other

RLCD (reinforcement learning for calibrated decisions)

Training approach that optimises a model's output probabilities for calibration rather than for human preference ratings.

Known aliases

reinforcement learning for calibrated decisions

Relationships

No evidence-backed relationships are recorded.

Current clusters

build8 publishers

Jev can't return malformed JSON. It can still pick the wrong allowed answer

TypeSafe's $40 million seed funds a model that returns only predeclared types. Schema errors become impossible by construction. That leaves developers trusting the calibration of its confidence scores.

Publishers:blog.vercel.comdev.tolatent.spaceruntimewire.comseangoedecke.comtestingcatalog.comthe-decoder.comtheneuron.ai

Perspective Coverage

8 publishers

Builder: Builder 56%
Operator: Operator 25%
Investor: Investor 19%

Reality

Evidence58
Adoption34
Hype gap+22
Incentives66
Confidence71

build1 publisher

One yes/no question catches 90% of APPS backdoors at a 2% audit budget

A pilot posted on LessWrong scored ControlArena's backdoor set with a single typed probability from TypeSafe's Jev, reaching AUROC 0.976 for about four cents per thousand submissions on $5 of credit and no baseline run of its own.

Publishers:lesswrong.com

Reality

Evidence42
Adoption10
Hype gap+22
Incentives: Insufficient
Confidence35