Skip to content

Evaluation-Centric Runtime — Backlog & Spec

Status: Draft · Created: 2026-06-21 · Companion:
Vision ·
FEP-0008 · ADR-009/010/011/012

Prioritized backlog for the Evaluation-Centric Runtime direction, derived from an arXiv corpus
survey grounded against Victor's eval/loop code. Thesis: Agent = Model + Harness — close the
evaluation loop and gate every change on it. Tags: feature / techdebt / vision. Effort:
S/M/L. This integrates the prior ../arxive roadmap (see "Integration" below); the internal
docs/roadmap.md carries the same epic for local planning.

ID Item Tag Pri Effort Depends Papers
EVR-1 Trajectory-eval harness in victor/evaluation/ (planning / tool-grounding / recovery / refusal, with CI confidence bounds) feature P0 M TRBench 2604.08178, 2605.10448
EVR-2 LLM-judge reliability gate (Krippendorff α / κ vs human-labeled set; order-swap ensemble) — ADR-011 feature P0 M EVR-1 AgentProp-Bench 2604.16706
EVR-3 Rubric-based completion evaluator + DimensionAwareFilter — ADR-009, FEP-0008 Phase A — Shipped opt-in; default NO-GO after SWE-bench-lite reliability failure; Prong-B A/B runner shipped for future candidates feature P0 L EVR-2 AdaRubric 2603.21362
EVR-4 Effect-grounded completion gate (verifiable state delta) — ADR-010 — Shipped opt-in (victor/framework/effect_gate.py) feature P0 M tools/verification/ HarnessFix 2606.06324, Harness-Bench 2605.27922
EVR-5 Regression-gated harness acceptance oracle + HTIR/ETCLOVG traces — ADR-012 — Shipped (victor/evaluation/{acceptance_oracle,htir}.py, promotion-gated) techdebt P0 M FEP-0007 batteries HarnessFix, Self-Harness 2606.09498
EVR-6 Online per-turn auditor (prefix-only continue/alarm) — FEP-0008 Phase C — Shipped opt-in; offline HTIR-oracle gate + digest-pinned producer + real-run review-pack exporter/finalizer shipped, independent review/frozen pack pending; default OFF feature P1 M edge model AgentForesight 2605.08715
EVR-7 Close credit→learner loop with segment-level process reward techdebt P1 L victor/agent/credit_assignment.py HISR 2603.18683, A²TGPO 2605.06200
EVR-8 Causal-frontier tool filtering (requires/effects/risk on BaseTool) feature P1 M tool registry CMTF 2606.06284
EVR-9 PruneTIR recovery ops (prune-resolved / resample-stuck / suspend-after-repeat) feature P1 M RecoveryService PruneTIR 2605.09931
EVR-10 Adaptive plan-depth in PLAN node feature P1 M TaskAnalyzer AdaPlan-H 2604.23194
EVR-11 Isotonic-calibrated routing/confidence (ECE-reported) feature P1 M smart routing UCCI 2605.18796
EVR-12 Cost-aware GEPA: add (quality, token-cost) to the Pareto objective feature P2 M victor/framework/rl/ MO-CAPO 2605.18869
EVR-13 Dual-rubric context pruning (semantic + dependency, graph-derived labels) feature P2 M graph index LaMR 2605.15315
EVR-14 Belief-entropy compaction guard (anchor-question probe) feature P2 S edge model MMPO 2605.30159
EVR-15 Runtime-supervisor sandbox tier via ASK policy (static/runtime split) feature P2 M sandbox + policy engine Sandlock 2605.26298
EVR-16 Speculative tool pre-execution on slack resources (sandbox-gated) vision P3 L sandbox/policy PASTE 2603.18897
EVR-17 Self-improving harness loop (mine → minimal edit → regression-gate) vision P3 L EVR-5 Self-Harness 2606.09498

Sequencing (measurement-first)

EVR-1 → EVR-2 → (EVR-4 ∥ EVR-3) → EVR-5 → EVR-6 → EVR-7 → …

EVR-3 must match-or-beat EnhancedCompletionEvaluator on the shipping-distribution reliability,
parity, characterization, and verifier-backed task-success batteries before becoming the default
completion_strategy. The 2026-08-05 SWE-bench-lite reliability re-gate failed, so the current
decision is a documented no-go: keep enhanced default and revisit only with a new candidate.

Integration with the prior roadmap

From ../arxive/agentic_runtime_roadmap_2026-04-27.md (standing P0s), reconciled — not replaced:

  • Cost-aware topology routing and the generative-optimization benchmark harness become
    consumers of the EVR-5 acceptance oracle (they are graded by it).
  • Experiment / proactive memory is fed by EVR-7 (segment-level process reward → learners).
  • Calibrated uncertainty (truth-aligned 2604.00445) is the substrate for EVR-2 / EVR-11.

Already delivered — do not re-propose

Perception + calibrated-confidence fusion, the PPAED loop (FEP-0007 unified), GEPA / MIPROv2 / CoT
distillation, FulfillmentDetector, semantic response cache, paradigm routing, tool-loop/spin
detection, offline AgenticExecutionTrace.

Meta-deliberation narration guard. _is_intent_only_response (in both
victor/framework/agentic_loop.py and victor/framework/enhanced_completion_evaluation.py)
now performs a full-response density check in addition to the legacy first-line prefix
check. When a response carries no payload (no fenced code block, no markdown table) and
contains 3+ distinct imminent-action markers ("Executing now", "Going now", "Calling now",
"Making the call", "no more deliberation"), it is classified as intent-only narration rather
than a substantial answer. This prevents the failure mode where the model narrates imminent
action without ever invoking a tool, which previously exited the loop before any tool ran.
Real answers carrying a code block or result table are never flagged.