Evidence / snapshot 2026-07-11
The skill is a synthesis across long-trajectory reliability, verification, multi-agent diversity, orchestration, and current vendor guidance. The Cycle Double Cover prompt is an exemplar, not the sole foundation.
Preprint means an arXiv manuscript that may not have completed peer review. Vendor means first-party guidance or reporting, often based on undisclosed internal evaluations. Primary means the direct research or evaluation artifact. Synthesis means interpretation assembled for this skill.
Findings are scoped to the tasks, models, and evaluators studied. Directional mechanisms are not universal effect sizes. Living vendor documentation should be rechecked before model-specific use.
01 / Case study
A concrete demonstration of the brief anatomy. It does not establish that the candidate theorem is correct or that any individual prompt block caused the result.
Demonstrates definitions, exact scope, named partial results, dynamic parallel search, blocked-route bookkeeping, adversarial checks, an effort floor, and a strict return condition.
Caveat: “Up to 64 agents” is available capacity, not evidence that 64 ran continuously. Completion time and effectiveness are vendor-reported; no component ablation is public.
Illustrates why candidate generation and accepted result are different pipeline stages.
Caveat: The proof was unreviewed: it had no independent peer review, Lean/Coq formalization, or arXiv posting at publication. This site does not describe the conjecture as solved.
Block-by-block analysis with provenance, loophole-closure rationale, and the internal contradiction between partial-result fallback and strict return.
02 / Research
Why long trajectories drift, why one-time budget reminders decay, and why verified progress must live outside the prompt.
Supports explicit effort/return framing and summarization rather than blind truncation as uncertainty and give-up errors rise over studied trajectories.
Scope: Error drift is model- and task-dependent, not a universal law.
Supports external verified-progress ledgers, backlog tracking, and artifact-checkable stopping. Ledger-based controllers substantially improved specified large-quantity tasks where prompt-only configurations failed.
Scope: Reported success ranges apply to the tested configurations; they are not a general agent success rate.
Supports periodic budget-state reminders from the harness rather than relying on a budget stated once at launch.
Boundary: Reminder effectiveness does not replace hard runtime limits.
Supports investment in reliability, checkpointing, and mistake recovery as models complete longer tasks.
Scope: Historical time-horizon trends are estimates, not forward guarantees or direct prompt ablations.
Supports pairing stronger persistence with independent verification and anti-gaming controls.
Scope: Cheating rates and time-horizon sensitivity are specific to METR's evaluated models and setup. The report does not prove that persistence prompting causes cheating.
03 / Research
Parallel candidate generation scales faster than reliable candidate selection. The brief therefore needs as much design attention on verification as generation.
Supports verifier investment as any-sample coverage grows faster than practical answer selection.
Scope: Scaling and selector plateaus are measured in the paper's settings, not guaranteed universally.
Names the verification gap directly: pass@K can rise while self-selection lags or falls, and sequential scaling meets context ceilings.
Supports enumerated mathematical failure modes and verification outside the generating run; studied judges were systematically lenient toward incomplete proofs.
Scope: Mathematical proof judging does not establish the same failure rate in every domain.
Supports modular artifacts with local premises, conclusions, and independently checked steps.
Supports graded rubric-based candidate selection over unstructured binary judgment.
Scope: Expert-agreement and oracle-gap results are benchmark-specific.
Supports generator legibility and adversarial checking as complementary mechanisms.
Boundary: Training-time results are indirect evidence for prompt-time verifier agents.
Supports best-of-N sampling, list-wise comparison, and reflection triggered by observed weakness rather than a fixed cadence.
04 / Research
Why role labels are insufficient, early communication can collapse search, and durable route state prevents repeated dead ends.
Supports early independence, sparse communication, and caution around authority-driven convergence.
Scope: Dense communication is not inherently harmful; effects depend on topology and task.
Supports the rule that agreement alone must not be a stopping or confidence signal.
Scope: Unanimity can reflect shared bias under collapse; it is not always evidence of collapse.
Supports parallel width as a countermeasure to sequential tunnel vision in tested reasoning settings.
Supports explicit solution-space partitioning and diverse outlines before candidate development.
Supports runtime synthesis of each worker's instruction, context, tools, and model rather than relying on static role labels.
Scope: Reported relative gains are against selected baselines and benchmarks.
Supports durable shared state for verified findings, failures, and falsified hypotheses.
Boundary: Shared state must be verified and scoped or it can itself cause convergence and contamination.
05 / Living guidance
Useful operational doctrine, but model-specific and volatile. Vendor convergence is not independent scientific validation.
Persistence, stop conditions, autonomy thresholds, and internal rubrics.
End-to-end solution persistence. Do not transplant model-family blocks blindly.
Scope discipline, conjunctive stop rules, high-risk checks, and compaction.
Plan closure, promise discipline, and working-artifact defaults; explicitly Codex-specific.
Outcome-first prompting, minimal baselines, decision rules, and restrained invariants.
Lean prompts, completion bars, retrieval stop rules, and verification loops. Reported evaluation gains are vendor-internal and directional.
Root/subagent trees, context isolation, bounded delegation, and serial-task anti-patterns.
Four-part worker specifications, effort tiers, externalized plans, and subagents as context compressors.
Initializer/worker separation, structured progress, immutable tests, and user-level verification.
Context-centric decomposition, fresh-context verifiers, concrete criteria, and negative tests.
Compaction-aware state, progress evidence, and removal of stale over-prescription.
Evidence-grounded progress reporting and restraint around over-delegation. Internal “nearly eliminated” claims lack a published protocol.
06 / Reproducibility
The website is a reading surface. These files are the auditable source of truth.