The arithmetic from unit §4. Classify 1,000 processes at 95% accuracy: 50 errors. Identity errors lose 50 edges and add 50 wrong ones. Direction errors each remove a true edge and add its reverse, so any direction-sensitive statistic takes double the damage. Worse, the bias has a direction: flipping 5% of edges in a population with a true 5% reciprocal-pair rate raises measured reciprocity — toward the publishable answer. Instructor script: "This is a bias, not noise, and it points toward the exciting result. That is the worst possible property for an error to have."
Fix the four families first. Confidence comes from which families agree, not from how many observations of the same kind you stack up. Family A, organelles: ribosomes effectively absent from axons — the single best cue when visible. Family B, geometry: axons hold caliber between boutons; dendrites taper, and daughter branches are thinner than the parent. Family C, polarity: vesicle cluster at a synapse -> presynaptic -> axon; PSD -> dendrite or soma. Family D, context: myelin, a fibre bundle, an AIS off a soma.
The taper rule: at a dendritic branch point the daughters' cross-sectional areas are systematically smaller than the parent's; at an axonal branch point caliber is roughly preserved. It is a Family B cue fully independent of organelle staining quality — valuable exactly when the ribosome cue has failed you.
Works as a live exercise: cover the right panel, ask the room to call five or six profiles on the left, then reveal. Note how many are genuinely hard. Instructor script: "The right panel is a six-class subcompartment model, not ground truth. Where you disagree with it, write down which cue family you are relying on — that is the evidence chain you would need to overrule it."
At 4 nm a Type I PSD is 8–12 pixels — the asymmetry is genuinely subtle, and H01's excitatory/inhibitory classifier is right about 85–87% of the time. Both calls are Bin B claims: "putatively excitatory (asymmetric)", "putatively inhibitory (symmetric)". Where the claim matters, the presynaptic cell type is usually the stronger evidence.
Myelin is a Family D context cue and a decisive one: dendrites are never myelinated. In the protocol it is step 3 — myelin or a fibre bundle -> axon, high confidence. Use the second bullet to calibrate expectations about resolution: knowing what you cannot see at 4 nm is part of reading the image honestly.
Spines are Family B: axons never bear them. In dense neuropil, though, you rarely see a whole reconstructed arbor like this — you see one cross-section. Run step 1 of the protocol first: a single clean synapse anywhere in view settles polarity faster than a long scroll, and the side carrying the vesicle cluster is the axon.
Ask learners to name three differences, label each with its family, then ask which they would still see at ten percent weaker membrane contrast. That ordering is their personal cue-robustness ranking — the calibration lab is where they measure it.
Resist stacking more of the same evidence: high confidence requires two cues from different families. If everything available belongs to one family, the honest output is medium confidence with the missing family named. Watch for over-read absence: no ribosomes in a 150–200 nm process is weak evidence, because it may be too thin to show a polyribosome in any one plane. Absence counts only when the feature would have been visible if present.
The PSD is why Family C works: a density on the receiving membrane means this process is postsynaptic. Then give the exceptions (unit §2) before anyone applies "PSD implies dendrite" mechanically: 1. Dendro-dendritic synapses — olfactory bulb, thalamus. 2. Axo-axonic synapses — chandelier cells onto the axon initial segment. The AIS is postsynaptic and still an axon; the tell is membrane undercoating plus fasciculated microtubules within roughly 20–60 µm of a soma. 3. Presynaptic dendrites in retina (amacrine cells). 4. Invertebrate neurons with mixed input/output regions on one neurite. Instructor script: "What tissue am I in? That is a question about the dataset, not the image, and it should be answered once, in the protocol."
The unit's local protocol, in cost order: (1) a synapse in view? which side? (2) ribosomes? (3) myelin or fibre bundle? (4) follow through z for 2–3 µm — beaded swellings with vesicles -> axon; steady taper or thinner daughters -> dendrite; a spine -> dendrite. (5) still unresolved -> UNCERTAIN, with the missing cue named. Cue conflict is a detector, not a tie-break: vesicle clusters and ribosomes in the same cortical process is a merge-error alarm. Inspect the path between the two observations for where two processes touch.