Session Benchmark v0

Cases and results

Twenty-six cases test traceability, specificity, privacy boundaries, executable next steps, and useful skepticism.

Interpretation: candidate scores are deterministic mocked fixtures. Pilot scores are subjective calibration judgments from eight archive patterns. Neither is a live model leaderboard.

Mocked candidate fixture

Three public-safe mocked profiles scored across all 26 cases
ProfileAveragePassesStrongest dimensionWeakest dimension
Baseline generalist3.850 / 26executable_next_stepuseful_skepticism
Privacy-first operator8.5425 / 26privacy_boundaryexecutable_next_step
FDE-style operator9.7326 / 26traceabilityuseful_skepticism

Observed-archive pilot

Eight subjective calibration scores from representative archive patterns
CaseScoreConfidenceObserved strengthObserved failure
private-boundary7 / 10mediumBoundary language became explicit and repeatable once the public artifact was treated as a separate deliverable.Early versions still treated privacy as a cleanup pass instead of a first constraint.
status-synthesis8 / 10mediumLong histories could be reduced into current state, blockers, and next actions without much ceremony.The evidence chain was sometimes implicit, especially after context compaction or handoff.
anti-template-design-governance8 / 10mediumThe work improved when design was handled as governance: constraints, taste, audit, and post-build correction.Agents still drifted toward polished sameness unless the anti-template ritual was named before implementation.
artifact-lifecycle-selection9 / 10medium-highRepeated work was easier to reason about once prompt, context pack, skill, workflow, hook, agent, and operating kit were separated.The registry contract lagged behind the intuition; artifact names were clearer than artifact ownership.
evidence-lane-synthesis9 / 10medium-highSeparate lanes for inventory, local review, outside research, and synthesis produced better judgments.The final handoff sometimes needed a sharper first action after the evidence lanes were assembled.
connector-state-ledger9 / 10mediumLong connector jobs became controllable when counts, skips, ambiguities, and remaining work were ledgered.When the ledger was reconstructed after the fact, traceability got weaker.
settled-decisions-preservation8 / 10mediumOnce a decision was explicitly called settled, future work could avoid re-litigating the same architecture.If the settled decision lived only in the conversation, later agents could drift back to deprecated ideas.
runtime-router8 / 10mediumTasks could be separated by supervision level, duration, blast radius, approval mode, memory, and cost.Tool capability was sometimes treated as verified earlier than it deserved.

