PlatPhormNews field notebook
A practical notebook for understanding model behavior: prompting habits, interpretability methods, verification rituals, and tools that make AI systems less mysterious.
Use this when the room needs one concept, one artifact, and one next action. The slides stay connected to entries, Dictionary, and Notes.
6 slides
Each slide pairs one concept, one artifact, and one next action
Start here / 01
Students cannot inspect model behavior if the key vocabulary is unstable.
Open with the words that matter: retrieval, receipt, confidence bucket, dry run, abstention, provenance. Once the class shares the terms, the demo becomes an evidence exercise instead of a vocabulary scramble.
AHA checkpoint
The glossary is not prep material. It is part of the operating system for the lesson.
Name the term before the trick.
Put the artifact on screen.
Ask what changed in the learner's next move.
Each animation is a deterministic concept trace, not a live model run. The goal is to make the hidden review move visible enough to use.
Uncertainty trace
This toy trace turns a polished answer into token confidence. The cold tokens are where a reviewer should slow down first.
A low-margin token is where the model nearly chose a different world.
The generated text looks continuous, but each token still carries a margin against alternatives.
Operator move
Read names, dates, citations, and refusal-boundary words with the token margin visible.
These tips turn PlatPhormNews root discovery surfaces into classroom explanations. Start with the front-door idea, then ramp into MCP, graph evidence, structured data, trace safety, and agent policy.
Review root siteInteractive tips
Root MCP manifest and trust policy
Source10th-grade version
Think of the public DMZ like the front desk of a school. Anyone can ask where things are, but they cannot change grades, unlock doors, or edit records.
Operator version
The root site publishes public discovery, network graph, stories, route compliance, and MCP introspection while reserving sync, validation, reports, and policy updates for protected actions.
Real-world view
Your job is to check that the front desk gives useful directions without accidentally becoming an admin panel.
AHA moment
Public does not mean powerful. Public means safely inspectable.
Safe boundary
Do not test destructive, private, protected, credential, admin, or mutation paths without explicit authorization.
Classroom move: Ask students to sort five actions into public read or protected mutation before showing any API response.
First-time learner
Explain the difference between public read, protected action, and private data.
Sort routes from robots.txt into safe-to-read and stop-sign groups.
Careful operator
Use sitemaps, RSS, llms files, OpenAPI, and JSON-LD to inspect a site without guessing.
Build a source-backed checklist for one public surface.
Agent workflow designer
Read MCP tools/resources/prompts and design a public-safe workflow with explicit stop states.
Write a dry-run plan for a read-only MCP inspection.
Real-world operator
Audit public automation behavior while respecting trust policy, robots policy, auth boundaries, trace redaction, and no-secret rules.
Produce a report that says what was inspected, what was skipped, why it was skipped, and what evidence supports each conclusion.
These are public-safe handoffs. This site teaches from Dictionary and prepares Notes prompts; it does not claim to publish definitions, store note bodies, or run protected workflows.
dictionary
Public terminology, definitions, provenance lookup, search, datasets, term graph, and MCP discovery.
Public reads
This notebook links concepts to Dictionary as concept receipts and glossary anchors; it does not publish definitions on the user's behalf.
Protected boundary
Moderation, import, admin, private submissions, audit logs, protected reports, model-assisted moderation, registry mutation, and destructive actions require PLATPHORM_API_KEY when enforcement is enabled.
notes
Public learning notepad for note enrichment, terminology detection, Dictionary lookup, and human-approved definition staging.
Public reads
This notebook gives students prompts and observation formats to use in Notes; it does not store note bodies or stage submissions directly.
Protected boundary
No visitor key entry. Backend calls to Dictionary, Docs, OpenContent, Sheets, Phorm, SVG, and other PlatPhorm services require PLATPHORM_API_KEY server-side.
Define retrieval and relevance before students score chunks.
Notes prompt: What retrieved chunk looked authoritative but did not answer the question?
Treat a source, command, row, calculation, or assumption as a claim handle.
Notes prompt: Which sentence in the answer needs a receipt before you would reuse it?
Show the action plan before any tool mutates state.
Notes prompt: What target ids, selectors, or rollback notes were missing from the tool preview?
