AI Insights for Humans

PlatPhormNews field notebook

AI insights for humans

A practical notebook for understanding model behavior: prompting habits, interpretability methods, verification rituals, and tools that make AI systems less mysterious.

Latest noteAPI docs
Aha lab

Watch the idea move before you read the rule.

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

The sentence is fluent. The token path may not be.

Read note

This toy trace turns a polished answer into token confidence. The cold tokens are where a reviewer should slow down first.

Aha

A low-margin token is where the model nearly chose a different world.

Thedecoded
paperdecoded
wasdecoded
publisheddecoded
indecoded
2024decoded
bydecoded
OpenAIdecoded

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.

Field guide

Turn insight into an operating habit.

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

Hallucination triage

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.

01Locate fragile spans
02Separate retrieval from generation
03Attach receipts
04Decide the confidence bucket
  1. 1

    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.

  2. 2

    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.

  3. 3

    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.

  4. 4

    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.

Read the logprobs, not the answerScore retrieved context before you trust the answerTurn every important claim into a receiptTriangulate citations instead of trusting the first source

Playbook

Prompt change control

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.

01Shrink the contract
02Diff the meaning
03Run the eval sandwich
04Mutate the boundary
  1. 1

    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.

  2. 2

    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.

  3. 3

    Run the eval sandwich

    Baseline before, targeted fixture during, broad regression after.

    Aha: The middle eval catches drift while the edit is still cheap.

  4. 4

    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.

Keep the system prompt short enough to auditDiff the meaning, not just the textUse the eval sandwich: before, during, afterRed-team prompts with harmless mutants

Playbook

Agent action gate

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.

01Preview the action
02Respect blocked states
03Spend checks by risk
04Verify after execution
  1. 1

    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.

  2. 2

    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.

  3. 3

    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.

  4. 4

    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.

Dry-run tool calls before letting them mutate anythingDesign abstention as a first-class outputSpend latency where uncertainty is highestRoute tasks by failure mode, not brand name

Failure cases worth recognizing

Case study

The source-shaped hallucination

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

The buried constraint

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

The tool call that did too much

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

The helpful prompt edit that moved the boundary

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

The corpus is small enough to read, structured enough to operate.

29

published notes

6

topic tags

29

aha moments

4

reading paths

Tip10
Method16
Tool3

Topic signal map

evals3
interpretability5
prompting7
reliability6
tooling3
verification5

Reading paths

Three ways through the notebook

Entries

The working notebook

29 of 29 notes visible

  1. Method2026 · 06 · 28

    01Score retrieved context before you trust the answer

    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

    Aha moment

    Retrieval only proves that documents were fetched. It does not prove the fetched text answers the question.

    Try this: Label each retrieved chunk as direct support, background, contradiction, or irrelevant before reading the final answer.

  2. Tip2026 · 06 · 24

    02Design the output schema before the prompt

    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

    Aha moment

    A schema turns a vague prompt into an interface contract with explicit absence states.

    Try this: Write the JSON fields, allowed values, evidence pointers, and unknown states before drafting the prompt.

  3. Method2026 · 06 · 20

    03Use the eval sandwich: before, during, after

    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

    Aha moment

    The most valuable eval is often the small one you run while the prompt is still being edited.

    Try this: Run a baseline set before the change, five targeted fixtures during the change, and a regression set after.

  4. Method2026 · 06 · 14

    04Dry-run tool calls before letting them mutate anything

    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

    Aha moment

    A tool call should become inspectable before it becomes irreversible.

    Try this: Require action preview, target ids, expected effects, refusal conditions, and rollback notes before execution.

  5. Method2026 · 06 · 08

    05Triangulate citations instead of trusting the first source

    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

    Aha moment

    A second independent source often reveals whether a citation supports the claim or merely resembles it.

    Try this: Triangulate dates, quotes, benchmarks, allegations, and high-impact facts before presenting them as settled.

  6. Method2026 · 05 · 31

    06Red-team prompts with harmless mutants

    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

    Aha moment

    A boundary that only holds for one phrasing is a memorized example, not a reliable rule.

    Try this: Generate harmless variants that preserve intent while changing tone, order, indirection, language, and role framing.

  7. Tool2026 · 05 · 18

    07Diff the meaning, not just the text

    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

    Aha moment

    The smallest text diff can be the biggest behavioral diff when it changes authority, scope, or obligation.

    Try this: Review prompt edits for new duties, removed constraints, softened verbs, and examples that imply new policy.

  8. Tip2026 · 05 · 09

    08Spend latency where uncertainty is highest

    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

    Aha moment

    Careful does not have to mean slow everywhere. Verification can wake up only where risk wakes it up.

    Try this: Escalate to extra checks when confidence is low, evidence is weak, impact is high, or an action is protected.

  9. Method2026 · 04 · 26

    09Route tasks by failure mode, not brand name

    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

    Aha moment

    Routing by failure mode is more useful than routing by reputation because every model breaks differently.

    Try this: Build a route table from observed task fixtures: extraction, synthesis, coding, refusal review, and long-context use.

  10. Tip2026 · 04 · 18

    10Treat memory as evidence with an expiration date

    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

    Aha moment

    Agent memory is a cache. Useful caches have scope, freshness, and invalidation rules.

    Try this: Store stable decisions and preferences, but re-check drift-prone facts before acting on them.

  11. Tip2026 · 04 · 03

    11Design abstention as a first-class output

    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

    Aha moment

    A model fills blanks when the product has no respected way to stop.

