SPB Git

spb/modelmap Public License

Internal cartography of local LLMs on Apple Silicon — registered, gated, negative-first. Public atlas at modelmap.io.

Python 66.3% JavaScript 24.5% CSS 8.1% Shell 0.7%

Phase 4: candidate ranking — 24 gaps scored on ten axes, 4 candidates selected

candidate_01 weight-only pre-screen (G06+G08/G03/G09), candidate_02
quantization deformation atlas (G01+G02), candidate_03 localization index
(G17+G11/G12), candidate_04 localvm working-set bridge (G24, reduced).
Mandatory substrate scheduled first: expH, expA/expD noise floors, map
cards v0, ablation-curves rule.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Simon-Pierre Boucher committed 4 h ago (Aug 12, 2026) parent d25cc06

Showing 6 changed files with +220 and −12

modified experiments/candidate_01/README.md +9 −4
@@ -1,6 +1,6 @@
1 1 ---
2 2 project: modelmap
3 document: candidate_01
3 +document: candidate_01 — Weight-only pre-screen
4 4 author: Simon-Pierre Boucher
5 5 contact: contact@spboucher.ai
6 6 website: https://modelmap.io
@@ -8,7 +8,12 @@ created: 2026-08-12
8 8 status: draft
9 9 ---
10 10
11 # candidate_01
11 +# candidate_01 — Weight-only pre-screen
12 12
13 Reserved for a Phase 7 prototype candidate — selected from `research/candidate_ranking.md` (Phase 4).
14 Not yet assigned: candidates are chosen only after the Phase 1-3 literature work.
13 +Assigned by Phase 4 (`research/candidate_ranking.md`). Core gaps: G06 (+G08 anchor, G03/G09 extensions).
14 +
15 +**First falsifier:** super-weight causal catalog on 2 models; then Pythia-410M weight-features-vs-probe-maps regression, kill at cross-validated R² < 0.2.
16 +
17 +Status: assigned — prototyping starts after the mandatory substrate
18 +(expH capture-cost frontier, expA/expD noise floors, map-cards v0).
19 +Hypothesis will be registered in `hypothesis.md` before any run (charter §10).
modified experiments/candidate_02/README.md +9 −4
@@ -1,6 +1,6 @@
1 1 ---
2 2 project: modelmap
3 document: candidate_02
3 +document: candidate_02 — Quantization deformation atlas
4 4 author: Simon-Pierre Boucher
5 5 contact: contact@spboucher.ai
6 6 website: https://modelmap.io
@@ -8,7 +8,12 @@ created: 2026-08-12
8 8 status: draft
9 9 ---
10 10
11 # candidate_02
11 +# candidate_02 — Quantization deformation atlas
12 12
13 Reserved for a Phase 7 prototype candidate — selected from `research/candidate_ranking.md` (Phase 4).
14 Not yet assigned: candidates are chosen only after the Phase 1-3 literature work.
13 +Assigned by Phase 4 (`research/candidate_ranking.md`). Core gaps: G01 (+G02, then G04/G05).
14 +
15 +**First falsifier:** Qwen3-0.6B FP16 vs Q4 probe drift vs seed/dataset noise floor: 10 properties × 5 seeds × 2 datasets.
16 +
17 +Status: assigned — prototyping starts after the mandatory substrate
18 +(expH capture-cost frontier, expA/expD noise floors, map-cards v0).
19 +Hypothesis will be registered in `hypothesis.md` before any run (charter §10).
modified experiments/candidate_03/README.md +9 −4
@@ -1,6 +1,6 @@
1 1 ---
2 2 project: modelmap
3 document: candidate_03
3 +document: candidate_03 — Localization index (Q1 instrument)
4 4 author: Simon-Pierre Boucher
5 5 contact: contact@spboucher.ai
6 6 website: https://modelmap.io
@@ -8,7 +8,12 @@ created: 2026-08-12
8 8 status: draft
9 9 ---
10 10
11 # candidate_03
11 +# candidate_03 — Localization index (Q1 instrument)
12 12
13 Reserved for a Phase 7 prototype candidate — selected from `research/candidate_ranking.md` (Phase 4).
14 Not yet assigned: candidates are chosen only after the Phase 1-3 literature work.
13 +Assigned by Phase 4 (`research/candidate_ranking.md`). Core gaps: G17 (+G11/G12 as measurement layers).
