spb/modelmap Public License

Only 1/4 fresh sources monotone (positive dose arm unstable); specificity
failed to replicate (random |delta| 3.36 vs bound 1.14 — run #5's 0.45 was
3-draw sampling luck). Halving held 4/4: necessary but not a reliable
additive handle. interventions/v1 (Level 2) stands as the final causal
claim. Gate record: refuse -> pass(L2) -> refuse -> refuse. Methodology
rule adopted: specificity controls need >=10 random draws + percentile
bound. Charts: x-axis customization for dose-response figures.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Conjunctive dose-response criterion failed: L12 passes everything
(strictly monotone +1.95..+6.79 across -2s..+2s, halved, specific 0.45
vs bound 1.11) but L04 is an overdose regime (any 2-sigma perturbation,
random included, collapses the margin) and L08 is not specific at 2s.
make_l3_mapcard.py refused v2; interventions/v1 stays Level 2. Gate
record: refuse (r3) -> pass (r4, L2) -> refuse (r5). Run #6 registered
ideas: layer-local dose calibration; minimal L12 single-layer L3 claim.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Early-band (2-15) mean specific damage +3.23..+3.35 across six fresh
direction sources (min +3.228 vs pre-registered bar 2.5; spread 0.12) on
a fresh behavioral bank (baseline +4.45). Late band unstable, no claim.
Published atlas/qwen3-0.6b-4bit/interventions/v1 (Level 2): a single
diff-of-means agreement direction, erased at any early-band layer,
removes ~73-75% of grammatical preference. The gate that refused run #3's
per-layer map passed the band version. Band map leads the home page.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Replication failed: mean pairwise rho 0.495 (<0.7), band claim fails on
2/5 direction sources; make_interventions_mapcard.py exits 1 and the
atlas stays at two entries. The refusal decomposes the object: early band
(2-15) replicates tightly (+2.98..+3.34 of +4.24), late band (20-27) is
estimator noise (-0.44..+2.77) -- retroactively explaining run #2's
anti-correlation. Run #4 pre-registers the narrower BAND claim.
Charts: interventions profile builder + dynamic y-axis (renders when a
map passes the gate).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

P1 CONFIRMED: bottom-5 skip damage was general (specificity 0.51 < random
mean 1.44), top-5 above mean but below p95 — skip never singles out the
probe layers. P2 FALSIFIED: direction-erasure profile anti-correlates with
the probe profile (rho=-0.136, p=0.76). Discovery: erasing the agreement
diff-of-means direction at ANY layer 2-15 destroys the behavior (up to
+3.97/+4.24 at L12, random-direction controls netted); late probe-ranked
layers carry little; L18/L22 suppressive. The load-bearing object is an
early-constructed DIRECTION, not a late place. Tap gains an edit hook.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Layer-skip ablation (Tap.skip): top-5 differential layers damage +2.17,
BELOW random-5 mean +3.14 (p95 +4.85); bottom-5 (layers 0-4) largest at
+4.99. Decodability != causal support (Hase-class dissociation, in-house).
Atlas probes/v2 records the failed check (interventions + confidence.md);
survival-rate ledger opened in expC analysis.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- v2 promptsets: structure-borne, token-balanced (overlap certificates);
  capture_pooled returns mean+last-token reps in one pass
- Strict twin gate falsified again (word_order twin acc 0.96 — subword
  statistics); gate retired, differential (real-twin) is the map
- agreement: 25/28 signal layers (max dSel +0.379); arith_valid: real
  0.86/0.90 vs twin 0.56, strongest last-token; word_order null-dominated
- First real noise floor: top-5 replication ~0.54 off ceiling; dataset
  shift > seed SD on only 3-8/28 layers (registered 1/3 not met)
- atlas probes/v2 published (Level 1, per-property verdicts); site charts
  generalized to any probes entry; home leads with agreement map

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- src/modelmap/capture: MLX Tap layer + random-init architecture twin
- promptsets v1 (6x240, sha256 manifest) — now the positive-control corpus
- Result: ceiling everywhere (acc 1.000); twin ALSO at 1.00 on all 28
  layers (max sel 0.56-0.88 vs registered 0.05 gate) — probe map reads
  tokenizer+architecture prior, zero trained-model signal. Shuffled-label
  control would not have caught it; only the twin null did.
- Published atlas/qwen3-0.6b-4bit/probes/v1 (negative_result=true,
  Level 1, replication 1.00) via publish.py gate
- Doctrine: probe maps only as REAL-TWIN differentials from now on

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Run #2 (cold cache, M3U96a, purge/repeat): FALSIFIED — warm ordering
inverts; zarr-uncompressed 0.62 GB/s beats raw-mmap 0.14 on cold random
batches (page-fault QD1 IO vs 32MiB chunk reads). Rule: IO granularity
decides, not the container. Store design revised; run #4 registered.

Run #3 (mlx-lm Qwen3-0.6B-4bit): CONFIRMED — retain 1.004x plain prefill
(capture free under lazy eval), retain+write 1.28x. First quantized-model
activation capture in Python tooling.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- mapcard v0 schema + publish gate (G23), stats core (bootstrap/FDR/
  replication), probe harness with built-in shuffled-label controls
  (unit test caught a fancy-indexing shuffle bug pre-science)
- expH hypothesis registered before run; run #1 on M5 Max 48GB:
  (A) mmap 3.2-10.8x zarr on random-batch reads (confirmed >=2x bar);
  (B) MLX capture nearly free (1.02x), torch-MPS retain 1.22x (confirmed)
- expA hypothesis registered (dataset-variance > seed-variance)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

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>

Seven clusters: quantization×internals, weight-only pre-screens,
replication/method-agreement, cross-model coordinates, localization
science, systems/tooling, atlas methodology + localvm bridge. Each gap
carries the six-field block with a smallest falsifying Mac experiment.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

~40 techniques, six families + cross-cutting instruments, eleven-field
blocks with epistemic flags; overlap analysis, already-tried combinations,
novelty traps, estimated Mac cost frontier (16/32/64 GB).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- research/notes/: 10 theme notes with verified sources, Apple Silicon
  status, failure modes, epistemic status per technique
- research/bibliography.md: ~200 sources with URLs + access dates
- research/LOG.md: sweep results — quantization gap confirmed (~5 shallow
  papers), weight-only pre-screen gap open, noise floors (30% SAE seed
  overlap, 1-5% neuron universality), atlas provenance gap verified

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- Repo skeleton per charter §3 with mandatory author headers everywhere
- tools/check_headers.py (gates commits), new_experiment.py, new_map.py
- benchmarks: hardware_manifest.py (sysctl-based) + harness skeleton
- experiments/micro A-H scaffolded with empty seven-field blocks
- site/: modelmap.io platform (Express, light theme, English), coherent
  with localvm-research/web — Home, Research, Experiments, Atlas (with
  confidence Levels 0-3), Results, Code, About; attribution footer

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>