Make unknown, partial, protected-required, and needs-review legitimate outcomes.
Notes prompt: Where should the model have stopped instead of filling the answer?
Keep examples that should pass, refuse, redirect, or require more evidence.
Notes prompt: Which prompt should become a fixture because it revealed a boundary?
Ask where a term or fact came from before teaching from it.
Notes prompt: What source or scope would change how you teach this definition?
Dictionary stabilizes shared terminology; Notes captures learner confusion; this notebook turns both into repeatable operating habits.
A public-safe handoff is still useful when it is honest about what it cannot mutate.
The best integration is not a button. It is a clear ownership boundary plus a next action.
The notebook is most useful when it becomes a repeatable review motion. These playbooks show what to do when answers, prompts, and agent actions start carrying real risk.
Playbook
A model produced a confident answer and nobody knows whether the weak point is retrieval, reasoning, or source support.
A reviewer can identify which claims are supported, which claims need receipts, and which answer parts should be withheld.
Locate fragile spans
Highlight names, dates, numbers, citations, and low-margin tokens before reading the answer as prose.
Aha: Hallucinations often hide where fluency is highest but token certainty is thinnest.
Separate retrieval from generation
Mark each retrieved chunk as direct support, background, contradiction, or irrelevant.
Aha: Bad context can make a good model sound wrong with confidence.
Attach receipts
Force each important claim to point to a source, row, command, calculation, trace id, or assumption.
Aha: The unsupported claim does not vanish; it becomes visible enough to reject.
Decide the confidence bucket
Assign high, medium, low, or unknown based on evidence quality and contradiction level.
Aha: Confidence matters when it changes the next action.
Playbook
A system prompt needs improvement, but every wording edit risks changing refusal, format, or evidence behavior.
The prompt change ships with a semantic diff, targeted fixtures, and a clear before/during/after eval record.
Shrink the contract
Separate durable role, hard boundary, output contract, and escalation rule from task-specific detail.
Aha: A smaller system prompt is not weaker; it is easier to audit.
Diff the meaning
List new obligations, removed constraints, softened verbs, and examples that imply new rules.
Aha: A one-word modal change can quietly become a policy change.
Run the eval sandwich
Baseline before, targeted fixture during, broad regression after.
Aha: The middle eval catches drift while the edit is still cheap.
Mutate the boundary
Test harmless variants that preserve intent while changing tone, order, language, and framing.
Aha: A boundary that only catches the obvious phrasing is not a boundary.
Playbook
An agent can call tools, but the operator needs confidence before it mutates, publishes, replays, deletes, or spends.
Every action moves through preview, inspect, authorize, execute, and verify without hiding protected states.
Preview the action
Require target ids, arguments, expected effect, refusal condition, and rollback notes.
Aha: Permission to plan is not permission to execute; the preview is where unsafe scope becomes inspectable.
Respect blocked states
Return unknown, partial, protected-required, or needs-review when evidence or authorization is missing.
Aha: A designed stop is better than a graceful guess because it preserves trust and gives the operator a next move.
Spend checks by risk
Wake up extra verification when impact is high, evidence is weak, or the action is protected.
Aha: Latency can be evidence-driven instead of uniformly expensive.
Verify after execution
Check the actual changed state, not only the tool's success response.
Aha: The world, not the model, gets the final vote, so post-action evidence is part of the action itself.
Case study
Setup: A model summarizes a report and includes a link beside every claim.
Failure: The links exist, but two citations point to documents that discuss the topic without supporting the exact sentence.
Hidden signal: The unsupported claims are entity/date claims with weak token margins and broad source references.
Repair: Use receipts per claim, then triangulate the high-impact facts with independent sources before publishing.
Aha: A citation beside a sentence is not the same thing as support under the sentence.
Case study
Setup: A long prompt includes a critical policy exception in the middle of retrieved context.
Failure: The answer follows the opening task and closing format while ignoring the exception.
Hidden signal: The missed fact was technically present but surrounded by lower-value context.
Repair: Move the task contract to the opening, repeat the exception at the end, and rank retrieved chunks before generation.
Aha: Presence in context is not the same thing as usable salience.
Case study
Setup: An agent is asked to clean a set of records and calls a bulk mutation tool.
