The most striking preliminary result is that zero-shot Claude Haiku achieved 77.8% top-1 accuracy with a macro F1 of 0.75, making it the strongest classifier on the partial annotation set. Meanwhile, the fine-tuned DeBERTa model, trained on expert narrative labels paired with behavioral text, collapsed to zero accuracy — a statistically significant difference at p=0.023. This apparent inversion of expected performance reflects a fundamental constraint in machine learning: a hundred-million-parameter transformer cannot be meaningfully trained on a hundred examples without catastrophic overfitting. The fine-tuning failure is not a methodological failure but rather a signal about the sample-size boundary. At the current dataset size, the pretraining knowledge encoded in Claude Haiku is already sufficient to extract archetype distributions from behavioral text alone. The expert labels, which represent Nina Harris's reading of corporate narrative material, are not yet winning against what an unsupervised reader can extract.

A Jungian brand archetype, as applied in strategy practice, is a contested construct. Mark and Pearson's twelve categories — Sage, Hero, Lover, Outlaw, Magician, Innocent, Explorer, Caregiver, Ruler, Creator, Jester, Everyman — form working vocabulary, but practitioners do not assign brands to single archetypes. Instead, they describe blends: a company might be seventy percent Creator and thirty percent Magician. Trained experts disagree at the edges of these distributions, which is not noise to average away but information to preserve. Treating archetype as a single categorical label discards what experts actually have. The stakes are whether soft-label expert annotation combined with machine learning can extract more stable archetype distributions from behavioral text than a zero-shot model that has never seen an expert's judgment. If supervised learning wins, we have validated a pathway to classification systems that respect professional nuance. If it cannot, we have learned something important about what modern language models understand without supervision.

The Jungian archetype framework, rooted in analytical psychology, has been operationalized in brand strategy through Mark and Pearson's taxonomic work. The twelve archetypes span psychological and behavioral patterns: the Sage embodies truth and clarity; the Hero represents courage and overcoming; the Lover seeks connection and intimacy; the Outlaw challenges norms; the Magician creates transformation; the Innocent pursues happiness and safety; the Explorer craves discovery; the Caregiver serves others; the Ruler commands order; the Creator generates innovation; the Jester brings humor; the Everyman seeks belonging. In brand practice, this vocabulary has proven intuitive and durable. Practitioners use it to describe corporate personality, messaging strategy, and stakeholder relationships. The framework is not rigid — brands are understood as blends, moving along continuums, and shifting across contexts. This flexibility is a strength when applied by trained experts; it is a challenge when reducing the archetype to a single categorical variable for computational analysis.

Traditional machine-learning classification would assign each company to a single archetype — a hard categorical label. This approach has an immediate problem: it forces the expert to make a binary decision at an artificial boundary, discarding the probability mass and uncertainty that the expert actually has. A practitioner who believes a brand is primarily Creator but has elements of Magician cannot express that judgment in a system demanding a single label. Soft-label annotation solves this by asking experts to assign probability distributions across all twelve archetypes, respecting that real brand identities are blended. BrandArchetype captures Nina Harris's expert annotations as probability distributions, preserving the information experts actually have. This is not merely more granular; it is fundamentally different from categorical assignment. When the classifier predicts soft distributions rather than hard categories, the cost function changes, the learning dynamics change, and the information content of the training signal changes. An expert who says a brand is sixty percent Creator, twenty-five percent Magician, fifteen percent Innovator provides much richer supervision than one forced to choose Creator.

The causal structure is a latent-variable model. At its center is A, the true archetype distribution for a company — unobserved, latent, the ground truth we infer. Two separate measurement processes generate noisy readings of A. First, N represents Nina Harris's expert annotation, derived from her reading of corporate narrative materials: annual reports, CEO letters, brand guidelines, public statements. This is a subjective, expert-informed reading of the latent archetype. Second, T represents behavioral text — the company's actual marketing and communication corpus from SEC filings, earnings calls, regulatory filings, and news archives. This is a separate, independent source of information about the same latent construct but arriving through a different channel with different noise characteristics. The classifier Ŷ learns the mapping from T to A. Supervised classifiers see both T and N during training; they predict A from T alone at test time. The zero-shot LLM receives only T and definitions; it never sees N. Critically, N and T are independent measurements of A. Their agreement builds confidence in A; their divergence creates the Tier 3 stratification.

One of the most powerful features is stratification by say/do alignment. Most companies show coherence between what they claim and what they do. A company narrating itself as Hero — overcoming obstacles, driving innovation, bold action — tends to have behavioral patterns consistent with that narrative. But some diverge sharply. A company claiming Caregiver identity (service, compassion, protection) while carrying regulatory violations, consumer harm, or anti-competitive behavior presents a different archetype in action. BrandArchetype handles this divergence explicitly rather than smoothing it away. The classifier is shown these misalignments as labeled examples. When narrative identity is Caregiver but behavior shows Enabler or codependent patterns, that gap is information — not error to average. Similarly, Ruler narrative paired with tyrannical behavior, or Magician narrative masking manipulative conduct, are shadow archetypes revealing the distance between presentation and practice. This stratification matters: the supervised classifier learns not only to extract archetypes from text but also to capture the specific ways narrative and behavior can diverge.