14 +
15 +**First falsifier:** the index must read ≈1 on Tracr-compiled circuits and ≈0 on random-init models; otherwise it is broken before touching real models.
16 +
17 +Status: assigned — prototyping starts after the mandatory substrate
18 +(expH capture-cost frontier, expA/expD noise floors, map-cards v0).
19 +Hypothesis will be registered in `hypothesis.md` before any run (charter §10).
added experiments/candidate_04/README.md +19 −0
@@ -0,0 +1,19 @@
1 +---
2 +project: modelmap
3 +document: candidate_04 — Working-set bridge to localvm (reduced scope)
4 +author: Simon-Pierre Boucher
5 +contact: contact@spboucher.ai
6 +website: https://modelmap.io
7 +created: 2026-08-12
8 +status: draft
9 +---
10 +
11 +# candidate_04 — Working-set bridge to localvm (reduced scope)
12 +
13 +Assigned by Phase 4 (`research/candidate_ranking.md`). Core gaps: G24 (+G09).
14 +
15 +**First falsifier:** domain-conditioned block-ablation working sets on a 3B model; < 10% reduction vs global working set kills per-domain prefetching.
16 +
17 +Status: assigned — prototyping starts after the mandatory substrate
18 +(expH capture-cost frontier, expA/expD noise floors, map-cards v0).
19 +Hypothesis will be registered in `hypothesis.md` before any run (charter §10).
modified research/LOG.md +41 −0
@@ -154,3 +154,44 @@ prototype candidates. Prior expectation (to be tested by the scoring, not
154 154 assumed): the quantization cluster and the weight-only pre-screen cluster
155 155 combine high novelty with low cost; G10/G23 are infrastructure that every
156 156 other candidate needs regardless of ranking.
157 +
158 +---
159 +
160 +## 2026-08-12 03:50 EDT — Phase 4: candidates ranked and selected
161 +
162 +**Question.** Which of the 24 gaps become prototype candidates, and in what
163 +order does everything else get built?
164 +
165 +**Action.** Wrote `research/candidate_ranking.md`: full 24×10 score table
166 +(convention: 10 always favorable, including complexity and risk), grouped
167 +written reasoning, and selection. Updated `experiments/candidate_01–04`
168 +READMEs with assignments.
169 +
170 +**Selection:**
171 +- **candidate_01 — Weight-only pre-screen** (G06 + G08 anchor, G03/G09):
172 + kill at cross-validated R² < 0.2 on Pythia-410M.
173 +- **candidate_02 — Quantization deformation atlas** (G01 + G02, later
174 + G04/G05): Qwen3-0.6B FP16-vs-Q4 probe drift vs noise floor.
175 +- **candidate_03 — Localization index** (G17 + G11/G12 as measurement
176 + layers): must separate Tracr (≈1) from random-init (≈0) first.
177 +- **candidate_04 — localvm working-set bridge** (G24 + G09), reduced scope:
178 + run the domain-conditioned ablation falsifier before committing.
179 +- **Mandatory substrate first, outside the ranking:** G20 (expH cost
180 + frontier), G10 (expA/expD noise floors), G23 (map cards v0), G13
181 + (ablation-curves rule, adopted at zero cost).
182 +
183 +**Interpretation.** The infrastructure quartet (G23 87, G20 86, G10 82,
184 +G13 75) tops the table by construction — every candidate presupposes it.
185 +The three main candidates fail on independent kernels (weights-regression /
186 +drift-vs-floor / index-validation), satisfying the charter's demand for
187 +genuinely distinct bets; each yields versioned artifacts even when negative.