Failure: The selector is broader than the user intended and changes records outside the target set.
Hidden signal: The model never previewed target ids or stated a rollback condition.
Repair: Add a dry-run gate, require explicit target ids, and make protected-required a first-class state.
Aha: Tool safety starts before execution, not after the audit log catches damage.
Case study
Setup: A prompt is edited to make answers friendlier and less formal.
Failure: The model starts complying with requests it previously redirected because the new examples imply looser policy.
Hidden signal: The text diff is small, but the semantic diff changes tone, authority, and refusal posture.
Repair: Run a semantic diff, then use refusal fixtures and harmless mutants before shipping.
Aha: Style examples can become behavioral instructions if you do not test the boundary.
Notebook state
55
published notes
13
topic tags
55
aha moments
7
reading paths
Reading paths
A facilitator path for turning notebook ideas into shared lessons, demos, checkpoints, and self-quizzes.
11 notes
Start with Explain it like grade 10, then tighten itHow to move from learner notes to shared vocabulary without pretending AI-generated definitions are approved truth.
7 notes
Start with Use Dictionary as the concept receiptPublic-safe patterns for linking PlatPhorm services while preserving ownership, auth boundaries, and honest handoffs.
9 notes
Start with Use integration cards instead of vague platform linksPractical habits for getting less brittle model behavior before adding heavier tooling.
9 notes
Start with Design the output schema before the promptWays to inspect internal signals and move from pretty explanations to causal evidence.
6 notes
Start with Read the logprobs, not the answerSmall rituals that keep AI-assisted work tied to evidence, tests, and explicit uncertainty.
13 notes
Start with Inspect one claim before judging the whole answerPatterns for agents that can act, route work, abstain, or spend extra verification budget.
9 notes
Start with Write the safe agent plan before the agent actsEntries
55 of 55 notes visible
Start with the version a smart first-time learner can repeat. Then add the sharper operator version: exact terms, evidence needed, boundary, and next action. Simple first is not dumbing it down; it is building the ramp.
A clear first explanation makes the advanced one land faster.
#classroom
Simple language is the ramp, not the ceiling. The expert version lands faster after the learner can repeat the plain one.
Rewrite one advanced AI idea twice: first as a tenth-grade explanation, then as an operator checklist with evidence and boundary.
Do not start by asking whether the AI answer feels good. Pick one important claim, ask what would prove it, and follow that evidence. One inspected claim teaches more than a vague reaction to a polished paragraph.
#verification
One inspected claim can reveal whether the whole answer is evidence-shaped or just confidence-shaped.
Pick the highest-impact sentence in a model answer and write the exact receipt that would prove or disprove it.
A useful review question points at something visible: a cited sentence, a JSON field, a tool argument, a trace id, a route, or a screenshot. Vibe questions create opinions. Artifact questions create inspectable answers.
#classroom
A review gets sharper when the question points to an artifact everyone can inspect together.
Replace one vague question like is this good with three artifact questions about a source, field, route, or tool argument.
Before using an AI tool or platform route, sort the action into three buckets: public read, protected action, or private data. Most unsafe automation begins when those buckets are mixed together.
#platform
Most automation mistakes become visible if you sort public read, protected action, and private data before the tool runs.
Take five routes or UI actions and label each one public read, protected action, or private data before designing the workflow.
A beginner can understand agents if you frame them as plans plus tools plus boundaries. Before an agent acts, write what it may inspect, what it may not touch, when it must stop, and what proof it must show.
#agents
An agent is easier to trust when its allowed moves, stop rules, and proof requirements are visible before execution.
Write a dry-run agent plan with allowed reads, forbidden actions, stop conditions, target ids, and verification evidence.
A Dictionary definition stabilizes what a word means. It does not prove that a model's claim is true. Use definitions to agree on language, then use receipts to prove claims.
#dictionary
Definitions stabilize language. Receipts prove claims. Mixing those jobs creates false confidence.
For one AI term, write the definition question and the proof question separately, then answer both with different sources.
A button should make a promise the product can keep. If it copies text, say copy. If it opens Notes, say open Notes. If it triggers a protected report, require PLATPHORM_API_KEY and show the boundary.
