Status Report — August 2026

DRAM-Native Bit-Logic Classification — 2026-08-09

August is the month the tooling caught up with the algorithm. One root cause — the float32 score accumulator drifts on large sums — explains a whole cluster of earlier anomalies (merge member counts, float-vs-int32 gaps, "the float tree is exhausted"). Fixing it (exact 64-bit accumulation, SCORE_TYPE default → double) made the merge deterministic, the archive self-describing, and the IFC — the delivery tool — finally able to run a heterogeneous ensemble end to end: 66.36% == 66.36% == 66.36% (merge == retrain == IFC import).

​ 1. The float32 Accumulation Drift — Precision is the Feature

The ensemble scores reach ~1e9 — far beyond float32's exact range (2²⁴ ≈ 16.7M). Each addition rounds, the error grows with every member, and the beam's gain/argmax evaluations become noisy. The same corpus searched with different merge accumulators tells the story:

Merge SCORE_TYPE	members	eval
float32	29	66.79%
int64 (-int32)	44	67.72%
double	44	67.72% (= int64 exactly)

The "32-bit is free / only top bits matter" assumption is refuted. The bit-exact algorithm is sensitive to score-accumulation precision — not to the container width. The DRAM chip accumulates popcounts exactly in integer arithmetic, so the int64/double behavior is chip-faithful; the float32 drift was an emulation artifact. SCORE_TYPE now defaults to double (-DSCORE_TYPE=float reproduces the legacy behavior).

Why this explains earlier puzzles

• The "float tree is exhausted" (28-33 members vs 88-116 historically): the historical results were made with the int32 merge (pre-float-default); the float merge drifted.
• --tries seemed necessary under drift (noisy gains → false attractors); with exact accumulation the single-try --max beam is deterministic.
• The float-vs-int32 gap (68.15 vs 69.61 on int32-selected members) was a selection + drift artifact, not an int32 advantage.

​ 2. Archive Format v12/v13 — Export Follows the Internal Format

The .ens archive used to store float32 (v11) even when the computation ran in double — losing precision between computation and storage. Per decision, the export now follows the internal format:

version	score type	bytes	note
v8	int32 (clamp)	4	legacy, read-only
v9-v11	float	4	unchanged
v12	double	8	double builds store losslessly
v13	int64	8	int64 mode, no int32 clamp

• The version encodes the score type; the export picks it from SCORE_TYPE.
• The .meta now records COUNTER_TYPE=/SCORE_TYPE=; the trainer errors and the merge warns+skips on mixed-type corpora — a float32 archive can no longer silently mix into a double corpus.
• A missed version bound in one reader path caused the "0 blocks filled" SEGV — fixed in all seven sites.

​ 3. lib/ki-ens.h — the .ens Format is Now One Module

All version logic was scattered across four duplicated read paths. It is now centralized in lib/ki-ens.h:

• Single version → {metadata layout, score width, type} mapping (ens_score_bytes, ens_score_type_str, ens_has_*, ens_version_for_score_type).
• EnsReader replaces the 4 duplicated read paths in merge-ensemble (read_ens_meta, load_scores_directory, --check, archive_*) — net −420 lines.
• ens_write + ens_verify (roundtrip, always-on): the trainer verifies every exported member at write time — a format regression fails on the first member, not after a full corpus.

​ 4. --max = Beam-as-a-Whole — Deterministic with Exact Accumulation

--max (alone) now activates the old tolerant formalism (stop only after beam_width consecutive dry steps) instead of zeroing --min-gain; --union was removed. With exact accumulation the single-try --max beam is deterministic:

corpus	single-try	--tries 10	Δ
M1-INT32	66.36% (51)	66.36% (51)	0
M3-INT32	65.25% (88)	65.25% (81)	0
M1-FLT32	66.88% (92)	66.88% (92)	0
M3-FLT32	65.44% (55)	65.44% (55)	0

​ 5. IFC v7 — the Delivery Tool Handles Heterogeneous Ensembles

model.otto v6 stored only W0/target/offset — it could not represent a heterogeneous ensemble (CIFAR import gave 12.81% = garbage; MNIST worked only because it is single-config). v7 adds a per-member section (channel, encoding, width, xform strings) after the header. A second bug was the missing enc-LUT initialization in the import (gave exactly 10.0% = random; enc_lut_get returned garbage). Both fixed:

--debug-member on the import prints per-member evals for verification ([ 3/51] ens=42.1% mem=32.4% G:sig8 xf=id).

​ 6. --no-ens-cache — Real-Live Inference on New Data

The IFC's purpose is inference on new data, so the file cache (data/prepped/cex_*.pre) is a research/dev feature. --no-ens-cache skips the file cache (read AND write) — inputs are computed from raw data every run; the in-memory xform cache stays (it builds from the run's data). Validated:

run	eval	time
--import member.exp (cache)	66.36%	5690ms
--import member.exp --no-ens-cache (real-live)	66.36%	6504ms

Only +814ms — the h0 scoring dominates, so the real-live performance IS the performance.