188 +Cross-model coordinate work (G14–G16) is deliberately blocked on noise
189 +floors — alignment claims without floors are the field's known failure mode.
190 +
191 +**Decision.** Phases 1–4 complete. Next: Phase 5 (experimental framework) —
192 +implement the substrate in order: (1) expH harness (hook overhead MLX vs
193 +MPS, storage formats on APFS, streaming throughput) reusing localvm
194 +patterns; (2) expA probe-reliability harness with the full controls
195 +doctrine; (3) map-cards v0 schema in src/modelmap/atlas enforced by
196 +tools/publish.py. Register hypotheses in experiments/micro/expH and expA
197 +before their first runs.
added research/candidate_ranking.md +133 −0
@@ -0,0 +1,133 @@
1 +---
2 +project: modelmap
3 +document: Candidate ranking (Phase 4)
4 +author: Simon-Pierre Boucher
5 +contact: contact@spboucher.ai
6 +website: https://modelmap.io
7 +created: 2026-08-12
8 +modified: 2026-08-12
9 +status: reviewed
10 +---
11 +
12 +# Candidate ranking — 24 gaps scored on ten axes
13 +
14 +Phase 4 deliverable (charter §7). Every gap from `research_gaps.md` scored 1–10
15 +per axis. **Convention: 10 is always favorable** — for *Implementation
16 +complexity*, 10 = simple; for *Risk*, 10 = low risk. Total = sum (max 100).
17 +Scores are prior judgments to be revised as evidence lands; the reasoning
18 +column is the commitment, the numbers are its summary.
19 +
20 +Axes: **SV** scientific value · **ES** epistemic soundness · **NV** novelty ·
21 +**FA** feasibility on Apple Silicon · **RP** reproducibility potential ·
22 +**GN** cross-model generality · **AV** public atlas value · **IC** implementation
23 +complexity (10=simple) · **SY** synergy with localvm · **RK** risk (10=low).
24 +
25 +## Score table (sorted by total)
26 +
27 +| Gap | SV | ES | NV | FA | RP | GN | AV | IC | SY | RK | **Total** |
28 +|---|---|---|---|---|---|---|---|---|---|---|---|
29 +| G23 map cards schema | 7 | 8 | 9 | 10 | 10 | 10 | 10 | 8 | 6 | 9 | **87** |
30 +| G20 capture cost frontier (expH) | 6 | 9 | 9 | 10 | 9 | 7 | 9 | 7 | 10 | 10 | **86** |
31 +| G17 localization index | 9 | 9 | 8 | 9 | 9 | 9 | 9 | 7 | 6 | 8 | **83** |
32 +| G10 noise-floor atlas (expA/D) | 9 | 8 | 8 | 8 | 10 | 9 | 9 | 7 | 5 | 9 | **82** |
33 +| G06 weight→activation regression | 10 | 7 | 9 | 9 | 9 | 8 | 9 | 7 | 8 | 5 | **81** |
34 +| G08 super-weight cartography | 6 | 9 | 6 | 10 | 10 | 8 | 8 | 9 | 5 | 9 | **80** |
35 +| G01 probe-drift across quantization | 8 | 7 | 8 | 9 | 9 | 8 | 9 | 8 | 5 | 8 | **79** |
36 +| G11 method-agreement maps | 9 | 9 | 8 | 8 | 9 | 8 | 9 | 6 | 5 | 7 | **78** |
37 +| G03 weight-geometry → quant sensitivity | 9 | 6 | 9 | 9 | 8 | 8 | 8 | 7 | 6 | 6 | **76** |
38 +| G12 correlational→causal survival rate | 9 | 10 | 8 | 7 | 8 | 8 | 9 | 6 | 4 | 7 | **76** |
39 +| G13 ablation-scheme curves standard | 6 | 8 | 6 | 9 | 9 | 9 | 8 | 8 | 3 | 9 | **75** |
40 +| G07 base↔instruct delta atlas | 6 | 4 | 7 | 10 | 10 | 8 | 8 | 9 | 4 | 8 | **74** |