#integration
A button is a product promise. If the action is only copy, link, or read JSON, the label should say exactly that.
Audit three UI buttons and write what each one actually does, what it does not do, and what proof the user receives.
Before asking AI to summarize a system, list the sources it is allowed to use and what each source can prove. The summary should be the last step, not the first guess.
#verification
A summary becomes safer when the source map exists before the prose starts.
Create a source map with source, what it proves, what it cannot prove, and checked date before asking AI for a summary.
The best teaching moment is often the wrong first guess. Let students make a safe mistake, inspect why it failed, and name the habit that would catch it next time.
#classroom
A safe mistake can teach the repair habit more clearly than a flawless demo.
Design one harmless failure case, let learners diagnose it, then write the habit that would catch it next time.
Students can ramp into advanced review with a simple readiness scale: green means enough evidence to proceed, yellow means inspect more, and red means stop because the boundary or evidence is wrong.
#classroom
Readiness labels matter when they change what happens next.
Score one model answer green, yellow, or red and require the reason plus the next action for that color.
A learner cannot evaluate a prompt, score a citation, or review a tool call if the vocabulary is foggy. Define the key terms first, then demonstrate the technique. The trick lands only after the words are stable.
A shared glossary lowers the cost of every later explanation.
#classroom
The vocabulary is part of the interface. If the term is unstable, the lesson is unstable.
Start each teaching block by naming three terms students must understand before they watch the model behavior.
When a lesson depends on specialized AI vocabulary, link the term to a public Dictionary definition. The definition becomes a concept receipt: not proof that the model is right, but proof that the class is using the same meaning.
#dictionary
A Dictionary link can be a concept receipt that proves the class is sharing a meaning before debating evidence.
Attach a Dictionary definition to each specialized term in a classroom slide, notebook entry, or review checklist.
The best glossary terms often begin as learner confusion. When several students write the same fuzzy phrase in Notes, promote that phrase into a candidate definition instead of answering it once and losing the signal.
#notes
Repeated learner confusion is not noise. It is demand for a clearer shared definition.
Collect fuzzy learner phrases in Notes, group repeated phrases, and promote only reviewed candidates to Dictionary.
A strong AI lesson has three beats: explain the mechanism, inspect a live or concrete artifact, then repair a failure. Students remember the loop because every abstract idea earns a visible consequence.
#classroom
Students remember a concept when they see the mechanism, inspect evidence, and repair a failure themselves.
Teach one topic with three beats: explain the mechanism, inspect a concrete artifact, then repair the broken part.
Do not leave a concept at the definition stage. Climb it through four rungs: word, example, failure, workflow. A student understands a term when they can use it to decide what to do next.
#classroom
A term becomes operational only when the learner can use it to choose the next move.
For one concept, write the word, a clean example, a failure example, and the workflow decision it changes.
A live demo creates cognitive load. Do the vocabulary work first so students can spend the demo watching the model behavior instead of decoding the instructor's words.
#dictionary
The demo is where students should watch behavior, not decode vocabulary under pressure.
List the lesson glossary before the first live output and link each term to Dictionary when a public definition exists.
AI can propose a definition, but shared language deserves review. Treat Dictionary publishing like a small release: draft, inspect, approve, publish, and leave provenance behind.
#governance
A bad shared definition compounds because future lessons and workflows inherit it.
Stage candidate glossary definitions as drafts, require human approval, and include source or scope notes before publishing.
A cross-site link should explain what the target service owns, what public data is safe to read, what requires PLATPHORM_API_KEY, and what the current page does not do.
A link without a contract is just navigation.
#integration
A cross-site link is safer when it says what the target owns and what this page will not do.
Write each integration card with owner, public reads, protected actions, and handoff links before adding buttons.
A lesson is not complete when the instructor finishes the content. Stop at the moment where the learner's model should change and ask them to say what they see differently now.
#classroom
The AHA moment is the measurable change in what the learner knows how to inspect next.
Pause after each major concept and ask students to name the first artifact they would inspect now.
When students inspect model behavior, ask them to write observations before conclusions. Notes should capture what happened, what surprised them, and what term they need clarified.
#notes
Observation-first notes preserve the evidence before the learner writes a story about it.
