# Mode B scope (claim boundary)

**Mode B (implemented in this repo):** TEDDY + our head as a **typed evidence source** → ANM **finite_field** decisions on CITE holdout (site4/test). Editable observer YAML (O0/O1/O2); LOO/flip attribution on **decisions**. With all markers entering at once, ANM's calls equal the fixed rule's for the same question; the claim is the declared, auditable layer.

**Scope in one sentence.** teddy_mm runs ANM's open-loop decision layer: TEDDY is a prescribed evidence source and does not read the field, so ANM's dynamics (state feedback, operator memory, delays) are not tested here.

This is **not** Mode A. Do not read current results as residual-stream / Jacobian / in-silico gene-perturb claims.

> **Registered v3 (1 Oct 2026) supersedes the notes below where they conflict** ([V3_REGISTERED_RESULTS](reports/V3_REGISTERED_RESULTS.md)). In the registered tests ANM's engine made exactly a declared rule's call on every test cell, on site4 and on external data (E1, E6), and none of its readouts added decision value. E3 rejected "the NK–T loss is in the readout or pooling and is repairable on frozen TEDDY". Mode A now has registered results: E5 (TEDDY's layers) stopped at its derivative check, and E5-M (sufficiency of the gene-mean for one token observer) was inconclusive. Experiment 2 was redesigned as v3 E2 and E4, Experiment 6 as E3, E5 and E5-M.
>
> **中文：** v3 预注册结果（2026-10-01）在冲突处取代下文：ANM 引擎在 site4 与外部数据上都与声明规则逐细胞判定相同，其读出未带来决策价值；E3 拒绝“可在冻结 TEDDY 上修复”；Mode A 方面 E5 停在导数检验，E5-M 对所测的一个观察者判定为不确定。

> **Corrections (29 Sep 2026): rerun done for Experiments 1, 3, 4 and 5.** See the notice at the top of the [README](../README.md#corrections-29-sep-2026-rerun-done-for-experiments-1-3-4-and-5) and the [site](https://danielchen26.github.io/teddy_mm/#corrections). Reports (official preprocessing, the numbers below): [`HARD_PROOF_OFFICIAL`](reports/HARD_PROOF_OFFICIAL.md), [`SCOPE_REFINE_PROOF_OFFICIAL`](reports/SCOPE_REFINE_PROOF_OFFICIAL.md), from `outputs/anm_cite_bridge_official/`; earlier run on the earlier preprocessing: [`HARD_PROOF_V2`](reports/HARD_PROOF_V2.md), [`SCOPE_REFINE_PROOF_V2`](reports/SCOPE_REFINE_PROOF_V2.md).
>
> - All markers now enter the engine at once (the first run entered them one per step in panel order, a hidden weight), and predictions use one train-median size factor, 0.92663, instead of a per-cell factor computed from measured protein. The answer key, the fixed rule and ANM share one scoring function.
> - With simultaneous markers ANM's lineage scores are a fixed multiple of the fixed rule's, so ANM makes the same call as TEDDY + fixed rule (re-coded per question) on every site4 cell for O0/O1/O2, by construction: "no calls" 272 / 1,493 / 2,551, accuracy of calls made 0.9429 / 0.9667 / 0.9545, exactness 0.9336 / 0.9312 / 0.8559. ANM's value in Mode B is a declared, auditable layer (written question, P_f/Q_f checks, LOO reasons), not better calls.
> - The O0 and O2 rule shifts (101 → 272 and 1,796 → 2,551 "no calls") come from the export's size factor and the official TEDDY-G preprocessing, not the scoring. O1 is redefined (weighted mean, key weight 2): 881 → 1,277 from the scoring, 1,277 → 1,493 from the size factor and the preprocessing. The first run's ANM accuracy edge (317 / 834 / 3,373 "no calls"; 0.9496 / 0.9723 / 0.9603) came only from declining more.
> - Experiment 3: the most frequent top marker is now CD2 (17.9%), CD5 close behind (23.6% → 17.3%); under simultaneous timing the top marker is the protein with the largest normalised value in the called lineage, so it is not ANM-specific. Flip rate 0.2999; within-lineage permutation p = 0.0099 for CD2's share (15.7% shuffled; the floor for 100 shuffles).
> - Experiment 4: soft_P never gates better than the fixed rule's margin at matched coverage (every difference ≤ 0; O2 at 40%: 0.9988 vs 0.9988); soft_P reaches 1.11, a score, not a probability. Hard cells are mostly out of scope (NK/ILC 3.1×, erythroid 3.4× enriched).
> - Experiment 5, on a disjoint split (11,725 / 5,025): labels to match ANM are logistic 500, MLP 1,000 (was 200), grid 50.
> - The answer key still has no NK or out-of-scope class; out-of-scope cells still get a call 73–95% of the time. Experiments 2 and 6 stay withdrawn pending redesign (prototypes: with degraded RNA, ANM with a trust per level = the fixed rule with a stricter bar per level; the TEDDY + head predictions do separate NK from T in the right direction but compress the differences to about a quarter to a third of the measured ones (CD335 0.88): 1,053 of 1,506 site4 NK–T pairs involve "gdT CD158b+" cells, which look NK-like in both predicted and measured protein, and on the other 453 pairs the median predicted within-pair gap for CD56 / CD94 / CD335 / CD3 is 0.148 / 0.100 / 0.364 / 0.205 vs measured 0.621 / 0.321 / 0.413 / 0.756; Mode A probes on the frozen embedding: a linear NK-vs-T probe separates NK from T at every TEDDY layer (held-out AUC 0.991–0.997); on these pairs a linear protein probe on the final embedding keeps about as much of the gap as our head (CD56 / CD94 / CD335 / CD3 0.28 / 0.40 / 0.40 / 0.28 vs 0.24 / 0.31 / 0.88 / 0.27), an MLP probe on the same embedding about twice as much (0.47 / 0.59 / 0.74 / 0.61), and 13 NK/T gene tokens more than the gene-mean, already at the input layer, which v2 read as part of the loss being in our head's readout and in mean pooling, part perhaps not in the detected RNA (v3 E3 rejected the repair reading; E5-M was inconclusive on pooling for the one observer tested); one site, probes are not the head's objective, pairs chosen for closeness in the embedding). "0 labels for a new question" holds for any written rule, including the fixed rule.
> - Phase 1 re-evaluated with the train-median factor, its head retrained on the official TEDDY-G embedding: mean test Pearson 0.603 (MLP), 0.582 (flow matching); was 0.610 / 0.596 (earlier preprocessing, per-cell factor). The earlier non-official preprocessing gave the same MLP Pearson (0.603).
> - Phase 2 retrained on the official embedding after the fixes (ADT transformed once, train-median size factor): test Pearson decoded directly 0.613 RNA-only, 0.767 protein-only, 0.763 joint; one flow-matching sample 0.211 / 0.216 / 0.219. The old low score came mostly from scoring a single flow-matching sample; decoded directly, RNA-only (0.613) is now slightly above phase 1 (0.603).
>
> Disclosures: donor 15078 is in training at other sites and is 32.6% of the site4 test cells; the "out-of-site check" is held-out donor 18303 at a training site, which was also the validation set; the 9 panel proteins were chosen by test-set Pearson; the numbers above use the official TEDDY-G preprocessing (counts / total × 10⁴, divided by TEDDY gene medians, top 2,048 rank tokens, no CLS, mean over gene tokens); the earlier `z_512` (table below) deviated from it.

