TEDDY’s embedding puts two cells almost together. They differ. Here’s how ANM’s loop catches it.
TEDDY’s embedding puts some myeloid and T cells at cosine 0.982, and the readout built on it (our head + the soft rule) falsely agrees 85% of the time. ANM’s loop vetoes the embedding for these pairs, swaps in measured protein as typed evidence with the strict rule, and checks the same pairs again, with no retrain.
Redesign prototype and disclosures
This page is Experiment 6, Look-alike cells; it was not rerun. The corrected rerun is done for Experiments 1, 3, 4 and 5: with all markers entering at once and a train-median size factor, ANM makes the same call as the fixed rule on every cell. Known issues here: markers were entered one per step in panel order, a hidden weight; predictions were scaled by a per-cell factor computed from measured protein; the answer key has no NK or out-of-scope class.
Redesign prototype (official TEDDY-G preprocessing; this page’s first-run numbers used the earlier, non-official one): 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). On site4, 1,053 of the 1,506 NK–T look-alike pairs involve “gdT CD158b+” cells, which look NK-like in both predicted and measured protein (an annotation problem, not a pipeline failure); on the other 453 pairs the median within-pair gap in predicted (TEDDY + head) 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 probe separates NK from T at every TEDDY layer (held-out AUC 0.991–0.997); on these pairs a linear protein probe keeps about as much of the gap as our head, but an MLP probe on the same embedding keeps about twice as much for CD56, CD94 and CD3 (0.47 / 0.59 / 0.61 vs 0.24 / 0.31 / 0.27), and gene-token states keep more than the gene-mean. v2 read this as: part of the loss is in our head’s readout and in mean pooling; part may not be in the detected RNA (one site; pairs chosen for closeness in the embedding). Superseded by v3: E3 rejected the repair reading, and the head’s compression on registered look-alike pairs is descriptive only (gap ratio 0.1169 flagged vs 0.3985 unflagged, overlapping intervals; 0.4729 vs 0.5137 without gdT CD158b+ pairs); E5-M, which tested whether gene-mean pooling discards what a token readout uses, was inconclusive for the one observer tested.
Donor 15078 is in training at other sites and is 32.6% of site4 test cells, and the 9 panel proteins were chosen by test-set Pearson.
1,490 myeloid↔T must-separate pairs · before: TEDDY + fixed rule (soft) · after: measured protein as typed evidence, strict rule · reverse_loop_small_stats.json
The loop, one step at a time
Click a step or use ← / →.
1 · What TEDDY + a fixed rule assumes
The readout treats cells with nearby embeddings as the same biology.
What the embedding is
z_rna_512.npy, 16,750 held-out cells).The implicit promise
2 · Veto: near-merged cells that differ
Nearest neighbours at cosine ≥ 0.98 whose measured proteins disagree on lineage are must-separate pairs.
in TEDDY’s embedding
Myeloid↔T pair cosine (1,490 pairs)
How big is the gap (cosine ≥ 0.98)
Of the 5,000 pairs used, how many differ in…
Myeloid↔T: measured protein difference on focus markers
TEDDY + fixed rule on these pairs (soft rule)
3 · Complement: three cheap fixes, no TEDDY retrain
Pick a fix: a better question, new evidence, or glued-on features.
4 · Verify: the same 1,490 pairs
Of 100 pairs with both cells called, how many wrongly share a call?
Scoreboard: every fix, every metric
First run. Hatched: the feature swap z⁺, distances only; key below.
Table view
What ANM adds over a feature swap
Glued-on protein separates one list; distances make no decisions.
Feature swap: z⁺ = embedding ⊕ protein features
Fixed myeloid↔T list, share still at cosine ≥ 0.98:
Across all cells (earlier global test, myeloid held out), z⁺ changes the must-separate rate by:
ANM decision loop
each input declares its source to ANM’s engine
soft → strict is a written change (0.12 → 0.28, key markers ×2)
declines on insufficient evidence
same pairs → false agreement, accuracy, soft separation
Claim boundary
“Reverse loop” is a decision loop: we change what the readout sees and asks, not TEDDY’s representation.
Open loop: TEDDY does not read the field, so ANM’s dynamics are not tested here; not yet, not impossible (why).
Claimed · the decision layer
- Veto → complement → verify on must-separate pairs
- The readout on TEDDY’s embedding falls short here
- Measured protein as complement evidence, this test only
- A rewritten question (strict rule), no retraining
- Normally measured protein stays the answer key
Not claimed
- TEDDY’s internals or representation response (Mode A)
- Pearson above phase-1 ≈ 0.60
- Fusion audit · clinical superiority
- ANM’s dynamics (feedback, memory, delays): open loop
- A final headline: withdrawn (redesigned as v3 E3, E5 and E5-M)
- Retraining TEDDY or our head (
best.pt, the head’s checkpoint)
Withdrawn pending redesign (29 Sep 2026) and the NK–T prototype
Withdrawn because markers entered one per step in panel order, predictions were scaled by a per-cell factor from measured protein, and the answer key has no NK or out-of-scope class. The redesign prototype (official TEDDY-G preprocessing) finds 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). On site4, 1,053 of the 1,506 NK–T look-alike pairs involve “gdT CD158b+” cells, which look NK-like in both predicted and measured protein (an annotation problem, not a pipeline failure); on the other 453 pairs the median within-pair gap in predicted (TEDDY + head) 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: a linear probe on the embedding is compressed about as much as our head, an MLP probe on the same embedding keeps about twice as much (CD56, CD94, CD3), and gene-token states keep more than the gene-mean, 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 (one site; pairs chosen for closeness in the embedding). Superseded by v3: E3 rejected the repair reading (v3 E3), and E5-M, the sufficiency test of the gene-mean, was inconclusive for the one observer tested (v3 E5-M).
Open loop, in full: 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. ANM’s step is any real processing order, not necessarily physical time. The nine marker predictions come out of one prediction with no order (entering them in panel order was the hidden weight); TEDDY’s transformer layers are a real order, so the natural test is Mode A, over layers.
Posters
Static first-run summaries, for slides.