41 +| G24 working-set maps (localvm bridge) | 9 | 8 | 9 | 7 | 7 | 6 | 8 | 5 | 10 | 5 | **74** |
42 +| G02 circuit survival under quantization | 8 | 8 | 9 | 7 | 8 | 8 | 8 | 6 | 4 | 7 | **73** |
43 +| G09 cheap maps → pruning tolerance | 7 | 7 | 7 | 8 | 8 | 7 | 7 | 7 | 9 | 6 | **73** |
44 +| G14 landmark-anchored alignment | 8 | 8 | 8 | 8 | 8 | 9 | 8 | 6 | 4 | 6 | **73** |
45 +| G21 shuffle-regime SAE replication | 8 | 8 | 9 | 7 | 9 | 7 | 7 | 6 | 5 | 7 | **73** |
46 +| G05 quantized-runtime capture (llama.cpp) | 7 | 7 | 9 | 7 | 8 | 9 | 7 | 4 | 7 | 6 | **71** |
47 +| G16 cross-scale emergence maps | 8 | 7 | 7 | 8 | 8 | 8 | 8 | 6 | 4 | 6 | **70** |
48 +| G22 MLX capture layer | 5 | 6 | 7 | 8 | 8 | 8 | 6 | 5 | 8 | 7 | **68** |
49 +| G18 FR/EN localization atlas | 7 | 8 | 6 | 8 | 8 | 6 | 8 | 6 | 3 | 7 | **67** |
50 +| G19 editing as perturbation probe | 8 | 8 | 8 | 6 | 7 | 7 | 7 | 5 | 4 | 6 | **66** |
51 +| G04 tuned-lens drift metric | 6 | 5 | 7 | 9 | 7 | 7 | 6 | 8 | 3 | 5 | **63** |
52 +| G15 Rosetta-mining on LLM pairs | 7 | 6 | 7 | 7 | 7 | 8 | 7 | 6 | 3 | 5 | **63** |
53 +
54 +## Reasoning (grouped; scores that drive the ranking)
55 +
56 +**G23, G20, G10, G13 — the infrastructure quartet.** These top the table not
57 +because they are the boldest science but because every other row *presupposes*
58 +them: G23 (map cards) is the charter's core deliverable and verified absent in
59 +the field (NV 9, AV 10); G20 is expH — nothing else's feasibility column is
60 +*measured* until it runs, and it reuses localvm's harness (SY 10, RK 10); G10 is
61 +expA/expD — the denominator of every future claim (RP 10); G13 is a cheap
62 +standard that upgrades everything else's epistemics. **Decision: these are not
63 +"candidates" to bet on — they are the mandatory substrate, scheduled first
64 +regardless of ranking.** Their risk is near-zero and their outputs are atlas
65 +artifacts in their own right (cost-frontier table, noise-floor table, schema).
66 +
67 +**G17 (83) — the strongest research candidate.** Q1 needs a *number*, not an
68 +adjective; the index is Tracr-falsifiable before any real model (ES 9), cheap
69 +(ablation curves, FA 9), general across models and granularities (GN 9), and
70 +its outputs are exactly the comparable, versionable artifacts the charter wants
71 +(AV 9). Main deduction: hydra-effect contamination requires resample + joint
72 +ablations (IC 7).
73 +
74 +**G06 (81) — highest scientific value in the table (SV 10).** If weight-only
75 +statistics predict activation-level maps, the accuracy/cost frontier of local
76 +interpretability moves for everyone; if they don't, a harsh pre-registered
77 +negative (R²<0.2) is still a Level-1 publication. RK 5 is the honest price of a
78 +real hypothesis. Absorbs G03/G08/G09 as sub-studies (same regression harness,
79 +different targets) — G08's causal verifiability (ES 9) makes it the anchor
80 +sub-study to run first.
