vh/ratatoskr
1
0
Fork 0
Code Issues 4 Pull Requests Actions Packages Projects Releases Wiki Activity

docs(diagnostics): R29 low-baseline step-response — hedonic asymmetry is emergent

Browse Source 
forseti (P0.239) + mimir (P0.615) ±impulse step-response resolves the lofn
ceiling confound. Findings: decay tau symmetric across signs (~0.9/turn);
raw impulse gain ~symmetric near neutral (forseti +0.034 vs -0.034); the
Frijda 'negatives loom larger' is emergent from the neutral anchor x non-zero
baseline interaction, not a primitive. Calibration fork answer: neither
asymmetric-decay nor negative-bias primitive is warranted — the levers are
overall gain + the neutral anchor.
This commit is contained in:
vh
2026-07-02 14:13:37 -07:00
parent 1819e05589
commit 61ff2da703
9 changed files with 1395 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/diagnostics/r29-pad-findings.md
+31
View File
@@ -122,6 +122,37 @@ re-stimulation). Series: `r29-pad-step-{pos,neg}-lofn.{json,csv}`.
ceiling-blocked), not the decay constant. A clean positive-τ / true Frijda test
needs a low-baseline agent (forseti/mimir, pos+neg from one baseline).
## Low-baseline step-response — the hedonic asymmetry is EMERGENT, not primitive
Ran ±impulse step-response on forseti (P0.239) and mimir (P0.615), up-headroom
baselines. Series: `r29-pad-step-{forseti,mimir}-{pos,neg}.{json,csv}`.
| agent (baseP) | POS impulse (max above baseline) | NEG impulse (isolated, neg−pos min) | decay retention/turn |
|---|---:|---:|---:|
| lofn (0.809) | +0.000 (ceiling) | −0.175 | ~0.92 both signs |
| mimir (0.615) | +0.000 (decay-masked) | −0.175 | ~0.90 both signs |
| forseti (0.239)| **+0.034** | **−0.034** | ~0.93 both signs |
- **Positive impulse gain is small but real (~+0.03)** — visible above baseline only
when the baseline is low enough that decay-to-0 doesn't immediately mask it
(forseti shows +0.034; at higher baselines the same push is swamped by faster
absolute decay). The "positive can't move" symptom is **gain + decay competition**,
not only the positive_p_cap ceiling.
- **Decay τ is symmetric across signs** (~0.9 retention/turn both) — no sign-dependent decay.
- **Near a neutral baseline the raw gains are ~symmetric** (forseti +0.034 vs −0.034).
The large apparent asymmetry at high baselines is the **decay-to-neutral ×
positive-baseline interaction** — decay OPPOSES positive excursions and REINFORCES
negative ones when baseline > 0.
**→ Calibration-fork answer (asymmetric decay vs ESM offset+bias):** the data supports
**neither** an asymmetric-decay parameter (τ symmetric) **nor** a strong negative-bias
primitive (raw gain ~symmetric near neutral). The Frijda "negatives loom larger" here is
**emergent** from the neutral anchor + non-zero baselines, not a primitive to calibrate.
The two real levers stay: **overall emotion→PAD gain** (too low, both signs) and the
**neutral decay anchor** (which manufactures the apparent valence asymmetry for any
persona whose trait baseline ≠ 0). Caveat: no base agent has a truly neutral baseline;
forseti (0.239) is closest and shows near-symmetry — a definitive gain-symmetry isolation
would need a neutral-baseline persona.
## What R29 asked for vs what ratatoskr can serve
| Item | Status |
|---|---|