Drop7 Research
← Experiments

Paired SCREEN of fair D4 with the survival-instinct root filter (strict and literal) against unchanged fair D4 on 128 fresh development games

completedtier SCREENalgorithmicpublic-developmentpublic-policyEX-20260822-survival-instinct-root-filter-8531cbd9
No explanation has been written for this experiment yet. Add web/content/research/EX-20260822-survival-instinct-root-filter-8531cbd9.mdx and it will appear here. The registered protocol is shown below.

The registered protocol

With the bit-exact fast fair-D4 search at five strata (cache 60,000, worst-case work bound, terminal utility -1,000,000, policy seed 0xd7075eed), three arms play the same 128 ordered seeds 0xa52c0000-0xa52c007f: 'none' (unchanged), 'strict' (refuse a root column when the next disc of value >= 3 would land above its vertical-clear height AND its landing row run already exceeds its value, unless no column remains) and 'literal' (refuse every vertically dead landing that does not clear on arrival). The strict arm's paired score delta versus 'none' passes the gate below; the literal arm is a secondary measurement. Per-arm coverage (triggered decisions, overrides of the unfiltered column, no-column fallbacks) is recorded.

Candidate
fast-d4s5-survival-instinct-strict
approaches/lifetime-objective/survival-instinct/run.cpp
Comparator
fair-d4
approaches/fair-expectimax/reference/fair-only-depth4.cpp

Primary metric

paired mean whole-game score delta, strict minus none, 128 games, depth 4 five strata

Statistical unit: whole-game

Pass criteria, fixed in advance

  • CHECK gates passed before the lease was opened: mask unit cases, parity of the all-allowed filtered search with FastSearch (0 mismatches), mirrored masks and actions, thread-count determinism.
  • incompleteDecisionsTotal 0 and illegalDecisionsTotal 0 in the artifact.
  • strict minus none: bootstrap 95% lower bound > 0 AND Student-t 95% lower bound > 0.
  • strict minus none: candidate Q25 >= comparator Q25 AND both halves > 0.
  • strict arm overrides >= 2% of its decisions (otherwise the result is a non-test of the rule and is recorded as inconclusive regardless of the deltas).

On pass: Record valid + pass; register a fresh-block replication and a seven-stratum arm as successor experiments; register the filter as a candidate leaderboard policy (playground only).

On fail: Record valid + fail for this exact filter; report the literal arm and coverage as diagnostics; open no further cohort; the theory is not-supported-as-tested.

Data and reuse

CHECK gates read previously opened development probe seeds 0xa5278000-0xa52784ff. The screen opens the new, never-read lease 0xa52c0000-0xa52c007f exactly once; the lease is disjoint from every range in docs/exploratory/lease-map.md and from the ./do-research allocator's future grants (the allocator scans lease records). No other seed is read.

seed leases: SL-20260822T060000Z-a52c0000

What happened

valid run · outcome: inconclusivemixedpublic-developmentRS-20260822T233343Z-12becce9

Held-out SCREEN of the survival-instinct root filter, depth 4 five strata, 128 paired never-read games 0xa52c0000+. STRICT (refuse a landing that is vertically dead AND entombed by its row): mean 346838 vs unchanged 348808, paired delta -1970 (bootstrap 95% bounds -27738 to +22313, paired sd 171933, floor 24999), W-T-L 41-26-61, moves 100.95 vs 101.32; the filter refused a column at 811 of 12922 decisions and overrode the unchanged search's column at 254 (1.97%), just under the preregistered 2% coverage bar, so the primary gate is recorded as INCONCLUSIVE by its own rule: the strict rule neither helps nor hurts measurably and rarely applies. LITERAL (the owner's wording, refuse every vertically dead landing that does not clear on arrival; secondary arm): mean 251744, paired delta -97064 (LB -136887, UB -59500), W-T-L 43-0-85, moves 75.67 vs 101.32, overrides 1243 of 9686 (12.8%): a clear, valid negative — forbidding vertically dead landings outright costs about a quarter of the game's length. Both arms: 0 incomplete and 0 illegal decisions.