​ 7. The 4-Way Technology Comparison (xform=id, 377 members)

Four corpora built identically (H=256, EP=10, seed 42, uniform W0, .meta type-bound), selection via --beam 10 --max:

corpus	cumulative (best-en)	beam-selected
M1-INT32 (maj1, int32)	62.63%	66.36% (51)
M3-INT32 (maj3, int32)	61.94%	65.25% (88)
M1-FLT32 (maj1, float+double)	62.70%	66.88% (92)
M3-FLT32 (maj3, float+double)	62.02%	65.44% (55)

• maj1 > maj3 consistently (+0.68-0.69pp cumulative, +1.11-1.44pp beam-selected) — the beam exploits maj1's better members more effectively.
• flt32-double ≈ int32 (±0.07pp cumulative; flt32 slightly ahead when beam-selected +0.19-0.52pp) — with exact 64-bit accumulation the container format is neutral: the Container Principle holds once the drift is gone.
• Winner: M1-FLT32 = 66.88% (maj1 + float counters + double accumulation).

​ 8. Member Lists are Setup- & Arithmetic-Bound

There is no universal member list. A selection is only valid for the exact setup it was made under (maj, H, EP, VN/HN, seed, W0) and the arithmetic it was scored with:

selection → retrain	eval	Δ
float-selected → float	67.85%	—
float-selected → int32	67.12%	−0.73pp
int32-selected → int32	69.61%	—
int32-selected → float	68.15%	−1.46pp

Each arithmetic performs best on its own selection; the coupling is asymmetric (int32-selected members hurt float more than the reverse). Always retrain with the corpus' exact flags — the .meta type binding now enforces the arithmetic side.

​ 9. Fashion-MNIST: The 91% Trap

→ Fashion-MNIST paper — a second, independent dataset that sharpened the August findings. The DRAM-native bit-logic stack (Otto Score + maj1, frozen random W0) reaches a 91.6% ceiling on Fashion-MNIST — and cannot break out of the 88.7–91.6% basin, no matter the search strategy:

Strategy	eval range	Finding
--beam 10 --max, H32..512, m1t 98..108	90.8–91.6%	H is flat (+0.14pp from H196→H512); the ceiling is architectural
--tries N --tries-no-lock	91.3–91.6%	full-pool search; MORE members ≠ better (pool-monotonicity disproven)
--tries 3 + --beam-bestN 1/2/3	91.4–91.5%	breaking the 1st-best fixation moves the attractor by ~0.1pp
--expansion-sort clarity	86.8–89.2%	member clarity alone is WORSE — diversity beats strength
Bit-Voting (no W0, no maj1)	88.7–89.0%	lower technology bound; the W0+maj1 advantage is maj1-driven

The 3.4pp gap to CNN SOTA (94–96%) is architectural — frozen random W0, no convolution, no trainable features. Every search lever moves the needle by tenths of a percent only. What speaks for the approach: speed — a 221-member sweep runs in ~1 minute (PRF trainer), orders of magnitude faster than CNN training. For a DRAM-native chip the question is not "can we reach 95%" but "how little hardware do we need for 91.6%".

Two real bugs surfaced along the way (both fixed): the IFC import started with CLI-default maj1-thresh (107) while the model was trained with 104 → every h0 bit differed → 36.6% instead of 91.1%. The model.otto format is now v8 (self-describing): it carries maj_mode, maj1-thresh, splitVN and splitHN, so --import needs no CLI flags. The --dry-run import also SEGV'd on the unloaded pixel buffer — now a clean "setup only" exit. See the Fashion-MNIST paper for the full leaderboard and the end-to-end 91.10% = 91.10% = 91.10% (merge = retrain = IFC) verification.

​ Summary

topic	result
float32 score accumulation	drifts at ~1e9 → refuted "32-bit is free"; SCORE_TYPE default → double
.ens archive	v12 (double) / v13 (int64); export follows the internal format
format code	centralized in lib/ki-ens.h; EnsReader + ens_write + ens_verify (−420 lines)
beam scan	--max = beam-as-a-whole, deterministic; --tries adds nothing; --union removed
IFC (delivery)	model.otto v7 (per-member metadata) + enc-LUT fix → heterogeneous 66.36% end-to-end
IFC self-describing	v8 carries maj_mode/maj1-thresh/splitVN/splitHN → import needs no CLI flags (Fashion 36.6% bug fixed)
real-live IFC	--no-ens-cache: compute from raw data, +814ms
technology comparison	maj1 > maj3; flt32-double ≈ int32; winner M1-FLT32 66.88%
Fashion-MNIST	91.6% ceiling / 88.7–91.6% basin — architectural, not search-limited; 91.10% end-to-end (merge = retrain = IFC), see Fashion paper