jspace/paper-B/PLAN.md

Paper B — Where Retrofitted Recurrence Attaches: Injection Cliffs, Read-Invariance, and a KV-Sharing Hazard

One-sentence claim. For loops retrofitted onto a frozen transformer, the write location is everything and the read location is nothing: performance falls off a one-layer cliff at the sensor→workspace boundary, is invariant to where the fed-back state is tapped, and — on KV-sharing architectures — becomes structurally impossible above the sharing boundary, a hazard that silently invalidates hook-based interventions far beyond this project.

Audience/venue. Interpretability/analysis venue (e.g. BlackboxNLP-class workshop, or an interp track); alternatively TMLR. Cites Paper A for training recipe, Paper D for the lens.

Claims and evidence manifest

#	claim	evidence	status
B1	Anchor cliff: inject L14→52%, L13→34%, L12→29%, L11→25% (monotone, one-layer onset at lens boundary)	results-band-{13_30,12_30,11_30}/eval*.json + reference	✅
B2	Tap invariance: taps 23/27/30/32/34 statistically flat (incl. motor/final layer); 40%-shorter loop body matches	results-tap*/eval*.json (note: 27/32/34 on H200, 23 on Spark, k0 anchors align)	✅
B3	KV-sharing hazard: E2B layers 15–34 reuse K/V from ≤14; band re-runs entered ≥15 are bit-exact nulls in training (zero grad) and behavioral nulls in eval (Δlogit up to ~11 with 0 argmax flips)	arm 17–27/24–34 logs; forensic probes (transcribe from session); results-node-final/results-kvtest/train.log	✅
B4	Causal demonstration: moving entrance one layer (14→15) reproduces the null on demand	results-kvtest/train.log (flat val@ all k)	✅ (eval tail missing — rerun 5-min eval for completeness)
B5	12B (no KV sharing) clean location test: wrong band 0.4–2.3% vs 32.8%	results-node-final/results-12b/eval_12b_gsm_band10*.json	✅
B6	Fixed-point dynamics: big first step, cos→1.000 by k≈4; explains accuracy/sharpening plateau and k-extrapolation flatness	dynamics data + eval_code_code_kgen.json	✅
B7	Discriminator: cliff = lens boundary vs last-full-attention-KV layer	MISSING — the (9,30) arm	❌ queued idea
B8	Residual-only perturbation impotence: large logit shifts w/o attention pathway don't flip decisions	probe from session (rerun scripted for reproducibility)	◐ script it

The hazard section (the paper's service to the field)

Write as a standalone, checkable warning: (1) detection recipe — check num_kv_shared_layers/equivalents before any layer-局部 intervention; (2) symptom catalogue — exact-baseline results, zero gradients, big-Δlogit/no-flip; (3) how our 3 pre-registered arms became silent nulls and how forensics found it (file diff → shared_kv_states kwarg → config). Affected model classes: KV-sharing efficiency variants (survey which public families have it).

Missing experiments (all cheap, Spark)

1. B7 (decisive): anchor (9,30) — L9 is full-attention+KV-computing but deep in sensor region. Pre-registration drafted in results-loop/PROTOCOL_UNIFIED.md §(to add). Prediction on record: lens boundary wins (arm 1 contained L9 and was catastrophic). ~70 min.
2. B4 completeness eval (5 min).
3. B8 as reproducible script + N=16 items (30 min).
4. Optional: channel-count sweep (anchor 12 with tap 28 length-match) if B7 surprises. Decide after B7.

Outline (workshop-length)

1. Setup: retrofit loop machinery (1 fig), what "write" and "read" are.
2. Read-invariance (tap curve fig).
3. Write-cliff (anchor curve fig + 12B clean replication).
4. The KV-sharing hazard (anatomy fig: which layers recompute K/V; symptom table; detection recipe).
5. Fixed-point account tying 2–4 together.
6. Implications for activation-intervention methodology.

Work plan

• Run B7 + B4-eval + B8-script (one Spark session, ~2h GPU total).
• Transcribe session forensics into paper-B/hazard_notes.md while fresh.
• Figures: anchor curve, tap curve, KV anatomy diagram.
• Draft after Paper A ships (it cites A's training recipe).