Training / Methodology

Reinforcement Learning for Calibrated Decisions (RLCD).

RLCD stands for Reinforcement Learning for Calibrated Decisions. It is the proprietary training methodology introduced by TypeSafe AI to train its System One decision models, including Jev.

Unlike traditional post-training methods that optimize models to produce human-preferred text, RLCD optimizes models to output discrete decisions accompanied by mathematically calibrated confidence scores and probabilities.

What is RLCD?

In standard machine learning pipelines, post-training methods like Reinforcement Learning from Human Feedback (RLHF) align autoregressive models to generate conversational responses that human evaluators rate highly. However, this optimization objective often rewards persuasive-sounding answers and overconfidence, even when the underlying reasoning is incorrect.

TypeSafe developed RLCD to address this limitation for automated software decisions. Rather than generating text, RLCD trains the model to satisfy two simultaneous objectives:

01Accurate decision classification: Selecting the correct categorical choice, rubric tier, or binary proposition truth status given input state.

02Calibrated probability estimation: Ensuring that the model's predicted probabilities reflect true empirical error rates across cohorts of predictions (e.g., when the model assigns 80% confidence, it should be correct approximately 80% of the time).

RLCD vs. RLHF vs. RLVR

Training Framework	Optimization Target	Primary Artifact	Known Vulnerabilities
RLHF (Human Feedback)	Human preference ratings & style	Conversational text	Sycophancy, verbosity bias, confident hallucination, mode dropping.
RLVR (Verifiable Rewards)	Binary programmatic ground truth (e.g. unit tests, math proofs)	Code / Reasoning chains	Reward hacking, limited to domains with automated verifiers, uncalibrated probabilities.
RLCD (Calibrated Decisions)	Decision accuracy + probability calibration	Typed decisions + confidence	Requires labeled decision distributions; does not generate explanatory text.

Understanding Epistemic Calibration

In statistical learning, calibration measures the degree of correspondence between predicted probabilities and observed empirical frequencies:

A model is well-calibrated if, across all instances where it predicts an event with probability p, the event occurs with empirical frequency p. Standard academic evaluation uses metrics such as Expected Calibration Error (ECE):

Calibration Definition (ECE)

// Expected Calibration Error (ECE) conceptual definition:
// ECE = Sum over bins b=1..B of (|B_b| / N) * |acc(B_b) - conf(B_b)|
//
// In an ideally calibrated system:
// When confidence is 90% across 1,000 predictions, exactly 900 (90%) are correct.

Calibration is distinct from raw accuracy:

•Accuracy: What percentage of total predictions were correct.

•Calibration: Does the model know when it is likely to be wrong? A well-calibrated model enables reliable thresholding for downstream automation (e.g., routing high-confidence decisions to automated code while escalating low-confidence decisions to human review).

What TypeSafe Has Publicly Disclosed

Through its launch report and technical primer, TypeSafe AI has disclosed the following facts regarding RLCD:

•Decision Primitives: RLCD is applied directly to optimize Jev's three evaluation primitives: Choice, Score, and Noul.

•Non-generative loss formulation: The training loss penalizes both classification mispredictions and probability miscalibration without involving token perplexity on generated text.

•Workflow evaluations: TypeSafe published internal comparative evals demonstrating that decomposing multi-step agent actions into calibrated System One calls yielded significant speed and cost advantages over LLM-based structured prompting baselines.

What Remains Undisclosed

TypeSafe has maintained proprietary boundaries around the specific engineering mechanics of RLCD:

•No public loss formula: TypeSafe has not published the exact mathematical equation, reward weighting parameters, or regularization terms used in RLCD training runs.

•No training dataset disclosures: The specific pre-training corpora, fine-tuning task distributions, and ground-truth validation sets remain unreleased.

•No base model architecture details: Parameter counts, foundation model lineage, and classification head architectures have not been publicly shared.

Current Evidence & Third-Party Evaluations

When evaluating claims about RLCD and Jev, developers should distinguish between vendor claims and independent benchmarks:

•Vendor claims: All performance figures reported in TypeSafe's launch post (such as internal cluster response times of 70ms–500ms and workflow cost savings) represent TypeSafe internal evaluations and remain attributed as vendor claims.

•Independent benchmarks: Third-party community evaluation suites (such as early academic evaluations and independent harness tests) provide empirical data on specific classification benchmarks, but multi-region independent replication across production workloads is ongoing.

•Hallucination scope: RLCD does not mathematically "eliminate errors." It produces calibrated confidence intervals so that software systems can systematically detect and handle uncertainty before taking automated actions.

← What is a System One Model?Apply Calibrated Outputs to Model Routing →Review TypeSafe Creator & Claims Ledger →