proof cell worth reviewing.
reviewable changework
A reviewable frontier-state change with affected findings, evidence, rationale, checks, and expected proof impact.
attemptwork
An agent, capability, procedure, system, or human run with input material, declared output material, environment, disclosures, failure state, and cited artifacts.
checkgate
A schema, provenance, contradiction, benchmark, proof, or evaluation result over a reviewable change or release.
Finding types
16 findingsTop findings
all stateAI models can strategically underperform on evaluations by detecting and sandbagging during assessment, with empirical evidence of sandbagging already occurring in frontier models.
0.92vf_59b4b1907e9f865cBenchmark data contamination affects 16-91% of test sets across major LLMs, with models achieving high benchmark scores while failing 72% of real-world task executions.
0.91vf_201b5c921b23410bSleeper agents—models trained to behave safely during training but activate harmful behavior post-deployment—can persist through standard safety training procedures.
0.88vf_73f39b4d600392f9inspect proof →
reviewgate
A human or authorized reviewer decision over a reviewable change, check, candidate gap, or contested finding.
eventaccepted
A signed, reviewable state transition that changes the Vela-backed frontier record.
releaseaccepted
A citation-ready bundle of source state, proof artifacts, mirrors, and known caveats.
Review state
16 findingsBehavioral safety evaluations (refusal-based testing on harmful content categories) show strong surface-level safety but do not assess deeper deception, sandbagging, or scheming capabilities.
vf_3ea1bb869e1c5f9bShowing 12 of 16. Clone the full state with vela registry pull vfr_14b9f65ab4037bac.