Three classifiers are evaluated in parallel on held-out companies. First, TF-IDF combined with support vector machines — bag-of-words with a linear classifier. This is cheap, interpretable, uses fixed vocabulary, and establishes whether surface lexical signal suffices. Second, zero-shot Claude Haiku, prompted with the twelve archetype definitions and behavioral text alone. No training examples. Whatever it knows about archetypes comes from pretraining, which plausibly already contains marketing-textbook analyses of well-known brands. Third, fine-tuned DeBERTa — a 100-million-parameter transformer trained on (T, N) pairs, minimizing cross-entropy loss against Harris's soft-label distributions. The identification claim has clean shape. If supervised models beat zero-shot on out-of-sample companies, the expert label N carries information the LLM does not extract from T alone — proving annotation is not redundant with pretraining. If zero-shot wins or ties, either N is redundant with consensus already encoded or supervised models are starved of training data.

The dataset comprises one hundred companies, a number reflecting the annotation bottleneck. Expert annotation of soft probability distributions across twelve archetypes, stratified by say/do alignment, is labor-intensive. Nina Harris is conducting the annotation, reading each company's narrative materials and assigning probability distributions reflecting her professional judgment about the archetype blend. Behavioral text is drawn from multiple sources: SEC filings (10-K, 10-Q), earnings call transcripts, regulatory decision records, and archived news coverage. These collectively provide a behavioral record — actual communication patterns, regulatory interactions, and public reputation from third-party reporting. The sample-size constraint is critical to understanding results. One hundred examples is too small to train a 100-million-parameter model without catastrophic overfitting. This is why fine-tuned DeBERTa achieved zero accuracy — not because the approach is wrong but because the data-to-parameters ratio is catastrophically unfavorable. The V1 manuscript acknowledges this and proposes a frozen-encoder architecture with lightweight head for the next iteration, a standard solution to the sample-size problem.

Preliminary results from the V1 manuscript reveal a striking pattern. Zero-shot Claude Haiku, given only archetype definitions and behavioral text, achieved 77.8% top-1 accuracy with macro F1 of 0.75. This is the strongest preliminary classifier on the partial annotation set. In contrast, fine-tuned DeBERTa achieved 0% accuracy with macro F1 of 0.00 — every prediction wrong. McNemar's test between zero-shot Haiku and fine-tuned DeBERTa returned p=0.023, indicating the gap is statistically significant. TF-IDF achieved intermediate accuracy, though exact numbers cut off in the manuscript. The zero-shot result is powerful: pretraining contains sufficient information to extract archetype distributions from behavioral text without any expert-labeled examples. The fine-tuning failure is unsurprising given n=100 and a 100-million-parameter model. This is a sample-size constraint, explicitly flagged in V1 as expected. The authors note that the next iteration will use a frozen encoder with lightweight prediction head, a standard architectural choice reducing free parameters and mitigating overfitting on small datasets.

BrandArchetype is fundamentally a variable-identification problem, not a confounding-adjustment problem. The research question is not: what causes archetype? Rather: what is the archetype, and can we measure it reliably? This is the vocabulary of Chapter 13. The case makes three points about causal inference and variable construction. First, soft-label expert annotation preserves information that hard categorical assignment destroys. When experts assign probabilities rather than a single label, they encode actual judgment, not forced binary choice. A classifier trained on soft distributions learns richer signal. Second, the latent-variable structure — two independent noisy measurements of the same construct — is central to causal modeling. Here, narrative N and behavioral text T are independent readings of A. Their divergence tells us something; their agreement builds confidence. Third, most subtly, pretraining's power means expert-label information is not fixed. It depends on what the base model already knows. If Claude Haiku has encoded marketing consensus through pretraining, then Harris's soft-label annotations may be redundant with information the model extracts without them. This is not a failure of supervision; it is a statement about what information exists in different channels.

The thesis emerging from BrandArchetype is that expert annotation is conditionally informative. It adds causal identification power only when the base model lacks access to the consensus or knowledge the expert embodies. At the current dataset size, with zero-shot Claude Haiku already achieving 77.8% accuracy, expert labels are not yet winning — a sign that pretraining has absorbed enough marketing-textbook knowledge about brand archetypes that additional supervision is redundant. This is not a failure of the expert or the annotation process. Rather, it reveals a fundamental truth about variable identification in the era of large language models: the information content of supervision is not absolute. It depends on what the model already knows. If this task were posed with a smaller, less-trained model, expert labels would almost certainly provide marginal identification power. If the task were novel or specialized, expert annotation would be irreplaceable. BrandArchetype thus contributes a more nuanced view of soft-label annotation. It is not universally superior to categorical assignment. Its value lies in preserving the full probability distribution of expert judgment, which matters when those distributions are not already encoded in the learner's pretraining. As research continues with larger annotation sets and different architectures, the relationship between pretraining knowledge and expert information will become clearer.