Skip to content

Academy Primitive

Decision Clinics

Decision Clinics are short artifact-first drills. They do not teach the whole workflow. They force one call under pressure: inspect the packet, choose the move, and defend it before you see the reveal.

Start With Clinic 01 First-Model Clinic Timed Clinic Research Clinic Back To Home

Step 1

Read The Packet

Start from artifacts, not explanation. Clinics should feel like opening someone else's run and deciding what matters first.

Step 2

Make The Call

Choose the model, the stop rule, the next move, or the refusal. The point is to commit before the answer is shown.

Step 3

Compare To The Reveal

After you write the note, compare your reasoning to the reference reveal and decide what evidence would change your mind.

All Clinics

# Clinic Focus
01 Public/Private Restraint Leaderboard restraint — visible gain vs. hidden risk
02 Leakage Or Signal? Feature availability — real signal vs. answer key
03 Review Budget Freeze Threshold policy under queue constraints
04 Overfit Or Underfit? Training curve diagnosis — opposite regimes, opposite fixes
05 Freeze Or Fine-Tune? Transfer strategy under small data budgets
06 Ensemble Temptation Marginal accuracy gain vs. operational complexity
07 Checkpoint Roulette Checkpoint selection from a training log
08 Threshold Under Asymmetric Cost Operating point when error costs are unequal
09 Metric Choice Under An Unusual Task Direct metric optimization vs. a stable proxy under deadline
10 Splitter Choice Under Ambiguity Random, grouped, or time-ordered validation when instructions are silent
11 Augmentation Choice Augmentation strength vs. weak-slice robustness
12 Mock Packet Triage Split, metric, and baseline choices in the first five minutes
13 First Model Defense Defending a score with a metric, baseline, and honest split
14 Tokenizer Choice Tokenizer choice driven by rare and code-mixed slices
15 Prompt Vs Retrieval Vs Fine-Tune Choosing one LLM adaptation lever from failure evidence
16 PEFT Depth Or Full Fine-Tune Adaptation depth under fixed data and compute
17 Kernel Choice Kernel selection after correcting a leaky split
18 Manifold Choice Matching PCA, UMAP, or t-SNE to the claim being made
19 Cluster Stability Stability across seeds before committing to a cluster count
20 IoU Threshold Or NMS Tune Detection operating-point gain vs. movement of the ruler
21 Embedding Reuse Or Retrain Frozen reuse vs. domain adaptation under limited data
22 Data Cleaning Choice Missing-value handling based on the missingness mechanism
23 Feature Selection Or Regularize Feature selection, regularization, and the cross-validation boundary
24 Published Result Trust Reproducibility judgment when a paper and a local run disagree

Why Clinics Exist

Topics teach one workflow move. Examples teach one runnable slice. Tracks teach the full connected workflow.

Decision Clinics do something different:

  • they start from artifacts instead of setup
  • they compress the lesson into one judgment
  • they train restraint, not just execution
  • they make hidden evaluation and weak-slice thinking feel normal

That makes them one of the clearest ways to keep AI Academy distinct from a general tutorial site.

Clinic Loop

Use the same loop every time:

  1. open the clinic
  2. read the artifact packet before the explanation
  3. write a short decision note
  4. reveal the reference answer
  5. state what evidence would justify changing your call

The note should stay short. Four to six sentences is enough if the reasoning is concrete.

What A Good Clinic Produces

A good clinic leaves behind:

  • one selected action
  • one rejected tempting action
  • one piece of evidence that drove the choice
  • one piece of evidence still missing
  • one short stop-or-continue rule

If the student only says which model "won," the clinic failed.

First Clinic

Start with Public/Private Restraint.

It is a strong first template because it trains three habits at once:

  • public gain is not proof
  • hidden evaluation matters more than visible rank
  • the right move can be to stop, not to keep searching

Suggested Sequences

First workflow pack — foundations, data, and honest validation:

  1. Clinic 13: First Model Defense — make the first score mean something
  2. Clinic 22: Data Cleaning Choice — argue from the missingness mechanism
  3. Clinic 23: Feature Selection Or Regularize — keep selection inside the fold
  4. Clinic 02: Leakage Or Signal? — enforce feature-availability discipline
  5. Clinic 10: Splitter Choice Under Ambiguity — choose the deployment-shaped split

Evaluation-under-pressure pack — metrics, budgets, leaderboards, and evidence:

  1. Clinic 01: Public/Private Restraint — resist visible-score pressure
  2. Clinic 09: Metric Choice Under An Unusual Task — optimize the grader you actually have
  3. Clinic 03: Review Budget Freeze — respect the queue constraint
  4. Clinic 08: Threshold Under Asymmetric Cost — use cost-driven operating points
  5. Clinic 12: Mock Packet Triage — make the opening decisions on a clock
  6. Clinic 24: Published Result Trust — calibrate trust when reproduction disagrees

Classical modeling pack — diagnostics, geometry, and complexity:

  1. Clinic 04: Overfit Or Underfit? — diagnose before you fix
  2. Clinic 06: Ensemble Temptation — balance accuracy and operational cost
  3. Clinic 17: Kernel Choice — repair the split before ranking kernels
  4. Clinic 18: Manifold Choice — match the projection to the claim
  5. Clinic 19: Cluster Stability — test stability before naming segments

Deep learning and modality pack — training, transfer, vision, and language:

  1. Clinic 07: Checkpoint Roulette — select from validation evidence
  2. Clinic 05: Freeze Or Fine-Tune? — choose transfer depth under data constraints
  3. Clinic 11: Augmentation Choice — optimize for the weak slice
  4. Clinic 20: IoU Threshold Or NMS Tune — separate model gains from scoring effects
  5. Clinic 14: Tokenizer Choice — inspect tokenization on failing slices
  6. Clinic 15: Prompt Vs Retrieval Vs Fine-Tune — choose one lever from the error buckets
  7. Clinic 16: PEFT Depth Or Full Fine-Tune — trade adaptation depth against iteration count
  8. Clinic 21: Embedding Reuse Or Retrain — compare reuse with domain adaptation

After Each Clinic

Route immediately into the matching workflow:

When To Use Clinics

Use a clinic:

  • after one example, before a full track
  • when the student keeps chasing the flattering score
  • when the weak slice is visible but the next move is unclear
  • when you want a short weekly judgment drill

Use a track instead when the student still needs the full workflow.