81 +
82 +**G01 (79) + G02 (73) — the quantization cluster.** G01 is the entry point:
83 +highest feasibility of the cluster (FA 9), clean controls, and the atlas story
84 +("the same model, mapped at four precisions") is uniquely ours (AV 9). G02
85 +follows once G01's noise floors exist; its FA 7 reflects the
86 +gradients-through-quantized-weights problem. G04 (63) and G05 (71) stay in the
87 +cluster's backlog: G04 until lens seed-variance is measured, G05 because IC 4
88 +(C-level engineering) makes it a tooling milestone rather than a first bet.
89 +
90 +**G11 (78) + G12 (76) — the verification science.** Both are epistemically
91 +top-shelf (ES 9–10) and both are largely *byproducts* of running G17+G10
92 +honestly: agreement rates and survival rates fall out of the same runs.
93 +Scheduled as measurement layers on candidate experiments rather than
94 +standalone prototypes — this is why their totals do not need to beat G17's.
95 +
96 +**G24 (74) — the localvm bridge.** SV 9 / SY 10 / NV 9, but RK 5 (working sets
97 +may not be domain-stable) and IC 5 (needs localvm's engine as consumer). Kept
98 +as the fourth candidate at reduced scope: run the falsifier (domain-conditioned
99 +block-ablation on a 3B model) before committing to the full bridge.
100 +
101 +**G07 (74) and G08 (80) — high totals, bounded ambition.** Both are cheap,
102 +reproducible atlas layers (RP 10) rather than research programs; they are
103 +folded into candidate work as first map deliverables (G08 → candidate 1's
104 +anchor; G07 → an early standalone atlas product to exercise the map-card
105 +pipeline end-to-end).
106 +
107 +**Deferred (G14 73, G21 73, G16 70, G05 71, G22 68, G18 67, G19 66, G15 63,
108 +G04 63).** Cross-model coordinates (G14/G15/G16) are blocked on G10's noise
109 +floors by design — alignment claims without floors are exactly the field's
110 +failure mode. G21 waits for a working SAE recipe (itself gated by G20's
111 +throughput numbers). G22 is built incrementally as experiments demand it, not
112 +as a big-bang library. G18 is a strong *application* of G17 once the index
113 +exists. G19 is deferred on port cost (CUDA-first editing stacks). Nothing here
114 +is rejected; the ranking sequences, it does not kill.
115 +
116 +## Selected prototype candidates
117 +
118 +| Slot | Program | Core gaps | First falsifier |
119 +|---|---|---|---|
120 +| **candidate_01** | **Weight-only pre-screen** — weight statistics vs activation-level ground truth | G06 (+G08 anchor, G03/G09 extensions) | super-weight causal catalog on 2 models; then Pythia-410M regression, kill at CV R²<0.2 |
121 +| **candidate_02** | **Quantization deformation atlas** — the same model mapped at FP16/Q8/Q4(/Q2) | G01 (+G02, then G04/G05) | Qwen3-0.6B FP16 vs Q4 probe drift vs noise floor, 10 properties × 5 seeds × 2 datasets |
122 +| **candidate_03** | **Localization index** — a Tracr-validated concentration measure for Q1 | G17 (+G11/G12 as measurement layers) | index must read ≈1 on Tracr, ≈0 on random-init; else broken before real models |
123 +| **candidate_04** (reduced scope) | **Working-set bridge to localvm** | G24 (+G09) | domain-conditioned block-ablation on a 3B model; <10% working-set reduction kills per-domain prefetch |
124 +
125 +**Mandatory substrate scheduled before/alongside candidates:** G20 (expH),
126 +G10 (expA/expD), G23 (map cards v0 co-evolved with expA's first output),
127 +G13 (adopted as schema rule, zero research cost).
128 +
129 +**Sequencing rationale.** The three main candidates share the substrate but
130 +fail independently (weights-regression, drift-vs-floor, index-validation are
131 +different kernels), satisfying the charter's demand for 3–5 genuinely distinct
132 +bets. Every candidate produces versioned map artifacts even in its negative
133 +outcome — the property the charter's §7 explicitly prefers.
134