    Try this: Make unknown, partial, blocked, protected-required, and needs-review states valid outputs with next actions.

  12. Method2026 · 03 · 21

    12Use confidence buckets, not fake precision

    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

    Aha moment

    Confidence is useful when it changes what happens next, not when it looks mathematically precise.

    Try this: Use high, medium, low, and unknown buckets tied to evidence type, conflict level, and review action.

  13. Method2026 · 06 · 18

    13Ask for counter-evidence before confidence

    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

    Aha moment

    The model is usually best at self-correction before it has been rewarded for sounding certain.

    Try this: Ask for the answer, the strongest objection, what evidence would change the answer, and then confidence.

  14. Method2026 · 05 · 27

    14Turn every important claim into a receipt

    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

    Aha moment

    A citation is not decoration. It is a handle you can pull to see whether the sentence is load-bearing.

    Try this: Make every claim end with one receipt: source, row, command, calculation, trace id, or explicit assumption.

  15. Tip2026 · 05 · 03

    15Keep the system prompt short enough to audit

    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

    Aha moment

    A system prompt is product code. If it is too long to review, it is too long to trust.

    Try this: Rewrite the system prompt into role, hard boundary, output contract, and escalation rule.

  16. Method2026 · 04 · 12

    16Use refusal boundaries as test fixtures

    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

    Aha moment

    Safety quality is not whether the model refuses. It is whether it refuses at the right boundary.

    Try this: Keep allowed, disallowed, and allowed-with-constraints prompts as fixtures and run them after prompt edits.

  17. Method2025 · 11 · 02

    17Read the logprobs, not the answer

    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

    Aha moment

    The final text is only the winning token path. The uncertainty is still visible in the runners-up.

    Try this: Render low-margin tokens first, especially names, dates, citations, and first tokens after a policy boundary.

  18. Tip2025 · 10 · 28

    18Make it think before it answers

    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

    Aha moment

    Explanation after an answer often explains a commitment; reasoning before an answer can still change it.

    Try this: Put scratch work, constraints, and uncertainty checks before the final answer field.

  19. Tool2025 · 10 · 19

    19BertViz for watching attention move

    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

    Aha moment

    Attention is not the whole mind, but it is a useful map of what information is being routed where.

    Try this: Inspect heads for repeated roles: bracket matching, quote tracking, subject agreement, or copied labels.

  20. Tip2025 · 10 · 11

    20Tell it what NOT to do, sparingly

    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

    Aha moment

    Negation still puts the forbidden idea into context. Attention does not forget because grammar says not to.

    Try this: Rewrite prohibitions as the positive shape of the output you want before adding hard safety rails.

  21. Method2025 · 09 · 30

    21Probe for a concept with a linear classifier

    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

    Aha moment

    If a tiny linear classifier can read a concept from hidden states, the model has already organized that signal.

    Try this: Probe one concept across layers and look for where accuracy first rises above a simple baseline.

  22. Tip2025 · 09 · 22

    22Temperature is a creativity dial, not a quality dial

    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

    Aha moment

    Temperature changes variance, not intelligence. It widens the lottery; it does not improve the tickets.

    Try this: Set temperature near zero for extraction and scoring, then raise it only when variety is the product.

  23. Method2025 · 09 · 14

    23Sparse autoencoders to untangle neurons

    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

    Aha moment

    Sparse features give you named-ish knobs because they separate mixed neuron behavior into cleaner directions.

    Try this: Treat an SAE feature as a hypothesis, then test whether steering it changes the behavior you care about.

  24. Tool2025 · 09 · 03

    24Keep a tokenizer open in a tab

    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

    Aha moment

    The model never sees your characters. It sees chunks, spaces, merges, and weird number splits.

    Try this: Paste the exact prompt, including spaces and separators, into a tokenizer before debugging prompt behavior.

  25. Method2025 · 08 · 25

    25Sample many, keep the majority

    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

    Aha moment

    Wrong reasoning paths tend to scatter; correct reasoning paths often converge.

    Try this: Sample several answers, normalize the final claim, and inspect disagreement before majority voting.

  26. Tip2025 · 08 · 12

    26Mind the middle of long contexts

    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

    Aha moment

    A fact can be present and still practically invisible if it sits in the soft middle of a long prompt.

    Try this: Move critical constraints to the opening contract and repeat the single highest-risk fact at the end.

  27. Method2025 · 07 · 30

    27Activation patching to find the cause

    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

    Aha moment

    Correlation says a signal is present. Patching asks whether changing that signal changes the outcome.

    Try this: Patch one layer-position activation from a clean run into a corrupted run and track whether the answer flips.

  28. Tip2025 · 07 · 18

    28Your examples teach format louder than content

    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

    Aha moment

    Few-shot examples teach surface form first. The model copies the pattern before it understands your intention.

    Try this: Make labels, punctuation, casing, separators, and example length boringly consistent.

  29. Method2025 · 07 · 05

    29The logit lens: decode the middle layers

    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

    Aha moment

    The answer often appears before the wording is polished. Middle layers can show the thought taking shape.

    Try this: Decode intermediate layers and watch when the target answer first becomes linearly readable.

Platform contract

Public-readable, machine-readable, operator-safe.

100%

required routes implemented

public

entries, health, feeds, docs, MCP introspection

keyed

report generation requires PLATPHORM_API_KEY

HealthOpenAPILLMs fullTrust