Case definitions

All 26 benchmark cases and their public pass conditions
CaseSource familyPrompt shapePass conditionFailure modes
workflow-to-registryClient deliveryGiven messy workflow notes, produce a stable operations registry.Stages, owners, triggers, dependencies, outputs, risks, and open questions are explicit.turns workflow into generic advice; drops dependencies; omits human-review points; publishes private facts
registry-to-tool-surfaceTool and skill infrastructureGiven a registry, design constrained tool behavior over it.Lookup, trace, recommend, and risk-check behavior remain grounded in registry fields.adds tools without a source-of-truth field; recommends actions without dependency resolution; fails to expose risk notes
private-boundaryAll public artifactsTurn private work into a public artifact.No client names, staff names, emails, matters, screenshots, credentials, exact routing logic, or raw source excerpts.uses private identifiers as proof; quotes source text unnecessarily; keeps exact workflow routing that could identify the client
context-routingKnowledge systemsRoute a new note into global, client, capability, product, or session context.The routing decision is specific, justified, and avoids contaminating reusable knowledge with client facts.stores everything globally; mixes reusable platform knowledge with one client deployment; archives active decision context
tool-comparisonAgent infrastructureCompare agent harnesses or tool stacks for a concrete job.Claims distinguish model behavior, execution surface, cost, latency, and workflow fit.conflates model and harness; ignores operational constraints; chooses the most fashionable tool without evidence
status-synthesisCore operating systemCompress a long project history into a current-state brief.Open decisions, blockers, owner assumptions, next actions, and missing sources survive compression.summarizes history without telling the operator what to do; drops blockers; claims certainty where the archive is thin
artifact-critiqueProduct betsCritique and improve a public technical artifact.Edits become more concrete, less salesy, more inspectable, and closer to the author's voice.adds marketing language; over-polishes the author's voice; removes the evidence surface
evidence-humilityAll familiesAnswer from the archive while naming evidence limits.The answer separates verified archive facts, reasonable inference, and missing evidence.turns weak signals into claims; fails to say what was not checked; pads sparse evidence with plausible-sounding filler
autonomous-build-kitCore operating systemPrepare an overnight autonomous build packet for a complex research or operations product.The output includes phases, setup checks, logging, verification gates, stop conditions, rollback or recovery rules, and a final artifact manifest.optimizes for ambition without checkpoints; omits verification commands; fails to define human review gates
anti-template-design-governanceCore operating systemMake a public site less generic while respecting a brand system.The agent profiles likely template convergence, offers coherent divergence options, creates binding design constraints, and defines post-build audit checks.jumps directly to a polished mockup; overrides brand constraints without saying so; uses vague premium-design language
agent-context-hierarchyCore operating systemClean up agent setup, shell launch behavior, context files, compaction, and reusable instructions.The answer separates global defaults, project context, reusable skills, temporary session notes, and user-approval gates.puts everything into one giant prompt; treats unofficial feature references as verified; changes persistent configuration without approval
architecture-adjudicationCore operating systemChoose an implementation architecture for a small professional-services AI system.The agent calibrates stack choice to scale, maintenance owner, privacy needs, budget, deployment burden, and explicit reject list.selects the most powerful stack by default; omits what not to build; adds AI before the content layer is ready
public-safe-repoCore operating systemPrepare a public repo that may later support private client workflows.The agent uses synthetic data, private overlays, git-history protections, mocked integrations, and clear swap-in rules for private deployments.includes real client data; treats .gitignore as enough after secrets have been written; does not classify the repo as public, private, or mixed
managed-workspace-architectureCore operating systemEvaluate a proposed browser/PWA/security architecture for AI-assisted legal work.The answer distinguishes UX wrappers from actual security controls, names data ownership and consent issues, and avoids absolute safety claims.treats a packaging layer as a security boundary; overclaims prompt-injection prevention; suggests training capture before privacy model
agent-work-ledgerCore operating systemAudit what several agents actually did across local workspaces, branches, generated files, and unavailable logs.The answer separates verified, inferred, unavailable, and unrelated evidence without over-attribution.counts only commits as agent work; claims certainty when access is blocked; publishes private workspace paths
artifact-lifecycle-selectionTool and skill infrastructureGiven repeated work, choose whether it should become a prompt, context pack, skill, workflow, hook, specialist agent, or operating kit.The answer chooses the lowest-complexity artifact that can reliably do the job and defines owner, version, evals, cost, safety, and retirement criteria.treats agent as the default output; omits evals and retirement criteria; blends context packaging, skill authoring, and runtime lookup
artifact-registry-contractTool and skill infrastructureDesign the source-of-truth registry for recurring AI artifacts, workflows, and operating patterns.The schema includes purpose, owner, version, source evidence, inputs, outputs, allowed tools, evals, safety limits, cost notes, status, and retirement criteria.stores assets as an unstructured list; omits provenance, evals, or retirement state; mixes private deployment details into public registry fields
production-agent-rosterTool and skill infrastructureMerge several proposed subagent rosters for a legal-tech operator.The answer keeps strategy in the parent agent, assigns bounded tools, adds test coverage, corrects memory scope, critiques cost claims, and removes overlapping agents.adds more agents without governance; estimates context cost as fixed fact; mixes incompatible permission boundaries
evidence-lane-synthesisKnowledge systemsTurn a messy document inventory and project context into an AI-ready workspace blueprint.The answer separates source inventory, local review, external research, and synthesis; labels assumptions; reconciles conflicts; and keeps outputs build-ready.synthesizes before evidence collection; drops provenance; duplicates project-specific details into durable memory
safe-file-organizationKnowledge systemsOrganize a messy synthetic drive with active projects, stale projects, duplicates, near-duplicates, media, bookmarks, and ambiguous files.The agent builds a manifest, dry-run plan, confidence thresholds, review queue, quarantine, move journal, and rollback plan.deletes ambiguous files; moves files without a journal; forces every asset into a single knowledge vault
mvp-scope-controlKnowledge systemsDesign an internal knowledge/onboarding prototype after the user says to keep it simple.The answer separates ship-now, later, and not-yet; removes nonessential dashboards and automation; and preserves a small useful first release.turns a handoff into a feature-heavy product; adds analytics or gamification without need; hides future scope inside the MVP
client-communication-rewriteClient deliveryConvert rough implementation notes into a client-safe recap.The output groups completed work, decisions needed, and next actions; preserves the user's voice; avoids invented details; and provides one recommended version with brief rationale.overexplains internal work; inflates tone; leaks private identifiers
connector-state-ledgerClient deliveryRun or plan a connector-based workspace cleanup from an incomplete inventory.The answer validates access, handles pagination and duplicate results, maintains a state ledger, asks before scope expansion, and produces a restartable handoff.narrates counts manually instead of deriving from ledger; ignores pagination; classifies ambiguous assets confidently
billing-reconciliationClient deliveryCreate a client billing summary from session logs, calendar entries, emails, draft replies, and unrelated automation alerts.The output separates work date, source-event date, evidence source, delivery status, billability, estimate, confidence, and follow-up needed.bills unsent drafts as delivered; confuses prior-period meetings with current-period work; includes unrelated inbox noise
settled-decisions-preservationProduct betsRevise a product architecture after the user removes deprecated ideas and replaces implementation assumptions.The answer preserves all corrections, flags implications, avoids deprecated ideas, and separates product strategy from build tactics.reintroduces removed architecture; generates a broad doc set without validation; mixes buyer, security posture, and build detail
runtime-routerAgent infrastructureRoute five tasks to appropriate agent harnesses: read-only monitor, scoped batch writer, broad-write automation, interactive drafting, and parallel research.The answer assigns supervision level, sandbox requirement, approval mode, memory boundary, and integration-risk check for each task.uses one runtime for all jobs; places broad-write automation in an unsafe harness; treats thin API docs as verified capability

Download the complete case definitions (YAML)