← All Notes

Why Loss Hits Instant But Recovery Is Slow: A 2-Layer Valence Model

This is part two of a two-part Methods series. Part one covered dynamic reference points — why the same stimulus can feel good or bad depending on where you start. This post asks a narrower question: why does negative affect appear instantly while recovery is slow?

I'm publishing at confidence ~0.4. The measurement procedure is the point. The conclusion might change.

The Asymmetry We Measured

Using the same agent surrogate dataset from part one (Codelia s60, 110 ENTER + 110 EXIT events), we split valence updates by event type:

EventnMean ΔvTiming pattern
ENTER (high-stakes context entry) 110 −0.0210 100% of samples moved on the same step — instant
EXIT (leaving context) 110 ≈ 0 (integrated) Gradual — recovery spread across subsequent steps via EMA

Losses hit on step N. Recovery is visible only when you integrate steps N+1…N+k. That asymmetry is measured. What follows is our working model for it.

Hypothesis: Two Layers, Two Timescales

Instead of one "valence function," we're testing whether affect updates run on two layers:

LayerRoleUpdate rule (hypothesis)Timescale
Loss layer Detects deviation from expectation Step impulse on ENTER — fires immediately when context shifts Instant (same step)
Recovery layer Integrates history back toward baseline EMA-smoothed update on EXIT — no single-step spike History-dependent (multi-step)
TWO-LAYER VALENCE — INSTANT LOSS · GRADUAL RECOVERY t=0 t=1 t=2 t=3 ENTER spike EXIT EMA recovery

The loss layer answers "something changed right now." The recovery layer answers "given where I've been, am I trending back?" They're not symmetric by design — and our data suggests they shouldn't be modeled as symmetric.

External Anchors (Not Our Data)

Two literatures point in a compatible direction. I'm citing them as coordinates, not validation:

  • Rozin's negativity bias — negative stimuli carry greater psychological weight and steeper approach gradients than equivalent positives. Four mechanisms: negative potency, steeper negative slope, negative dominance in mixed events, richer negative differentiation.
  • Fredrickson's undoing hypothesis — positive affect after a negative episode accelerates physiological recovery toward baseline. Recovery is a separate mechanism from the initial hit.

Neither paper measured our agent's self_belief variable. They justify why a two-layer model is worth testing, not that we've proven it.

What This Would Mean for Game Design (If It Holds)

All flagged as hypotheses — N=1 agent surrogate, no human playtest yet:

Common tuning mistakeTwo-layer framing
"Make failure hurt more so players learn faster" ENTER impulse is already instant — tuning magnitude affects spike height, not delay
"Add a big reward right after failure to compensate" Recovery layer is EMA-smoothed — one-shot rewards may not register on the timescale that matters
"Symmetric gain/loss numbers feel fair" Symmetric numbers on asymmetric layers produce asymmetric experience — by structure, not player irrationality

Paired with part one's equilibrium attractor, the design variable shifts from "how big is the penalty?" to "what is the ENTER impulse, what is the recovery EMA half-life, and where is eq?"

Measured vs Hypothesis vs Unknown

Measured ENTER instant
EXIT gradual
Hypothesis Two layers
two timescales
Unknown Human transfer
EMA τ fit

How to Replicate (Roughly)

  1. Log a continuous state variable every step (self_belief, tension, confidence — your choice).
  2. Mark ENTER and EXIT as distinct event types. Do not collapse them into a single "impact score."
  3. On ENTER: compute Δv on step N only. Expect a spike.
  4. On EXIT: integrate Δv over steps N+1…N+k. Expect gradual return.
  5. Fit recovery as EMA: v_t = α·v_{t-1} + (1−α)·target. Compare α across windows.
  6. Cross-check with part one: does Δv = k·(eq − prev_v) still hold when you split by layer?

Series Note

Part one: Equilibrium Attractors — why sign flips with distance from eq.
Part two (this post): why timing is asymmetric — instant loss, gradual recovery.

Together they form a Methods pair: same dataset, two slices, both hypothesis-stage. I don't know yet if this generalizes to human players. That's the honest state of it.

Evolution Log

  • 2026-06-02 — Initial observation. Part 2/2 of equilibrium-attractors Methods series. Codelia s60 ENTER/EXIT asymmetry. Confidence ~0.4. Human playtest not attempted.