---

## z_512 loading factor (required definition)

| Item | Value |
|---|---|
| Symbol | `z_512` |
| Source | Frozen TEDDY-G last-layer hidden states |
| Pool | **Mean-pool over real tokens** (attention mask), not token 0, not a disease token |
| Context length | **512** (CITE embed; `scripts/03_embed_rna.py --preprocessing legacy` reproduces it; the stored embeddings reproduce at 512 with cosine 1.000000, vs median 0.970 at 1024) |
| Not | TEDDY pretrain context **2048** |
| Not | A dedicated disease / CLS token readout |
| Normalisation | **No TEDDY gene-median normalisation** (the official pipeline divides each gene by its TEDDY median before ranking) |
| Dim | `d_model = 512` after mean-pool |
| Deviation from official TEDDY-G | The official model returns token 0 by default (no CLS token is added, so token 0 is the top-ranked gene); the official tutorial mean-pools over all 2,048 positions, padding included, after gene-median normalisation. With the medians applied (512 tokens, same masked mean) the cosine to the stored `z_512` is **0.649** (median over 48 site4 cells; fully official 2,048-token setups: 0.205 token 0, 0.419 tutorial mean). The results above are now rerun with the official preprocessing, which gave the same phase-1 Pearson; this table documents the earlier embedding. |

Bridge sidecar `outputs/anm_cite_bridge/z_rna_512.npy` stores **L2-normalized full** `z_512` (export row order). Compact `z_rna_export.npy` keeps only `z_keep=32` leading dims for typed-event JSONL — **not** the reverse-step1 probe space.