Use the format I saw, I expected, I think it means, and I need a definition for after every demo.
If a term definition will anchor a lesson, inspect its source and scope first. A definition can be useful for one domain and misleading in another.
#dictionary
A definition is scoped evidence. It helps only when its source and domain match the lesson.
Before teaching a term, inspect whether the definition names the domain, source, and intended audience.
Instead of asking whether the answer is good, ask which claim would fail first, which receipt is missing, which term is undefined, and what evidence would change the decision.
#evals
The question teaches the habit. Ask about failure, missing receipts, and next evidence instead of asking if it is good.
Give students four prompts: what fails first, what receipt is missing, what term is undefined, and what evidence changes this.
Repeated misunderstandings are not interruptions; they are product signals. Track them as missing entries, missing definitions, missing examples, or missing classroom exercises.
#classroom
Misconceptions are content backlog with evidence attached.
Track repeated misunderstandings as missing entries, missing definitions, missing examples, or missing exercises.
A teaching surface should remove sidebar noise when the room needs one idea. Full-screen mode turns the notebook into a slide-like sequence without losing links back to entries, Dictionary, and Notes.
#classroom
Full-screen mode is useful when it narrows the room to one concept, one artifact, and one next action.
Use full-screen mode for kickoff, live demo, recap, or AHA checkpoint, then return to the notebook for details.
Terms drift when teams reuse them in new workflows. Schedule periodic review for classroom terms, Dictionary definitions, and notebook entries so old wording does not teach stale behavior.
#governance
Shared vocabulary is a dependency, and dependencies drift.
Review classroom terms after model, tool, or platform boundary changes and update Dictionary plus notebook links together.
Every entry should produce a recall question, an artifact question, and a transfer question. If learners can answer all three, the insight has moved from reading to practice.
#classroom
Learning sticks when recall, artifact inspection, and transfer all work.
Write one recall question, one artifact question, and one transfer question for each high-value entry.
A retrieval system can hand the model a pile of documents and still miss the actual question. Score each chunk for direct relevance, answer support, and contradiction before you let the final answer inherit authority from the word 'retrieved'.
Retrieved is not the same thing as relevant.
#verification
Retrieval only proves that documents were fetched. It does not prove the fetched text answers the question.
Label each retrieved chunk as direct support, background, contradiction, or irrelevant before reading the final answer.
When the output has fields, types, allowed values, and failure states, the prompt stops being a prose wish and becomes an interface contract. The model has less room to improvise because the answer has a shape to satisfy.
#reliability
A schema turns a vague prompt into an interface contract with explicit absence states.
Write the JSON fields, allowed values, evidence pointers, and unknown states before drafting the prompt.
One eval at the end is a smoke alarm, not a control system. Run a small eval before the change, a targeted eval during the change, and a regression eval after. The sandwich catches drift while the work is still cheap to fix.
The middle eval is where most prompt regressions get caught.
#evals
The most valuable eval is often the small one you run while the prompt is still being edited.
Run a baseline set before the change, five targeted fixtures during the change, and a regression set after.
A tool call is not just an answer; it is an action. Before a model can send, delete, replay, publish, or spend, make it produce a dry-run plan with inputs, target, expected effect, and rollback path.
#tooling
A tool call should become inspectable before it becomes irreversible.
Require action preview, target ids, expected effects, refusal conditions, and rollback notes before execution.
One citation can prove that a sentence has a source-shaped object nearby. Two independent citations can show whether the claim survives contact with another account. For consequential facts, triangulation beats citation decoration.
#verification
A second independent source often reveals whether a citation supports the claim or merely resembles it.
Triangulate dates, quotes, benchmarks, allegations, and high-impact facts before presenting them as settled.
If a safety or quality rule only works for the exact phrase you tested, you do not have a rule. Create harmless mutants that preserve intent while changing tone, language, order, and indirection.
#evals
A boundary that only holds for one phrasing is a memorized example, not a reliable rule.
Generate harmless variants that preserve intent while changing tone, order, indirection, language, and role framing.
A prompt edit can look tiny and change the whole behavior. Run a semantic diff that lists new obligations, removed constraints, softened verbs, and changed authority before shipping instruction changes.
