Augmentor Policy Handoffs

{
  "query": "offline reinforcement learning trajectory augmentation stitching deterministic dynamics conservative return-improving bridge",
  "categories": [
    "cs.LG",
    "cs.RO"
  ],
  "date_from": "2018-01-01",
  "max_results": 10
}

{
  "title": "Validate summed-action detour certificates",
  "research_question": "When a redundant deterministic logger segment is replaced by its summed action, does logged action mass predict endpoint drift, and can the resulting drift budget support a conservative return-improvement certificate without simulating every candidate at acceptance time?",
  "method": "Build a seeded deterministic active-positioning simulator with memoryless state-dependent actuation x_next=x+D(x)a and a reliable scripted logger. Enumerate natural and controlled detour segments, replace each by one feasible summed command or the minimum bounded chunking, and exactly roll out the bridge and the logger continuation for evaluation. Log the complete augmented controller state and resume the planned suffix mode. Implement true time compression, the bridge's actual reward, and a total endpoint budget containing distortion-composition and clipping residual.",
  "controls": [
    "constant-field exactness before clipping, with clipping and infeasible sums reported separately",
    "smooth scalar gain, anisotropy, spatial rotation, and discontinuous or narrow-recovery distortions",
    "exact deterministic return oracle",
    "action-mass-only and exact-endpoint-distance gates",
    "local reachable-tube continuation lower envelope when a global value-Lipschitz assumption is invalid"
  ],
  "evaluation": "Primary geometry readout: true endpoint drift versus Q and calibrated B, by distortion. Primary decision readout: certificate precision, recall, and false-positive rate against exact time-compressed bridge-plus-suffix return. Any mismatch between the written splice semantics and executed rollout is a failed methodological control, not a scientific null.",
  "experiments": [
    "constant-distortion exactness and clipping audit",
    "distortion-family drift sweep",
    "return-certificate evaluation against exact counterfactual return",
    "discontinuous-logger or narrow-recovery assumption stress"
  ],
  "deliverables": [
    "raw per-candidate results",
    "drift-versus-proxy figures",
    "certificate table",
    "proof and assumptions note",
    "accepted and false-positive examples"
  ],
  "compute": {
    "gpu_type": "CPU",
    "num_gpus": 0
  }
}

{
  "proposal_id": "P-4a7c"
}

{
  "run_id": "R-91de",
  "section": "metrics"
}

{
  "run_id": "R-91de",
  "section": "controls",
  "detail": "constant-field failures, reward timing, clipping, time compression, and logger continuation audit"
}

{
  "run_id": "R-91de",
  "reason": "Do not extend or tune the invalid return-certificate sweep. The geometry question is answered, while the return comparison violated the intended reward, clipping, time-compression, and handoff semantics. Retain all drift artifacts for reuse."
}

{
  "title": "Summed-action mass tracks endpoint drift but does not certify return improvement",
  "statement": "In deterministic state-dependent active positioning, redundant action mass Q is a strong risk-ranking statistic for summed-action endpoint drift, but it is not a standalone return-improvement certificate. The attempted return test was methodologically invalid because it mismatched reward timing, time compression, clipping residual, and logger controller state.",
  "evidence": "16,800 candidates; Pearson r(Q,drift) 0.82-0.95 and empirical bound coverage 0.96-1.00. The evaluated certificate had recall 1.0 but precision 0.27-0.49 and FPR 0.93-1.00. Constant-field audit found clipping residual under B=0, zero-action waiting, wrong shortcut reward, and phase-reset continuation.",
  "caveats": "The logger's cancelling detours were synthetic. No valid conclusion about realistic shortcut yield or a correctly indexed return theorem follows from this run."
}

{
  "query": "summed action drift certificate active positioning reward timing logger state",
  "max_results": 5
}

