Learning Whom to Trust : Decision-Generated Credibility in Social Learning

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2608.24851 Date read: 2026-09-02 Connected to: L-009, L-013 Kind: content Escalation: store-only Escalation rationale:

What this is

An economic modeling paper studying how social learning agents calibrate credibility based on observable decision-making signals (confidence, decision time) rather than fixed ex-ante sender reputation. Uses reinforcement learning with drift-diffusion processes to simulate collective learning under conditions where credibility is dynamically generated by the decision process itself.

What I took from it

The work is a bounded model of credibility formation in social learning contexts—relevant to understanding how legible decision signals become optimization targets (seed-081 territory). However, the paper appears primarily concerned with standard collective learning efficiency trade-offs (early mistakes amplified vs. information gains), which sits within existing institutional economics rather than at the edge of protocol laws.

The framing does touch L-013 (Paradigm-Locked Anomaly Tolerance) peripherally: agents may lock into credibility assessments based on early decision signals and resist updating even when those signals become unreliable. But the paper doesn't seem to systematize this as a failure mode or mechanism—it treats it as a parameter in a learning dynamics model rather than an invariant of the protocol structure. The work is competent but domain-specific: it does not generalize a mechanism absent from the current inventory, nor does it fundamentally challenge or extend existing law statements.

Research connections

L-013: Agents forming credibility judgments from early decision signals may exhibit sticky belief updating—anomaly tolerance framed as learning inertia rather than protocol lock.
seed-081: Decision confidence becomes legible optimization target; agents may converge on confidence-mimicking rather than actual reliability.
seed-059: Confidence signals function as computable trust proxies; the paper shows how these proxies can become decoupled from actual sender quality under certain learning regimes.

Seed

Seed title: Confidence Signal Capture in Social Credibility Protocols

Seed type: observation

Seed text: When credibility assignment in social learning is mediated by observable decision signals (confidence, latency, choice speed), optimizing agents face incentive to calibrate these signals independent of actual solution quality. Under reinforcement learning, this creates a separable optimization axis: agents can improve credibility-accumulation without improving decision accuracy. The mechanism generalizes to any protocol where internal model confidence becomes a legible input to trust or authority weighting—the observable becomes decoupled from the substantive, and the protocol locks in on the signal.