engineering/arborist
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
arborist/docs
History
russell@unturf.com da5af2dfaa
#000052 §3.2.1 round-2: extend size spectrum to 7 models (30MB-560MB, ~20×) — real model science from a wider sweep 
7 models on the 4090 against the 26-pair candidate-bench:

  bge-reranker-large    560MB  +6.285  ← new primary (most cushion)
  ms-marco-electra-base 110MB  +5.349  (cost alternate — 5× smaller, 87% the margin)
  MiniLM-L-4-v2          50MB  +3.187
  bge-reranker-base     280MB  +3.076
  MiniLM-L-6-v2          80MB  +3.053  (was primary before §3.2.1)
  MiniLM-L-12-v2        130MB  +2.663  (deeper ≠ better — same pattern as §7 #24)
  MiniLM-L-2-v2          30MB  -1.970  ← capacity floor (NOT separable, 10/12 catch)

Findings:
- Biggest is best WITHIN a family (bge-large > bge-base, 2× size → 2× margin).
- Across families: electra-base (110MB) beats bge-base (280MB) — 60%
  smaller, 75% more margin. Architecture/training corpus > parameter count.
- Depth non-monotonic within MS-MARCO MiniLM: L-4 > L-6 > L-12.
- Capacity floor between L-2 (30MB, fails) and L-4 (50MB, separates).
- 6 of 7 clean-separate the candidate-bench. The 'reranker catches
  Zionist-style mis-cites' claim is a property of competent rerankers
  as a class — once above the floor.

Manifest primary moves to bge-reranker-large for cushion. demote_below_score
STAYS null — clean candidate-bench doesn't predict real-pipeline behavior
(§7 #18→#27 = 6 verdict flips). Expect a walk-back at §3.2.2 step 2.
2026-05-13 10:05:56 -04:00
..
_source #000035 closed — v7 substrate spec into the repo; § 9.10 folded in 2026-05-11 17:06:19 -04:00
diagrams docs/diagrams: rename aborist-modules → arborist-modules (post-rename cleanup) 2026-05-08 07:50:28 -04:00
tickets #000052 §3.2.1 round-2: extend size spectrum to 7 models (30MB-560MB, ~20×) — real model science from a wider sweep 2026-05-13 10:05:56 -04:00
bench-maxing.md aborist/arborist 2026-05-07 09:31:49 -04:00
benchmarks.md aborist/arborist 2026-05-07 09:31:49 -04:00
calculator-test-patterns.md ticket #000036 Tier-2: dav1d Option B (conservative B1 envelope) applied in v1 2026-05-11 07:06:50 -04:00
cti-architecture.md aborist/arborist 2026-05-07 09:31:49 -04:00
lexical-first-rationale.md docs: lexical-first-rationale.md — why the cheap retrieval path is the default 2026-05-12 09:23:50 -04:00
mesh.md aborist/arborist 2026-05-07 09:31:49 -04:00
pi-star-composition.md pi_star: land ticket #000015 (π* domain library + composition algebra) 2026-05-07 16:51:33 -04:00
qa-modes-bench.md #000048 step 2.4 — parse_pointer_claims clause segmentation 2026-05-11 17:09:06 -04:00
seven-point-program.md tests/doc_counts: regression test for numeric claims in docs/ (4x drift fix) 2026-05-10 16:15:52 -04:00
soft-hash-channel-analysis.md docs/#000018 §9.2: mark resolved — φ_PRG = HMAC-SHA-512 (#000035 closed) 2026-05-11 17:20:48 -04:00
soft-hash-channel-t3-bound.md ticket #000036: add KAT-regen tooling + close 2026-05-11 08:02:25 -04:00
spec-methodology.md docs: land ticket #000019 (spec methodology for π*, V, policy fields) 2026-05-07 16:53:28 -04:00
TICKETS.md #000054: acronym-parens concept extractor (closes abbreviation→expansion retrieval gap) 2026-05-13 07:00:25 -04:00
tool-action-dag-design.md docs: add tool-action-dag-design.md research path (pre-ticket) 2026-05-07 19:47:50 -04:00
v7w-frontier-catalog.md #000013 closed: v7-W spatial-temporal substrate paper + namespace 2026-05-09 15:00:05 -04:00
v8-fork-score.md CLI: arborist v8 score → arborist substrate score 2026-05-10 09:12:34 -04:00
warrant-substrate-cookbook.md docs: bump warrant-substrate-cookbook AUTOCOUNT 20 -> 28 for #000054 tests 2026-05-13 07:01:02 -04:00
zk-frontier-bench.md #000016 parked: ZK frontier-proof bench plan + wire protocol 2026-05-09 15:05:08 -04:00
zk-wire-protocol.md #000016 parked: ZK frontier-proof bench plan + wire protocol 2026-05-09 15:05:08 -04:00