{
  "title": "Re-test exact and contraction-gated shortcut inequalities",
  "research_question": "Under post-transition negative-distance reward and a deterministic distance-improving logger, do exact direct bridges satisfy the non-worsening inequality, and does the approximate-landing gate accept only when the fixed-mode handoff is locally contractive and covered by strict nominal slack?",
  "method": "Use complete augmented state z=(x,m), a deterministic time-homogeneous logger with non-increasing waypoint distance, true time compression, and post-transition reward. For exact bridges, preserve the planned physical and controller states and reject or minimally chunk infeasible sums rather than silently clipping. For approximate bridges, measure a conservative local contraction ratio over the certified endpoint-error tube and accept only when rho_hat<1 and B/(1-gamma*rho_hat) is below the strict nominal margin.",
  "controls": [
    "tail-decomposition and monotone-distance identity audit",
    "constant-field feasible exact bridges with no hidden clipping",
    "minimum bounded multi-step chunking with correctly indexed bridge return",
    "separate clipping counterexample audit",
    "analytically contractive handoff map",
    "analytically non-contractive handoff map that must reject nonzero-drift approximate bridges",
    "exact B=0 bypass in both contraction regimes"
  ],
  "evaluation": "For exact bridges, report acceptance and theorem violations. For approximate bridges, compare contractive and provably non-contractive regimes on acceptance, accepted-edit outcomes, and false accepts; verify the finite-difference contraction audit against the analytic factor.",
  "experiments": [
    "exact bridge theorem control",
    "feasible chunked bridge control",
    "clipping/infeasibility diagnostic",
    "contractive versus provably non-contractive assumption stress"
  ],
  "deliverables": [
    "raw per-candidate results",
    "proof with reward and indexing conventions",
    "contraction audit",
    "accepted and refused bridge summary"
  ],
  "compute": {
    "gpu_type": "A40",
    "num_gpus": 1
  }
}

{
  "proposal_id": "P-b261"
}

{
  "run_id": "R-7fc0",
  "section": "controls",
  "detail": "green pilot before full contraction stress sweep"
}

{
  "run_id": "R-7fc0",
  "instruction": "The green pilot validates the harness. Run the pre-specified full seeded repetition sweep without changing the contraction factors, margin rule, exact B=0 bypass, or nominal logger. Keep contractive and provably non-contractive conditions separate and report acceptance, accepted non-worsening rate, false accepts, and maximum contraction-audit residual."
}

{
  "run_id": "R-7fc0",
  "section": "metrics"
}

{
  "title": "Contraction gate rejects provably unsafe approximate shortcut handoffs",
  "statement": "Under post-transition negative-distance reward and true time compression, exact direct bridges are non-worsening without a contraction assumption. Approximate bridges can be admitted by a local fixed-mode contraction certificate and strict-margin gate; the same gate refuses nonzero-drift edits when contraction is provably absent.",
  "evidence": "72,000 candidates. Contractive nonzero-drift acceptance 66.01% with 0 false accepts; non-contractive nonzero-drift acceptance 0%. Exact bridges accepted 100% in both regimes with no violations. Maximum contraction-audit residual 2.7e-14.",
  "caveats": "Controlled deterministic construction only; realistic logger redundancy, bridge-class coverage, and gate yield remain unmeasured."
}

{
  "title": "Measure shortcut yield against a redundant-segment oracle",
  "research_question": "Is the contraction-gated shortcut rule merely safe, or does it recover a useful fraction of genuinely redundant segments produced by realistic deterministic active-positioning loggers?",
  "method": "Build a stateful time-homogeneous scripted positioner with deterministic partial observations and full simulator state for evaluation. Generate natural redundancy through filtered visual servoing, waypoint or approach-direction correction, settle-and-retry logic, and bounded action and jerk. For every logged segment, compute an evaluation-only best feasible deterministic bridge under the same constraints and planned suffix-mode handoff. Label oracle-positive redundancy by strictly improved true time-compressed return and safe completion. Independently run the online exact-or-contraction-gated minimal bridge policy without oracle access.",
  "controls": [
    "nominal logger on matched initial states",
    "evaluation-only oracle bridge policy",
    "unconstrained minimal-bridge proposer with exact deterministic return evaluation, isolating analytic-gate conservatism",
    "exact-endpoint-only gate",
    "smooth gain, anisotropic or cross-coupled, spatial rotation, and capture-boundary distortions",
    "logger-mechanism ablations removing filtering, waypointing, or settle-and-retry logic"
  ],
  "evaluation": "Report environmental redundancy prevalence among all segments, proposal-class coverage among oracle positives, gate recall overall and conditional on representability, gate precision, and a rejection waterfall among oracle positives. Also compare online success, return, steps, cumulative distance, action or jerk cost, constraint violations, and accepted edits per episode. A near-zero yield is valid and must not be tuned away.",
  "experiments": [
    "distortion-suite redundancy prevalence and gate recall",
    "logger-mechanism ablation",
    "nominal versus gated versus oracle online utility",
    "capture-boundary refusal stress"
  ],
  "deliverables": [
    "three-denominator prevalence, expressivity, and recall table",
    "coverage-versus-safety frontier",
    "oracle-positive acceptance and rejection visualizations",
    "exact audit of every accepted edit"
  ],
  "compute": {
    "gpu_type": "A40",
    "num_gpus": 1
  }
}