Trader evolution - genome families

15 families (6 alive, 9 closed) - updated 2026-08-14 15:55 UTC - every family carries its HOPE (the live reason to keep working it) or its cause of death. back to the trader page

shipped running contested dead

Core pipeline shipped 60 attempts | best: champion: blind -4.0 holdout n=99; the live trader

hope: alive - new information feeds keep joining it (BAR 2026-08-14)

blind (informed-person) shipped c1

standard (note+read+clamp) shipped c1, -3.1

champion (+modules, 3-read, clamps) shipped c1-2, -4.0

BAR equity/commodity barriers surveillance 2026-08-14, -5.86 p=.0625 exact; kill-switch armed

Deep research (v2) dead-spawned 12 attempts | best: v2direct tails 62/62 side-agreement

hope: none as champion-beater (reader subtracts; depth hurts mid-band) - but its tails spawned Selection

v2direct dead +1.0 vs champ

v2read dead reader subtracts, CI-clean

mid-band 2nd pass dead +3.3 CI-clean wrong way

Selection / gating live-contested 15 attempts | best: tail-gate KILLED by its own prereg (decisive n=63 ~-0.14 vs -0.5 floor); selection v2 shipped

hope: live-trigger routing only (tailGate died correctly; costRouterShadow + DEF-H accruing now); offline tables refused budget per sign1

tail-gated hybrid contested LIVE; kill block running, floor -0.5

cost router blind<0.10 pending ties champ at 1/10 cost; ship held

1536 mechanical gates dead LOO-CV ~0 (sel1)

sd-gate held approved-as-amended, awaiting kill verdict

tailGate live block dead preregistered kill: pooled decisive n=63 (BAR-era -1.76 n=48, HARD +5.06 n=15) misses -0.5 floor; reverted same hour

ADJ1 claim-audit adjudicator candidate downward range-exits -5.67 exact p=0.0277 (evidence-reading reaches where read statistics cannot: venue outside reads range on 80% of flagged); prereg-failed on silence-rule defect; ADJ2 guards running

ADJ2 guarded adjudicator saturated prereg FAIL (mean -1.92 vs -3.50, flip 9-9); guards fixed regressions but suppressed gains; downward-exit kernel real twice at design floor

DEF-H marker filter candidate marker minus self-contradicted headers: -7.62 flip 16/6 OOS, beats CONST control; shadow-journaling live, adjudicates at k>=39 clusters

SIGN1 routing-premise adjudication validated resolved null at fleet scale: no offline detector transfers (750 pairs, planted control excludes >=2pt); Integration law amended - live measured-fire triggers + constant-rule co-primary only

Ensembles / aggregation dead 10 attempts | best: median-of-5 blind styles -0.97 (not clean)

hope: NONE - correlated-error dilution measured repeatedly; closed

median-of-3 researchers dead dilutes best member

LMSR internal market dead killed twice

logit pooling dead null; compression is in the reads

Post-hoc corrections dead 14 attempts | best: S2 extremity -0.465 on set20 VAL only

hope: NONE cross-population (LOSO identity, xb2); S2 live prereg runs out its terms

extremity recal / shift population-local same effect found twice; halves off-population

correction layer (9 genotypes) dead fits identity, LOSO

coherence normalization dead killed 3x

Computed info sources shipped 9 attempts | best: SOURCE FETCH shipped: -16.68 CI[-23.57,-10.66] VAL confirm, flip 19-1, 0 regressions (largest validated effect of the campaign)

hope: host-by-host expansion through the same prereg door (billboard failed entry at +2.27); BAR still under surveillance

crypto barrier math shipped c1

BAR (yahoo) surveillance shipped 2026-08-14

REC archive rates dead +2.46 harm as lines

NN retrieval dead similarity != structure

I7 resolution-source fetch running dispatched 2026-08-14

i7src source-fetch (ISSUE-7) shipped fetch the resolution source the market itself names: VAL ACCEPT -16.68 n=28/21 fams, flip 19-1; allowlist wunderground+gol.gg; surveillance contract rides the ship commit

Venue modeling alive 8 attempts | best: favorite-underpricing law verified (all splits)

hope: ALIVE - venue-prior candidate (vm2, honest double-label) + truthbet3 live probes

venue calibration law validated 0.5-0.7 favorites +16-35pp

venue-prior (same question history) candidate -0.86/-0.89, needs clean prereg

inverse-calib transform dead champion IS a price estimator

truthbet3 probes running shadow+live, class adjudication n>=15

Joint / coherence estimation saturated 6 attempts | best: joint TRAIN -1.0, VAL -1.2 (never clean)

hope: THIN - K=3 screen rejected (flip 10-7, one +20.4 family); revisit only with a new mechanism

joint family estimation rejected i4jf K=3: mixed passes, regression

axiom projection dead estimates already coherent (0.795)