A one-word modal change can be a policy change.
#prompting
The smallest text diff can be the biggest behavioral diff when it changes authority, scope, or obligation.
Review prompt edits for new duties, removed constraints, softened verbs, and examples that imply new policy.
Do not add slow verification everywhere. Spend extra model calls, retrieval, self-consistency, or human review only where the first pass shows low confidence, high impact, or weak evidence.
#reliability
Careful does not have to mean slow everywhere. Verification can wake up only where risk wakes it up.
Escalate to extra checks when confidence is low, evidence is weak, impact is high, or an action is protected.
The best model for a job is the one whose failure mode you can tolerate. Route extraction, synthesis, coding, refusal review, and creative exploration by observed behavior, not by leaderboard aura.
#reliability
Routing by failure mode is more useful than routing by reputation because every model breaks differently.
Build a route table from observed task fixtures: extraction, synthesis, coding, refusal review, and long-context use.
Memory makes an agent feel continuous, but stale memory can be worse than no memory. Store decisions, constraints, and stable preferences; refresh facts that drift; never let memory outrank current evidence.
#verification
Agent memory is a cache. Useful caches have scope, freshness, and invalidation rules.
Store stable decisions and preferences, but re-check drift-prone facts before acting on them.
If the only valid-looking answer is a filled answer, the model will fill one. Give it a clean way to say unknown, insufficient evidence, unsafe, or needs human review, and make those states useful.
#reliability
A model fills blanks when the product has no respected way to stop.
Make unknown, partial, blocked, protected-required, and needs-review states valid outputs with next actions.
A model saying 87% confident usually means it learned that numbers sound calibrated. Buckets like high, medium, low, and unknown force confidence to be tied to evidence quality and decision impact.
#verification
Confidence is useful when it changes what happens next, not when it looks mathematically precise.
Use high, medium, low, and unknown buckets tied to evidence type, conflict level, and review action.
A model that only explains why its answer is right will happily polish a mistake. Ask it to list the strongest evidence against its answer before it scores confidence. The missing objection is often the thing that saves you.
Confidence after counter-evidence is more useful than confidence after a sales pitch.
#reliability
The model is usually best at self-correction before it has been rewarded for sounding certain.
Ask for the answer, the strongest objection, what evidence would change the answer, and then confidence.
For high-stakes answers, make the model attach each claim to a source, calculation, line number, trace id, or explicit assumption. Unsupported claims do not disappear, but they become visible enough to reject.
#verification
A citation is not decoration. It is a handle you can pull to see whether the sentence is load-bearing.
Make every claim end with one receipt: source, row, command, calculation, trace id, or explicit assumption.
A giant instruction stack feels powerful until nobody can tell which rule caused the behavior. Keep the non-negotiables small, move task detail into the user prompt, and test the system prompt like code.
If you cannot diff it cleanly, you cannot debug it cleanly.
#prompting
A system prompt is product code. If it is too long to review, it is too long to trust.
Rewrite the system prompt into role, hard boundary, output contract, and escalation rule.
Do not only test the happy path. Save examples that should be refused, partially answered, or redirected, then run them every time you change instructions. Safety behavior drifts just like product behavior.
#evals
Safety quality is not whether the model refuses. It is whether it refuses at the right boundary.
Keep allowed, disallowed, and allowed-with-constraints prompts as fixtures and run them after prompt edits.
The text a model gives you is the tip of the iceberg. Ask the API for token logprobs and you can see where it was confident versus where it basically flipped a coin. A fluent sentence built on 51% tokens is a guess wearing a suit.
Confidence ≠correctness. But low confidence is a great place to look for hallucinations.
#interpretability
The final text is only the winning token path. The uncertainty is still visible in the runners-up.
Render low-margin tokens first, especially names, dates, citations, and first tokens after a policy boundary.
Put the reasoning before the conclusion, never after. If you ask for the answer first and the explanation second, the explanation is just a story it invents to justify a choice it already made. Order matters more than wording.
#prompting
Explanation after an answer often explains a commitment; reasoning before an answer can still change it.
Put scratch work, constraints, and uncertainty checks before the final answer field.