What it had to pass
  • CHECK gates passed before the lease was opened — observed: RUN-20260822T063817Z-a5d4c143: 7 of 7 (mask units re-run after a corrected test expectation)
  • incompleteDecisionsTotal 0 and illegalDecisionsTotal 0 — observed: 0 / 0
  • strict minus none: bootstrap and Student-t 95% lower bounds > 0 — observed: -27738 / -27150
  • strict minus none: Q25 >= and both halves > 0 — observed: Q25 delta -13513; halves +1721 / -5661
  • strict arm overrides >= 2% of decisions (else inconclusive) — observed: 254 of 12922 = 1.97%
Recorded metrics
strict
candidate
name
strict
games
128
score
mean
346838.4609
median
285244.5000
sd
234288.6063
q25
175358.2500
min
86,228
max
1,312,546
moves
mean
100.9531
q25
55
numberedClearsPerMove
2.0076
coverRevealsPerMove
1.1163
maxChainDepth
13
meanOccupiedCells
23.8341
censoredGames
0
incompleteDecisions
0
illegalDecisions
0
gamesAtOrAboveMillion
1
reference
name
none
games
128
score
mean
348808.3047
median
268,837
sd
246206.6767
q25
188871.5000
min
103,221
max
1,698,396
moves
mean
101.3203
q25
55
numberedClearsPerMove
2.0040
coverRevealsPerMove
1.1149
maxChainDepth
13
meanOccupiedCells
24.0289
censoredGames
0
incompleteDecisions
0
illegalDecisions
0
gamesAtOrAboveMillion
3
pairedScore
n
128
meanDelta
-1969.8438
pairedSd
171932.7446
bootstrapLower95
-27738.2355
bootstrapUpper95
22313.0406
studentTLower95
-27150.1196
studentTQuantile
1.6569
detectionFloor
24998.8202
detectionFloorNote
1.645 * sd / sqrt(n), the z-based floor defined in docs/research/status.md
wins
41
ties
26
losses
61
firstHalfMeanDelta
1721.0313
secondHalfMeanDelta
-5660.7188
q25Delta
-13513.2500
largestSwings
  1. -844,875
  2. -805,407
  3. -548,601
  4. 505,964
  5. 486,921
bootstrapSeedHex
0xb0071eaf
resamples
20,000
pairedMoves
n
128
meanDelta
-0.3672
pairedSd
46.6503
bootstrapLower95
-7.3516
bootstrapUpper95
6.2109
studentTLower95
-7.1993
studentTQuantile
1.6569
detectionFloor
6.7829
detectionFloorNote
1.645 * sd / sqrt(n), the z-based floor defined in docs/research/status.md
wins
29
ties
60
losses
39
firstHalfMeanDelta
0.7031
secondHalfMeanDelta
-1.4375
q25Delta
0
largestSwings
  1. -227
  2. -220
  3. -150
  4. 135
  5. 134
bootstrapSeedHex
0xb0071eaf
resamples
20,000
coverage
decisions
12,922
triggeredDecisions
811
overrides
254
noColumnFallbacks
22
literal
candidate
name
literal
games
128
score
mean
251744.0781
median
236792.5000
sd
112378.7779
q25
159800.2500
min
85,922
max
628,660
moves
mean
75.6719
q25
50
numberedClearsPerMove
1.8787
coverRevealsPerMove
1.0197
maxChainDepth
11
meanOccupiedCells
24.6577
censoredGames
0
incompleteDecisions
0
illegalDecisions
0
gamesAtOrAboveMillion
0
reference
name
none
games
128
score
mean
348808.3047
median
268,837
sd
246206.6767
q25
188871.5000
min
103,221
max
1,698,396
moves
mean
101.3203
q25
55
numberedClearsPerMove
2.0040
coverRevealsPerMove
1.1149
maxChainDepth
13
meanOccupiedCells
24.0289
censoredGames
0
incompleteDecisions
0
illegalDecisions
0
gamesAtOrAboveMillion
3
pairedScore
n
128
meanDelta
-97064.2266
pairedSd
265747.3813
bootstrapLower95
-136887.3000
bootstrapUpper95
-59499.6051
studentTLower95
-135984.0518
studentTQuantile
1.6569
detectionFloor
38639.3588
detectionFloorNote
1.645 * sd / sqrt(n), the z-based floor defined in docs/research/status.md
wins
43
ties
0
losses
85
firstHalfMeanDelta
-71884.9844
secondHalfMeanDelta
-122243.4688
q25Delta
-29071.2500
largestSwings
  1. -1,577,399
  2. -823,508
  3. -787,758
  4. -774,508
  5. -674,807
bootstrapSeedHex
0xb0071eaf
resamples
20,000
pairedMoves
n
128
meanDelta
-25.6484
pairedSd
72.1942
bootstrapLower95
-36.4305
bootstrapUpper95
-15.4375
studentTLower95
-36.2216
studentTQuantile
1.6569
detectionFloor
10.4970
detectionFloorNote
1.645 * sd / sqrt(n), the z-based floor defined in docs/research/status.md
wins
42
ties
10
losses
76
firstHalfMeanDelta
-18.7656
secondHalfMeanDelta
-32.5313
q25Delta
-5
largestSwings
  1. -425
  2. -225
  3. -212
  4. -210
  5. -185
bootstrapSeedHex
0xb0071eaf
resamples
20,000
coverage
decisions
9,686
triggeredDecisions
3,470
overrides
1,243
noColumnFallbacks
497
Limitations
  • Single 128-game development-tier screen; the strict arm's effect, if any, is below the 25,000-point floor of this cohort.
  • The strict coverage of 1.97% is within rounding of the 2% bar; the inconclusive label follows the preregistered rule literally rather than the spirit, and the deltas are reported in full.
  • The screen ran after the evolution's own screen on the same otherwise idle machine; decisions are deterministic, wall times are not timing-grade.
  • The literal arm's negative rejects the rule as a hard root constraint at depth 4; a soft penalty the search can weigh is a different candidate.