---

## Implemented (Mode B)

- RNA→ADT predictions from our head on frozen TEDDY's embedding (`best.pt`) as typed δu evidence into ANM finite_field
- Editable observer YAML: `bridge_anm/readouts/cite_lineage_O{0,1,2}.yaml`
- Missing-modality abstain demos (honest silence under weak channels; Experiment 2, withdrawn; redesigned as v3 E2 and E4)
- Closed-form LOO / top-1 flip attribution on **declared decisions**
- Scope gate (on the v2 key never better than the fixed rule's own margin; v3 E1.4a inconclusive) / zero-label observer transfer vs TEDDY + trained classifier (see proof reports)
- Mode B reverse **decision** loop on must-separate pairs (veto→complement→verify; Experiment 6, withdrawn; redesigned as v3 E3, E5 and E5-M)

Verified numbers only — cite [`HARD_PROOF_OFFICIAL`](reports/HARD_PROOF_OFFICIAL.md) and [`SCOPE_REFINE_PROOF_OFFICIAL`](reports/SCOPE_REFINE_PROOF_OFFICIAL.md) (official preprocessing); earlier run: [`HARD_PROOF_V2`](reports/HARD_PROOF_V2.md), [`SCOPE_REFINE_PROOF_V2`](reports/SCOPE_REFINE_PROOF_V2.md) (corrected rerun, earlier preprocessing); first run: [`HARD_PROOF`](reports/HARD_PROOF.md), [`SCOPE_REFINE_PROOF`](reports/SCOPE_REFINE_PROOF.md), [`MISSING_MODALITY_ANM_DEMO`](reports/MISSING_MODALITY_ANM_DEMO.md) (withdrawn).

## NOT claimed (Mode A and adjacent)

- Mode A: residual stream as field, layer Jacobian, in-silico gene perturbs (Mode A results so far: v2 probes on the frozen embedding for the NK–T look-alike pairs: [linear](reports/MODE_A_LINEAR_OFFICIAL.md), [MLP / kNN](reports/MODE_A_NONLINEAR_OFFICIAL.md), [token level](reports/MODE_A_TOKEN_OFFICIAL.md); `outputs/mode_a_official/`; and the registered v3 E5, stopped at its derivative check, and E5-M, inconclusive, in [V3_REGISTERED_RESULTS](reports/V3_REGISTERED_RESULTS.md))
- ANM's dynamics: TEDDY does not read the field, so state feedback, operator memory and delays are not exercised
- Gated ±ε / ±ε/2 G1–G4 protocols
- Mode-A reverse closed loop (representation response to perturbs)
- Fusion audit
- ATAC / chromatin foundation

## What current conclusions *are* about

Current proofs speak to **decision-layer sensitivity to the TEDDY + head evidence** (edit observer → abstain / P_f / Q_f / LOO flips), **not** to TEDDY representation response under gene/protein perturbs.

They **cannot** yet answer whether `z_512` (or the export sidecar) is **sufficient for protein readout** in the strong sense. A reverse step-1 “must-separate” probe (near-identical z, different true lineage/protein labels) is a **sufficiency-for-readout** diagnostic — still Mode B / evidence geometry, **not** Mode A perturb-response.

## Honest weak-model note

**ADT-only** was a weak channel in the first-run missing-modality demo (panel Pearson drops; the fixed rule's abstain can stay 0 while Q is mediocre). The abstain story is intentionally honest: ANM silence under weak evidence beats silent over-answer. Not clinical. Experiment 2 is withdrawn (redesigned as v3 E2 and E4). Official prototype: with the default export (one flow-matching sample) the protein-only panel Pearson is 0.35; decoded directly (`--phase2-decode direct`) it is 0.95 (joint 0.95, RNA-only 0.90 via phase 1). These masks feed the measured protein into the encoder, so they are reconstructions and not evidence for the redesign. Do not claim Pearson > phase-1 full-ADT **~0.60** (0.603 with the train-median size factor); phase 2's protein-only / joint Pearson (0.767 / 0.763) are reconstructions, not predictions.

## Cheap next probes (Mode B only)

Four cheap probes ranked in [`docs/reports/CHEAP_PROBES_RANKING.md`](reports/CHEAP_PROBES_RANKING.md). Still **not** Mode A. Gated ±ε G1–G4 here are **decision-layer** nudges on typed evidence, not residual-stream / gene-perturb protocols.


## Reverse decision loop (Mode B — implemented; Experiment 6 headline withdrawn, redesigned in v3)

Script: [`bridge_anm/reverse_loop_small.py`](../bridge_anm/reverse_loop_small.py)

**Veto → complement → verify** on reverse_step1 `z_512` must-pairs. Winner = true-ADT typed evidence + O1 (not `z+` alone). Deep writeup + NotebookLM posters: [`docs/reports/ANM_HELPS_TEDDY_LOOP.md`](reports/ANM_HELPS_TEDDY_LOOP.md). Still **not** Mode A perturb-response.

## Reverse step 1 (sufficiency probe only)

Script: [`bridge_anm/reverse_step1_must_separate.py`](../bridge_anm/reverse_step1_must_separate.py)

Reads full `z_rna_512.npy` (not compact `z_keep=32`) + true ADT panel, finds near-identical z pairs with different lineage/protein labels (**must-separate**). Report: [`docs/reports/REVERSE_STEP1.md`](reports/REVERSE_STEP1.md).

Still Mode B / evidence geometry — **not** Mode A perturb-response.