Automatic score-decay ablation — v1

Decision: retain-original. 16704 measured encounters. Only automatic score decay changes, from 0.025 to 0; sampled actions still update. No production change.

Suite/case	Phase	No-decay − original	No-decay online − boundary frozen	Original online − boundary frozen
legacy/reverse	switch	3.38 [1.75, 5.25]	19.00 [15.13, 23.13]	17.00 [12.62, 21.50]
legacy/feint	switch	1.25 [0.63, 1.88]	-4.63 [-5.88, -3.38]	-4.88 [-6.13, -3.63]
legacy/cover	switch	0.00 [0.00, 0.00]	0.00 [0.00, 0.00]	0.00 [0.00, 0.00]
synthetic/bait-switch/staggered-lanes	switch	0.17 [-3.50, 3.50]	9.00 [4.00, 14.50]	8.58 [3.66, 13.83]
synthetic/bait-switch/staggered-lanes	return	2.75 [0.25, 5.76]	2.00 [-2.75, 6.75]	0.25 [-3.75, 4.50]
synthetic/bait-switch/offset-islands	switch	0.50 [0.00, 1.50]	3.00 [-3.50, 8.25]	2.50 [-3.75, 7.25]
synthetic/bait-switch/offset-islands	return	0.75 [0.00, 2.25]	6.50 [1.25, 12.00]	6.25 [0.75, 12.50]
synthetic/reactive/staggered-lanes	switch	-1.50 [-4.50, 0.00]	1.75 [-2.75, 5.75]	2.25 [-4.25, 7.00]
synthetic/reactive/staggered-lanes	return	1.50 [0.00, 3.50]	3.50 [1.00, 6.25]	4.50 [0.75, 9.25]
synthetic/reactive/offset-islands	switch	0.00 [-2.00, 1.50]	3.25 [0.75, 6.25]	2.75 [0.00, 6.00]
synthetic/reactive/offset-islands	return	0.00 [-1.00, 1.25]	2.25 [-1.25, 6.50]	2.00 [-1.50, 6.25]
synthetic/burst-patrol/split-gates	switch	0.00 [0.00, 0.00]	4.08 [0.17, 8.58]	3.83 [0.08, 8.08]
synthetic/burst-patrol/split-gates	return	0.58 [0.00, 1.75]	2.75 [0.58, 5.08]	2.42 [-0.58, 5.50]
synthetic/burst-patrol/shelves	switch	0.25 [-0.50, 1.00]	2.75 [0.00, 5.50]	3.33 [-0.25, 8.00]
synthetic/burst-patrol/shelves	return	0.00 [0.00, 0.00]	7.42 [1.75, 14.00]	6.17 [0.75, 12.42]
synthetic/threat-retreat/split-gates	switch	0.25 [0.00, 0.75]	1.25 [-2.25, 6.00]	1.75 [-2.25, 6.50]
synthetic/threat-retreat/split-gates	return	0.75 [0.00, 2.25]	4.75 [1.50, 9.00]	3.75 [0.25, 7.25]
synthetic/threat-retreat/shelves	switch	0.00 [0.00, 0.00]	9.83 [1.08, 19.34]	9.08 [1.00, 17.50]
synthetic/threat-retreat/shelves	return	-1.25 [-3.75, 0.00]	2.00 [-1.50, 5.50]	5.00 [0.75, 10.00]
synthetic/fresh	switch	0.13 [-0.13, 0.38]	4.48 [2.67, 6.48]	4.50 [2.35, 6.79]
synthetic/fresh	return	0.02 [-0.79, 0.71]	4.23 [1.81, 6.81]	4.33 [1.98, 6.83]
synthetic/known	switch	-0.21 [-1.42, 0.94]	4.25 [1.56, 6.94]	4.02 [0.81, 6.94]
synthetic/known	return	1.25 [0.25, 2.44]	3.56 [0.69, 6.56]	3.25 [0.56, 6.25]

Values are HP damage, late eight rounds for legacy and late four for synthetic. Brackets: 95% paired seed-bootstrap intervals. Primary: fresh switch aggregate; all other intervals exploratory.

Locked gates

fresh_gain: FAIL
fresh_post_switch_learning: PASS
legacy_feint_gain: FAIL
legacy_feint_learning_nonnegative: FAIL
no_large_case_regression: PASS

Simulation budget

Suite/variant	Encounters	Steps
legacy/original	2016	1943011
legacy/no-decay	2016	1932557
synthetic/original	6336	7137253
synthetic/no-decay	6336	7135609

Equal encounter histories per variant; record actual simulation steps and process CPU. No fixed-compute efficiency claim.

Limits

Neither decay nor lack of decay repairs missing joint-role credit or adds route memory.
Fresh behaviors/maps form a small prospective screen and become known regression fixtures after this run.
Direction changes are scripted, bots do not model or learn the enemy, and no human enjoyment is measured.
Reward, representation and runtime remain unchanged. No fixed-policy superiority claim is tested here.
Original and candidate are separately trained: the primary effect covers the complete configured learning history, not just post-switch decay.
Geometry validity and unit tests may run before the source freeze using non-evaluation seeds; no reserved combat may guide implementation.

Reproduce

npm run evaluate:retention -- --out target/retention-replication

npm run verify:retention replays all combats and separately reconstructs all core feedback from physical evidence.