README: the composite series complete — exp010-exp014 all shipped

exp014_toggles: the toggle registry live — gates bit-exact (all-off==frozen, strip==identity); span gating BEATS all-on 1.0 vs 0.875; inverted control 0.667; denominator-damping characterized; 3-segment registry demo ledgered

exp013_collective: 10-anchor collective — 3-for-3 preregistrations (infill 0.549 predicted 0.5-0.8; simesh passenger; tokendiff 0.886); full NxN matrix vs frozen row; task_1 starvation boundary + subcap attribution inconclusive (honest findings); mixed-batch per-position dispatch

README: full series online — exp010 certifications (two-regime 2-seed; MATCH seed-dependence; OOD cell confirmed), exp011/exp012 shipped, exp013 running, exp014 queued

exp009 amendment: two-regime law s1 certification (story usage 0.821 vs 0.820; decouples replicate) + the preregistered OOD-band cell CONFIRMED (tokendiff usage 0.918 predicted >0.9; denoiser preserved to 3 decimals)

exp008 amendment: 2-seed certification (denoiser replicates to 2nd decimal; sampler law both seeds; class-cond honest two-seed noise)

exp005 amendment: seed-1 rows — the s0 MATCH verdict is SEED-DEPENDENT (s1 nospec falls 0.917/0.708/0.563 while real composite holds 0.938/1.0/1.0): specialists = seed-robustness, echoing heredity-as-robustness

exp012_bridges: CLIP bridge — first demonstrated cross-modal conditioning (0.273 vs 0.170 shuffled); BERT bridge honest negative x2 (predictability principle); roster consequence ledgered

exp011_newanchors: infill anchor (+18pt over real LM floor, dose-response) + simesh similarity (infonce 0.494 recall@1 over 465 distractors; SIGN-CODE Hamming retrieval 0.359 — contrastive pressure reorganized the discrete surface; recon_target boundary marker 2:1)

article_ft2 v2: full public citation pass — 29 external refs fetched+verified (0 hallucinated), 11 internal links verified, 3 corpus attributions corrected (Adam discipline -> memory-ft3; selector failure class -> ft1; attractors claim scoped), References + corpus sections added

article_ft2: The Aleph Moves Into a Pretrained Trunk — full writeup of exp001-009 (142 numbers ledger-audited; two-regime dispatch law; predictability principle; termination lesson; the family)

README: the overnight family — exp005 (control decides) / exp006 / exp007 / exp008 / exp009 shipped; two-regime dispatch law + predictability principle

exp008_tokendiff: discrete token diffusion via gated adapter — denoiser PASS (beats identity all levels, 189x at pure noise); greedy mode-collapse ledgered + sampling fix (diversity 0.92); class-conditionality NOT demonstrated (honest negative)

exp004 amendment: seed-1 certification (truncated 0.0 both seeds, ok 1.0/0.938; over-emission replicates 1.667/1.311) — 2-seed rider satisfied

exp009_family: 6 anchors, 11 sub-competences through one dense signed dispatch; domain-level modularity clean (story usage 0.82, surgical decoupling both directions); tokendiff absent-disclosed

exp005_taskmoe: routed family serves all registers; capacity control DECIDES — random frozen stacks match (0.979/0.979/1.0): register-level specialists were passengers; cross-ref exp009 domain-level load-bearing

exp007_math: math anchors, 4 ask-methods — format lock-in 4/4, no correctness tax, held-out f(x) 0.688->0.938; sep peak 0.496 @L23

exp006_story: story anchors, 4 request registers — cont/inst/json 1.0, kw UNMOVED 0.083 (format learning != constraint compliance); strongest register sep at the time (L8 0.462)

README: exp004_composite shipped (row-aligned SFT passes the gate 16/16); exp005 running

exp004_composite: composite construction gate — packed-window termination FAIL (ledgered) -> row-aligned SFT v2 PASS 16/16 structural_ok 1.0; memory rider (checkpointing + chunked CE)

README: experiment line updated with exp001-003 statuses, links, headline numbers + exp004/005 plan

exp003_instruct: hallucination reduction quantified (specialist grounding 0.979/0.958 vs frozen 0.542, >10x fewer failures); register capture found (specialist answers every register in its own); multi-task resolves it (all registers >=0.875); register-differentiation hypothesis CONFIRMED on the sign-code surface (2x deep-layer separation vs the wikitext stack, 2 seeds)

exp002 Wave B seed-1 folded in: ordering budget-stable certified 2 seeds (mlp 13.899/13.889 vs relay_pw 13.983/13.932); relay convergence seed-dependent — claim corrected

exp002_refine: the aleph-addressed patchwork consumer wins (13.997 mean, 2 seeds, first relay under 14); width/strobe saturated; coverage beats concentration (-0.48); wide MLP diverges 1/2 seeds vs 0% for all relays; ordering budget-stable at 6k

exp001_relay: the retrofit works at 0.5B (frozen 17.80 -> relay 14.09 / mlp 13.93, <0.6% trainable, 2 seeds); GPT-2 ppl ordering inverts (substrate-scoped) while the gate mechanism replicates (relay gates grow, MLP gates shrink) - the gate-vs-ppl dissociation

geolip-aleph-qwen: the AlephLM generation stake (certified-component composite on Qwen2.5-0.5B; experiment numbering restarts at 1)

initial commit