RADICAL: agentic trader (T1) alive 1 attempts | best: PARITY with champion at iteration zero (+0.62/+0.80, within noise, 2.1x cost); agent-decided tool use VALIDATED as mechanism (-4.4/-2.4 vs its own no-tool arm, flip-p .009/.024 both populations)

hope: STRONG - zero-iteration parity implies headroom; G1 = champion-grade base + the agentic loop (the validated mechanism on the stronger anchor)

AT1 (ReAct, 4 steps, mechanical blindness) candidate parity; floors missed properly

agent-decided tool use (mechanism) validated c1 revival condition met; flip co-primary clean 2 populations

G1: champion-base + agentic loop queued dispatched 2026-08-14

RADICAL: program-synthesis (T2) saturated 1 attempts | best: architecture flawless (154/154 programs, 0 trivial); approach REJECTED +5.76 CI-clean, flip 22/45 p=.007

hope: CLOSED as a general trader - computation wins only where the question IS computational (the modules family already harvests that class); narrow revival = code-when-computational routing only

T2 full trader rejected +5.76 CI-clean worse; saturation prereg closed round 2, 0.6 CC returned

RADICAL: scenario-sim (T3) saturated 1 attempts | best: trees parsed 99%; all six aggregators lose; best +3.54 CI-clean at TRAIN screen

hope: CLOSED - externalized arithmetic over model-stated step probabilities inherits every step's compression; consistent with T2 and the forced-product grave

T3 full trader (6 aggregators) rejected +3.54 CI-clean at screen; never reached VAL

RADICAL: case-based (T4) alive 1 attempts | best: ZERO-COST flip-parity with champion on BOTH populations (33/31, 42/41; mean +1.66/+1.61 not separable); negative controls fail = real retrieval

hope: STRONG - free champion-tier baseline; model-in-loop arm prereg'd awaiting funds; obvious cross with T1 loop

pure precedent retrieval (no model) candidate 0.000 CC/mkt, flip-parity 2 populations

model-in-loop arm pending-funds prereg'd, TRAIN sighting only

RADICAL: simulated crowd (T5) saturated 1 attempts | best: crowd 0.0579 vs champ 0.0599 (null, flips 19-24); 5 clearing mechanisms x real diverse-diet fleet arms; floors unmet both populations

hope: CLOSED - aggregation cannot beat its best member even with structurally diverse diets; the strongest test the c1 law ever faced, passed

T5 clearing engine (5 mechanisms) rejected free; graveyard law upheld at its limit

Model axis (M1) saturated 7 attempts | best: m3 never-seen-population prereg: gemini-all -0.31 [-2.54,+2.02] flip p=0.12, ROUTED -0.13 - resolved nulls

hope: CLOSED (ISSUE-9 saturated): m2 gains were population overlap (69/88 shared fams). Revival: new model generation, or cost-only case after the token meter is fixed (luna reports 3 input tok/call - broken)

M1 sweep queued blind sweep -> role sweep -> full config

m1val avg@5 confirm rejected RETRACTED by m2 instrument audit (meter-cap trip, family-key bug, cost sign inverted: gemini 8.17x MORE per read)

m2 routed prereg validated gemini-all -2.34 [-4.16,-0.99] flip p=.0006, blocked by 8 fams >+5; geo route rejected; history router -2.99, blockers 8->2

m3 no-overlap prereg dead 170 never-seen clusters: engine and history router both resolved nulls (band 1.11 vs floor -1.50); m2 r=+0.678 was shared-family artifact

Fitted prior (FIT1) dead 1 attempts | best: EB table beat class-prices (validated) but lost to blend-CONST in both forms (dead1 routed replication)

hope: CLOSED: the prior adds a constant, not information, on our population. Revival: an archive whose population matches liquid-coherent books

EB structural table (H1) validated vs class-prices: set20 VAL -13.64 [-23.46,-4.27] flip p=.0498; set21 -11.71 [-21.13,-3.89] flip p=.0246

blend into champion (H2) rejected blend-CONST 0.014 control matches it (-0.22 vs -0.20): the blend added a constant, not information

GBM/ridge rungs dead plain EB 0.1468 beats GBM-LAD 0.1586 / logit-GBM 0.1719 / ridge 0.3191 on 13.8k+17.5k held-out; L2-vs-MAD objective mismatch measured

dead1 routed replication dead flagged-slice prior -6.28 but CONST-blend -6.90: routing does not rescue it

Selection rule: budget goes only to families with stated hope; "many attempts, no hope" closes a family regardless of how clever the next variant looks. Full records: the lab's GRAVEYARD, ISSUES and papers.