An open-source visualizer that shows which tokens each attention head is looking at, layer by layer. You start to notice heads that only track quotation marks, or ones that follow subject-verb agreement across a whole paragraph.
Most heads are boring. The interesting 5% are worth the dig.
#tooling
Attention is not the whole mind, but it is a useful map of what information is being routed where.
Inspect heads for repeated roles: bracket matching, quote tracking, subject agreement, or copied labels.
Negative instructions are weak anchors. 'Don't mention X' often makes X more likely because you just put X in the context. Prefer describing the thing you DO want. Reserve negatives for hard safety rails, not style.
#prompting
Negation still puts the forbidden idea into context. Attention does not forget because grammar says not to.
Rewrite prohibitions as the positive shape of the output you want before adding hard safety rails.
Freeze the model, grab the hidden activations for a pile of examples, and train a tiny linear probe to predict some property (is this text about money? is it past tense?). If a simple line separates them, the concept is already represented in there, plain as day.
Cheap, fast, and weirdly revealing.
#interpretability
If a tiny linear classifier can read a concept from hidden states, the model has already organized that signal.
Probe one concept across layers and look for where accuracy first rises above a simple baseline.
Turning it up doesn't make answers smarter, it makes them less predictable. For extraction, classification, and anything with a right answer, keep it near zero. For brainstorming, let it wander. People mix these up constantly.
#prompting
Temperature changes variance, not intelligence. It widens the lottery; it does not improve the tickets.
Set temperature near zero for extraction and scoring, then raise it only when variety is the product.
Individual neurons are polysemantic, one neuron fires for 'the Golden Gate Bridge' and also 'feeling trapped' and also semicolons. Train a sparse autoencoder on the activations and you pull those mixed signals apart into cleaner, single-meaning features.
This is roughly how the 'Golden Gate Claude' demo worked.
#interpretability
Sparse features give you named-ish knobs because they separate mixed neuron behavior into cleaner directions.
Treat an SAE feature as a hypothesis, then test whether steering it changes the behavior you care about.
Half of weird model behavior is a tokenization story. ' apple' and 'apple' are different tokens. Numbers split in unintuitive ways. Pasting your prompt into a tokenizer viewer takes ten seconds and explains a surprising number of bugs.
#tooling
The model never sees your characters. It sees chunks, spaces, merges, and weird number splits.
Paste the exact prompt, including spaces and separators, into a tokenizer before debugging prompt behavior.
For reasoning tasks, generate the same answer five or ten times at moderate temperature and take the most common result. A single greedy answer is fragile; the consensus across samples is far more reliable. Slow, but it works.
#reliability
Wrong reasoning paths tend to scatter; correct reasoning paths often converge.
Sample several answers, normalize the final claim, and inspect disagreement before majority voting.
Models attend strongly to the beginning and end of a long prompt and get foggy in the middle. Put the instructions and the most important facts at the edges. If something keeps getting ignored, it's probably buried in the soft center.
Sometimes called 'lost in the middle'.
#prompting
A fact can be present and still practically invisible if it sits in the soft middle of a long prompt.
Move critical constraints to the opening contract and repeat the single highest-risk fact at the end.
Run the model twice, once on a clean prompt and once on a corrupted one, then copy a single activation from one run into the other. If the output flips, you just found a component that carries the behavior. It's causal, not just correlational.
The closest thing in this field to a controlled experiment.
#interpretability
Correlation says a signal is present. Patching asks whether changing that signal changes the outcome.
Patch one layer-position activation from a clean run into a corrupted run and track whether the answer flips.
In few-shot prompts the model copies the shape of your examples first: the punctuation, the casing, the length, the label words. Get the format dead consistent and the task often solves itself. Sloppy examples teach sloppiness.
#prompting
Few-shot examples teach surface form first. The model copies the pattern before it understands your intention.
Make labels, punctuation, casing, separators, and example length boringly consistent.
Take the hidden state from an intermediate layer and run it through the model's output head early. You get the model's 'best guess so far' at that depth. Watching the prediction sharpen layer by layer is like a time-lapse of a thought forming.
#interpretability
The answer often appears before the wording is polished. Middle layers can show the thought taking shape.
Decode intermediate layers and watch when the target answer first becomes linearly readable.