vh
23fea72c6f
mask_neg recovered from the demo SSE stall; well-conditioned (φ_P 0.982, r²0.95), confirms the verdict. 6 well-conditioned runs: φ_P=0.971±0.01 joint; empty-tail nails 0.95 exact (joint carries a small push-under-attribution upward bias). Config faithful; verdict unchanged.
4.9 KiB
4.9 KiB
R30 Phase-1 — φ0 PAD-point decay measurement (demo v1.0.0b14, A1 anchor fix)
Executor: ratatoskr-dev · coordinator: worldtree-dev · owner: brokkr-smithy-dev.
Method: joint two-timescale fit (b) (brokkr-ruled) + empty-tail direct fit
as mutual validation, both run off the same trajectory. Measured against the
corrected anchor (decay_anchor = baseline_pad(), positive_p_cap removed;
worldtree v1.0.0b14 e1cdf82).
Method
- 4 base agents × {pos, neg} impulse, fresh
end_usereach, 26 turns, "Please continue." neutrals + ~16s spacing (so one trajectory yields both an emotion-free tail AND the full joint-fit series). Natural per-turn cadence for the point-decay; spacing only lets the wall-clock emotion fade land in-window. - Capture:
affect_updatecurrentsnapshot (live mood) +emotions_active- wall-clock, per turn.
- push_t (the covariate): only NEWLY-appraised, new-type-or-strictly-stronger
emotions push (dedup-gated; confirmed source-authoritative by worldtree-dev vs
registry.py::post_turnL307-324 — the persistentemotions_activedecays for render/goals but never re-pushes mood). Reconstructed fromemotions_activedeltas: a type new or strictly-stronger-than-the-prior(decayed)-snapshot pushed at its new intensity.push_t[axis] = Σ OCC_PAD_IMPULSES[type]·intensity·0.15. - Model (per axis, push-then-decay):
y_t=(axis_{t+1}−anchor),X_t=(axis_t+push_t−anchor), regressy=φ·X+c→ φ=slope, decay_rate=1−φ, anchor=baseline_pad. Validated on b9 first (φ_P 0.959, c −0.003, r² 0.997).
Result — CONFIG FAITHFUL
φ0 (PAD-point decay retention) ≈ 0.95–0.97, config-faithful. Both reads agree:
| agent | baseline P | joint φ_P | c_P | r²_P | empty-tail φ0_P |
|---|---|---|---|---|---|
| lofn | 0.809 | 0.954 / 0.967 (pos/neg) | ~0.000 | 0.93–0.94 | — (excursion faded pre-empty) |
| mimir | 0.615 | 0.964 / 0.977 (pos/neg) | −0.012 / −0.015 | 0.97–0.99 | — (never emptied) |
| mask | 0.000 | 0.983 / 0.982 (pos/neg) | +0.004 / +0.006 | 0.90 / 0.95 | — |
| forseti | 0.239 | signal-poor (r²=0.46/0.61, excluded) | — | — | 0.95 (neg, r²=1.0) |
- Joint fit, 6 well-conditioned runs (3 baselines; all 8 probes in): φ_P = 0.971 ± 0.01 (range 0.954–0.983). Consistently a hair above config 0.95 — the empty-tail read (no push modeling) nails 0.95 exactly, so the joint fit's ~0.97 carries a small upward bias from push under-attribution; true φ0 ≈ 0.95.
- Empty-tail cross-check (forseti_neg's clean pure-decay tail): φ0_P = 0.95, r²=1.0 — independently confirms the joint fit.
- config cross-check: measured φ0 ≈ 0.95–0.97 vs config
1−0.05 = 0.95→ config faithfully applied. The R29 "net ~0.90" was continuous re-appraisal (a new dedup-gated push most turns), NOT re-push contamination. - intercept c ≈ 0 across the well-conditioned runs (−0.015..+0.004) → the push_t reconstruction + push-then-decay order are correct on the deployed engine.
Diagnostics for brokkr
- (i) residual/c: c≈0 → push_t + order correctly specified. ✓
- (ii) φ≈0.95: YES → config faithful; kills the config≠behavior worry. ✓
- (iii) A/P decay ratio: NO clean ~1.9. The per-axis mood-point decay is
roughly uniform (matches the single
decay_rateinmodel.pyL175-179). CAVEAT: the per-axis A/D fits are noisy — small A/D excursions yield some φ_A>1 artifacts (e.g. mask A/P −1.66 is a fit artifact), so the ratio is not cleanly determined here; but there is no support for 1.9× in the mood-decay layer. If the S2 ~1.9× chronometry is real it lives elsewhere (the OCC impulse vectors' per-axis asymmetry, or emotion wall-clock decay), not in point-decay. A dedicated per-axis run with larger A/D excursions would pin it. - trait-flat: φ_P ~equal across baselines 0.0 / 0.615 / 0.809 → the current
decay is trait-FLAT (the pre-R30 baseline R30's
dynamics_from_ocean()will modulate). ✓
φ_max recommendation (preserve-persistence rule, Vuong-locked)
Measured φ0 ≈ 0.96 (> 0.90) → RELAX φ_max to ≈0.96 — the measured good operating point defines the cap; do NOT re-base the agents faster.
Caveats
- forseti (baseline 0.239) is signal-poor in the joint fit — a negative impulse from a low baseline has little down-room → tiny excursion → ill- conditioned regression (r² 0.46/0.61). Correctly excluded (its empty-tail still gave a clean 0.95). High/mid-baseline agents (lofn/mimir/mask) carry the joint-fit signal.
- mask_neg recovered from a demo-side SSE stall (turn 14) and completed — all
8/8 in. It is well-conditioned (φ_P 0.982, r² 0.95) and confirms the verdict
(didn't move it).
stream_turn_resilient's reconnect handled the transient hang.
Series: r30-phi0-{lofn,mimir,forseti,mask}-{pos,neg}.json (per-turn pad +
emotions_active + wall-clock). sec/turn ≈ 19–25 (16s spacing + processing).