Validate Automated Codes Before They Enter ENA
Use the shared synthetic design-talk dataset to practise evaluating whether model-generated labels are accurate enough for relational analysis and how classification errors propagate into edges, document the decision, test a defensible alternative, and keep every interpretation tied to source evidence.
Der geprüfte Tutorialtext ist derzeit auf Englisch verfügbar.
Methodenquellen
Vollständiges Tutorial
Validate Automated Codes Before They Enter ENA addresses evaluating whether model-generated labels are accurate enough for relational analysis and how classification errors propagate into edges. This decision belongs in the evidence model, not in a late formatting pass. It affects which coded observations can contribute to a network, what the plotted coordinates represent, and how confidently a comparison can be interpreted. The lesson extends the same eight-team synthetic design-talk case used throughout the Academy so that only the target decision changes while the broader learning context remains stable.
You will work from the source CSV, codebook, and blank decision record. The synthetic file is intentionally small and balanced; it supports transparent inspection but no real-world inference. Your goal is to leave an audit trail that another analyst can rerun: the primary choice, its theoretical rationale, one prespecified sensitivity check, the resulting network evidence, and the claim boundary.
Lehrbeispiel
Validate Automated Codes Before They Enter ENA in the design-talk case
Eight fictional teacher-design teams discuss goals, student evidence, strategies, tradeoffs, and revisions. Four are labeled baseline and four scaffolded, with six ordered utterances per team. The recurring question is whether the two conditions organize these ideas differently. For this lesson, the analytic focus is evaluating whether model-generated labels are accurate enough for relational analysis and how classification errors propagate into edges; the tiny balanced file makes every contributing row inspectable while preventing invented population claims.
Übungsdatensatz
Schritt 1
Restate the relational estimand
Write one sentence naming the unit network, the comparison, and the contextual relation represented by a connection. Then explain how evaluating whether model-generated labels are accurate enough for relational analysis and how classification errors propagate into edges can change that estimand. Keep the distinction between raw observation, coded evidence, accumulated connection, normalized vector, projected coordinate, and substantive claim visible. If the target choice cannot be linked to the question, pause before opening a model.
PrüfpunktThe decision record names one unit, one bounded context, one comparison, and the exact relation a weighted edge represents.Schritt 2
Freeze the shared baseline specification
Record the dataset filename and hash; team_id as unit; conversation_id as conversation; line_number as order; condition as metadata; the five binary code columns; the current window, weighting, normalization, and rotation; exclusions; and software version. This frozen baseline prevents an apparent change from being attributed to evaluating whether model-generated labels are accurate enough for relational analysis and how classification errors propagate into edges when another setting changed at the same time.
PrüfpunktA second analyst can recreate the baseline model without relying on interface memory or undocumented defaults.Schritt 3
Implement the primary decision
Apply this prespecified procedure: retain prediction scores, create a human-labeled evaluation set, report per-code precision and recall, inspect subgroup performance, choose thresholds before ENA, and review uncertain or out-of-domain cases. Save the configuration before inspecting the preferred group pattern. Export unit coordinates, accumulated and normalized edge tables, group summaries, and diagnostics rather than keeping only a screenshot. Mark any manual judgment with a date and rationale so later review can distinguish planned configuration from reactive adjustment.
PrüfpunktThe primary run has a dated configuration, machine-readable outputs, and a rationale written before substantive interpretation.Schritt 4
Inspect units before group averages
Review all eight unit networks, their coded-row exposure, connection magnitudes, and positions. Identify all-zero, sparse, or influential cases and verify that every unit belongs to the intended condition. A mean network can look coherent even when only one team supplies an edge, so record unit coverage for the strongest visible relations and inspect whether within-group heterogeneity contradicts a simple group story.
PrüfpunktEvery interpreted group edge has a unit-coverage count, and no case was silently dropped because its pattern was inconvenient.Schritt 5
Run the planned sensitivity check
Without changing unrelated settings, rebuild networks under plausible thresholds and with human-corrected labels for the evaluation subset to identify edges that are especially vulnerable to false positives or negatives. Save the primary and alternative outputs side by side and compare connection ranks, unit coordinates, group means, diagnostics, and the wording of the conclusion. Sensitivity analysis is not a search for the most separated plot; report stable, weakened, reversed, and ambiguous patterns with equal visibility.
PrüfpunktThe record states exactly what changed, what stayed fixed, and whether the main interpretation survives the alternative.Schritt 6
Return from edges to evidence
Close the interpretive loop: trace each interpreted automated-code edge to source text and classifier outputs, including errors and manual overrides rather than only successful examples. Read enough surrounding rows to recover the conversational context and examine counterexamples, not only quotations that fit the network. Preserve team, conversation, and row identifiers in the audit output. If the edge interpretation conflicts with the excerpts, revise the interpretation or investigate the coding and window rather than trusting the graph.
PrüfpunktThe interpretation cites traceable rows from multiple units and records at least one contradictory or boundary case.Schritt 7
Write the decision and claim boundary
Write three layers in order: the observed source evidence, the modeled pattern under the primary specification, and the explanation the synthetic design cannot establish. Include the sensitivity outcome and this boundary: Good aggregate classification performance does not guarantee reliable rare-code connections or fairness across groups. ENA can magnify systematic co-classification errors into persuasive network patterns. End with the exact files needed to reproduce the lesson and a note that the exercise demonstrates workflow logic rather than a finding about real teachers or scaffolds.
PrüfpunktThe final paragraph separates observation, model output, interpretation, uncertainty, and non-causal limitation in explicit language.