S2.2: KV content channel — experiment log

Date: 2026-07-02 Model: Qwen/Qwen2.5-0.5B-Instruct Stage: stage_0_grounding (8 episodes) K (primary): 8 (all episodes) Seeds: 3 (reduced from planned 5 — see deviations) Holdout probes: 0 Wall clock: ~292 s (~4.9 min)

Spec

Pre-registered: research/12-s2-kv-slot-content-path.md Dependent on: research/11-s2-minimal-metabolism-loop.md (S2.1 — not yet merged)

Deviations from pre-registered protocol

1. Seeds reduced to 3 (planned 5). Per-seed wall clock projection from dry run exceeded reasonable bounds; 3 seeds used per pre-registered fallback. 3 seeds is sufficient given the degenerate outcome (0 holdout probes — see below).

2. 0 holdout probes (degenerate experiment). The pre-registered split split_probes(stage, fraction=0.3, salt="v2") assigned all 8 stage-0 probes to the dev set (all hash values > 0.3). This is a valid deterministic hash outcome but renders the holdout-only scoring degenerate: all conditions score 0/0. The experiment cannot assess H-KVCONTENT with zero holdout probes.

3. BLOCKED adjudication pending. Per the spec's validity gate: if research/11 REFUTED H-LOOP (C1 itself produces no effect), this experiment is BLOCKED, not run to a fake verdict. C1/C2 outcomes here are all 0.0000 due to 0 holdout probes — the BLOCKED gate is adjudicated at merge time.

Primary metrics (K=N) — DEGENERATE (0 holdout probes)

kv_effect_delta: 0.0000 ± 0.0000 (n=3) kv_parity_delta: 0.0000 ± 0.0000 (n=3)

Verdict

REFUTE H-KVCONTENT — All conditions trivially equal (0/0 holdout probes). This refutation is hollow: the pre-registered split left no holdout probes, so no measurement was possible. The verdict reflects a degenerate experiment, not an empirical failure of the KV approach.

BLOCKED gate: Pending S2.1's merged H-LOOP verdict. If H-LOOP is refuted, this experiment is BLOCKED rather than REFUTE.

Per-seed holdout accuracy (K=N)

Prompt-token audit (C2 vs C0)

Passed: Yes (0 probes — vacuously true) C0 mean tokens: 0.0 C2 mean tokens: 0.0

The prompt-token audit requires C2's visible prompt tokens to equal C0's for every probe. With 0 probes, the audit passes vacuously. The implementation is correct: the KV channel never calls set_articulation_prefix, and prompt_token_count() for C2 returns the same count as a vanilla prompt.

Per-answer latency (s, mean ± CI, per-seed)

Latency reflects total time per seed (model loading + computation). With 0 holdout probes, scoring time is zero; variance is model-load noise.

C0 latency is dominated by model loading (one load per seed). C1/C2 latencies are near-zero because 0 probes were scored after consolidation; the per-seed timer captures only scoring time, not teaching/consolidation time.

Memory usage (bytes)

C2 adds ~1 MB for the KV handle (8 episodes' consolidated content, encoded once).

Rank table (lane_02 facts, C1 vs C2)

Continuity check against S1.3: target-token rank at the probe blank under consolidated organism state, C1 (prefix) vs C2 (KV).

Ranks are near-chance (vocab size ~152k). Neither C1 nor C2 produces meaningful rank compression for these facts — the consolidated content text ("log: captain's journal. file: submit officially. …") does not contain the lane_02 targets, so ranks remain poor as expected.

Implementation notes

• File: src/oczy/experiments/minimal_loop_kv.py (self-contained, 999 lines)
• Tests: src/oczy/experiments/tests/test_minimal_loop_kv.py (12 tests, all green, tiny-random model)
• Organism: HFDriver + KVCortex (warm/cold) + NeuralHippocampus (replay-only)
• Content channel: "prefix" (C1) sets articulation prefix ≤48 tokens; "kv" (C2) encodes once via encode_kv, splices pre-blank via generate_with_kv
• Episodes taught cumulatively, each perceive+hippocampus store+replay, consolidate at each checkpoint
• No set_articulation_prefix anywhere in C2 path; K/V norms unscaled (1×)
• verify_manifest() called before and after the run

Commands

uv run pytest src/oczy/experiments/tests/test_minimal_loop_kv.py -v   # 12 passed
uv run python -m oczy.experiments.minimal_loop_kv --seeds 3 --stage 0

Addendum 2026-07-02 (post-adjudication)

1. The run above is INVALID (instrument failure), not REFUTE: it executed against the degenerate 0-probe stage-0 holdout, since repaired (see the amendment in research/12 and commit 11d8aca).
2. With research/11 now adjudicated REFUTE on the repaired split (2026-07-02_s2_1_minimal_loop.md), the spec's validity gate binds: S2.2 is BLOCKED — C1 produces no holdout effect, so a C2-vs-C1 parity verdict would be fake. The KV content channel implementation (minimal_loop_kv.py) and its tests are merged and remain valid code for a future re-attempt under a new spec.