AETRE & The Governed Agent

Comprehensive pre-submission diagnostic scorecard for authors and researchers, calculating crowd novelty percentile, reviewer disagreement risk, and prescriptive refinement actions.

Batch applies disclosed, uncalibrated lexical indicators to a cohort, computing heuristic ranks, VOI ranks, and demonstration stream allocation (Stream A Fast-Reject, Stream B Deep Review, Stream C Fast-Pass).

Calculates the exact Bayesian Value of Information (VOI) for crossing a top-K selection boundary under Gaussian conjugate updates.

Fits Platt logistic scaling on continuous model scores and binary labels, returning slope, intercept, Expected Calibration Error (ECE), and Brier score.

Evaluates evaluator queue load using Kingman's Heavy-Traffic approximation and returns governor throttle recommendations when utilization exceeds rho >= 0.85.

Optimizes reviewer-to-proposal assignment by maximizing domain/keyword affinity while enforcing Kingman queue utilization constraints (rho <= 0.85) on individual reviewer workloads.

Calculates Bayesian posterior mean and uncertainty under correlated multi-agent evaluator noise (rho_corr), preventing artificial overconfidence from redundant LLM outputs.

Simulates submitter entry equilibrium under generative AI generation costs and refundable submission deposits to curb spam floods.

Calculates the unbiased Horvitz-Thompson exploration audit estimator (H_hat_D) and 95% confidence intervals on deprioritized candidates to catch false negative breakthroughs.

Fits the decision boundary to a venue's own calibration data by sweeping thresholds and ranking candidates by boundary VOI at each. The boundary determines whether triage beats chance, and a default carried from another corpus generally does not.

Calculates Generalized Pareto / Heavy-Tailed Value of Information (VOI) to optimize selection pipelines for positive black swan breakthrough discovery.

Runs a multi-policy held-out review allocation backtest across 8 triage policies under fixed review budget K, evaluating true decision flips, precision, recall, and paired bootstrap intervals.

Evaluates a multi-specialist heterogeneous reviewer network, identifying bottleneck domains and generating capacity rebalancing actions.

Runs the venture dealflow triage benchmark: generates a synthetic heavy-tailed cohort of deals and compares status-quo preliminary-score screening against AETRE heavy-tailed VOI triage under a fixed diligence budget.

Computes multi-attribute Bayesian Value of Information across orthogonal proposal evaluation dimensions (Novelty, Rigor, Impact, Feasibility), outputting composite VOI and optimal dimension-specific review targets.

Calculates Proposition 1 theoretical recall ceiling R_N <= min(1, K_N / H_N) to determine if a pipeline is mathematically capacity-constrained.

Calculates super-linear anti-sybil staking deposit requirements (Stake(m) = S_0 * m^gamma) to deter mass AI spam submissions while preserving human entry.

Calculates the Proposition 1 theoretical recall decay curve across arrival expansion scales (e.g. 1x, 2x, 5x, 10x, 20x, 50x) demonstrating capacity collapse points.

Calculates optimal dynamic Bayesian stopping boundaries for sequential reviews (Accept, Reject, or Solicit More Reviews) based on posterior decision confidence and boundary VOI.

Runs a paired-cohort Monte Carlo simulation across all 4 screening regimes, comparing Quality Throughput, FDR, Unconventional Recall, and Human Reviews with central 95% run-to-run outcome intervals (not confidence intervals for the mean).

Sweeps the submission fee across a range and returns the submitter equilibrium at each point, showing how entry volume and low-quality deterrence respond to the staking fee rather than evaluating a single fee.

Comprehensive system introspection returning AETRE architecture, bundled synthetic fixtures, optional data adapters, connectors, mathematical tools, and institutional tiers.

Applies transparent, uncalibrated lexical routing indicators to proposal text, then calculates a VOI index and demonstration stage route (FAST-PASS, FAST-REJECT, or DEEP REVIEW). Not a validated estimate of scientific quality.

Pillar I Bellman Governor: evaluates Bayesian dynamic programming stopping policy over multi-stage pass lattices under asymmetric loss stakes (L/R), returning optimal action (CONTINUE, HALT_AND_COMMIT, HALT_AND_REJECT), expected utility, VOI, and critical threshold p*.

Micro-level loss evaluation comparing autonomous execution against deferral to human review, under an asymmetric loss matrix weighted by irreversibility.

Road A Tiered Verification Gate: evaluates host pre-checks (Tier 0 syntactic AST parsing and structural validation) on candidate Python code or pull request patches to short-circuit broken submissions before expensive container CI escalation.

Evaluates declarative safety invariants against a runtime state snapshot before a state change is committed, returning every violated rule rather than the first.

Pillar V Knapsack Queue Controller: packs candidate pull requests into the review queue under capacity budget K using the c-mu rule (density rho_i = E[U_i] / k_i) and calculates the dual capacity shadow price lambda_K.

Fits the recall saturation curve Recall(K) = 1 - exp(-gamma K) to observed budget and recall pairs by least squares, returning the fitted gamma and the budget beyond which marginal recall stops paying for itself.

Section 4.1 Tripartite Review Boundary: evaluates whether an autonomous coding candidate should be AUTO-admitted, sent to human REVIEW, or ABSTAINED based on reviewer effort cost and knapsack capacity shadow price lambda_K.

Generates a tamper-evident SHA-256 decision receipt over an execution's identifier, payload and timestamp, so a gated agent action can be verified after the fact.

Computes the marginal shadow price of scarce reviewer time (lambda_K) from queue backlog and review capacity under a heavy-tailed utility distribution, and the factor by which it elevates the admission cutoff.

Solves the finite-horizon optimal stopping problem by backward induction: at each step the agent either takes the terminal payoff or pays the continuation cost for one more step of evidence.