# AI Insights for Humans

A working notebook of practical AI interpretability, prompting, reliability, and tooling notes for people who want evidence instead of hype.

Purpose: Practical AI insight notebook for human operators.
Entries: 29
Tags: evals, interpretability, prompting, reliability, tooling, verification

Public surfaces:
- HTML notebook: https://ai-insights-for-humans.platphormnews.com/
- RSS: https://ai-insights-for-humans.platphormnews.com/rss.xml
- API docs: https://ai-insights-for-humans.platphormnews.com/api/docs
- OpenAPI: https://ai-insights-for-humans.platphormnews.com/openapi.yaml
- MCP: https://ai-insights-for-humans.platphormnews.com/api/mcp

Protected actions require PLATPHORM_API_KEY.

## Route Standard
Implemented required routes: 20/20
- GET /: implemented; Public AI insights notebook and reading paths.
- GET /api/health: implemented; Platform health payload for the publication.
- GET /api/v1/health: implemented; Versioned platform health payload.
- GET /api/docs: implemented; Machine-readable API and route documentation.
- GET /api/v1/entries: implemented; Read-only JSON index of notebook entries.
- GET /api/v1/entries/{id}: implemented; Read-only JSON detail for one notebook entry.
- POST /api/v1/reports: implemented; Protected generation of a real notebook/platform summary report.
- GET /api/mcp: implemented; MCP metadata and JSON-RPC usage information.
- POST /api/mcp: implemented; JSON-RPC 2.0 MCP endpoint for real notebook tools, resources, and prompts.
- GET /openapi.yaml: implemented; OpenAPI 3.1 contract for public and protected routes.
- GET /llms.txt: implemented; Concise LLM-readable site summary.
- GET /llms-full.txt: implemented; Complete LLM-readable notebook and platform summary.
- GET /llms-index.json: implemented; Structured LLM index generated from real entries.
- GET /robots.txt: implemented; Crawler policy and sitemap pointer.
- GET /sitemap.xml: implemented; Canonical URL sitemap for the notebook.
- GET /sitemap-index.xml: implemented; Sitemap index for platform crawlers.
- GET /rss.xml: implemented; RSS feed generated from notebook entries.
- GET /feed.xml: implemented; RSS-compatible feed alias.
- GET /atom.xml: implemented; Atom feed generated from notebook entries.
- GET /manifest.webmanifest: implemented; Web app manifest for install and preview surfaces.
- GET /.well-known/mcp.json: implemented; MCP capability discovery manifest.
- GET /.well-known/agents.json: implemented; Agent-facing route and policy discovery manifest.
- GET /.well-known/security.txt: implemented; Security contact and policy file.
- GET /.well-known/trust.json: implemented; Trust, auth, public/protected, and exposure policy.

## Trust Policy
Web dashboard, public-safe discovery, browser-based operations, trusted-domain discovery, standard route compliance, Vercel metadata capture, trace inspection, and agentic workflow discovery are intentionally supported for public read-only debugging and operator workflows. Mutating, administrative, ingestion, replay, fork, remediation, deployment, sync, test-triggering, reporting, and write actions require PLATPHORM_API_KEY.

## Field Guide Playbooks
### Hallucination triage
A model produced a confident answer and nobody knows whether the weak point is retrieval, reasoning, or source support.
Outcome: A reviewer can identify which claims are supported, which claims need receipts, and which answer parts should be withheld.
1. Locate fragile spans: Hallucinations often hide where fluency is highest but token certainty is thinnest.
2. Separate retrieval from generation: Bad context can make a good model sound wrong with confidence.
3. Attach receipts: The unsupported claim does not vanish; it becomes visible enough to reject.
4. Decide the confidence bucket: Confidence matters when it changes the next action.

### Prompt change control
A system prompt needs improvement, but every wording edit risks changing refusal, format, or evidence behavior.
Outcome: The prompt change ships with a semantic diff, targeted fixtures, and a clear before/during/after eval record.
1. Shrink the contract: A smaller system prompt is not weaker; it is easier to audit.
2. Diff the meaning: A one-word modal change can quietly become a policy change.
3. Run the eval sandwich: The middle eval catches drift while the edit is still cheap.
4. Mutate the boundary: A boundary that only catches the obvious phrasing is not a boundary.

### Agent action gate
An agent can call tools, but the operator needs confidence before it mutates, publishes, replays, deletes, or spends.
Outcome: Every action moves through preview, inspect, authorize, execute, and verify without hiding protected states.
1. Preview the action: Permission to plan is not permission to execute; the preview is where unsafe scope becomes inspectable.
2. Respect blocked states: A designed stop is better than a graceful guess because it preserves trust and gives the operator a next move.
3. Spend checks by risk: Latency can be evidence-driven instead of uniformly expensive.
4. Verify after execution: The world, not the model, gets the final vote, so post-action evidence is part of the action itself.

## Failure Case Studies
### 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.

### 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.

### 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.

### 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.

## AHA Ladder
### From answer to signal
Before: Read the model's sentence and judge whether it sounds plausible.
After: Inspect token margins, retrieved evidence, receipts, and contradiction before judging.
Operator question: Where would this answer break first?

### From prompt to contract
Before: Write instructions until the behavior feels described.
After: Define the output schema, failure states, hard boundary, and test fixtures.
Operator question: What is the model allowed to leave blank?

### From tool to action
Before: Let the agent call a tool when it appears to understand the task.
After: Preview arguments, inspect target ids, authorize mutation, execute, then verify changed state.
Operator question: What would stop this action before it runs?

### From confidence to decision
Before: Ask the model for a percentage and decide whether it feels high enough.
After: Bucket confidence by evidence quality, conflict level, and downstream impact.
Operator question: What does this confidence label change?

## Entries
# Score retrieved context before you trust the answer
Kind: Method
Tag: verification
Published: 2026-06-28
RAG failures often look like model failures because the model speaks with the confidence of the context it was given. The first move is to separate retrieval quality from generation quality.
Give each chunk three labels: directly answers, adjacent background, or contradiction. The answer should cite directly answering chunks first and name contradictions instead of smoothing them away.
A good retrieval review is boring and mechanical. It asks whether the evidence was even in the room before debating whether the model reasoned over it correctly.
## 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.

## Watch For
- Answers citing adjacent background as if it were direct support
- Contradictory chunks hidden by smooth synthesis
- High source count used as a substitute for relevance
Note: Retrieved is not the same thing as relevant.

# Design the output schema before the prompt
Kind: Tip
Tag: reliability
Published: 2026-06-24
A schema makes ambiguity visible. If you cannot name the field, type, and allowed absence state, you probably do not know what you are asking the model to produce.
Start with the downstream consumer. Does it need a label, a span, a confidence bucket, a source pointer, or an abstention? Write that contract first and then prompt the model to fill it.
The aha is that schemas are not just for machines. They force human reviewers to decide what counts as complete, partial, unknown, and unsafe.
## 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.

## Watch For
- Required fields that force the model to invent missing data
- Enums that do not include blocked, unknown, or partial
- Fields whose consumer or decision impact is unclear

# Use the eval sandwich: before, during, after
Kind: Method
Tag: evals
Published: 2026-06-20
Before-change evals establish the baseline you are trying not to lose. During-change evals are small and sharp: the five examples most likely to break because of the edit. After-change evals make sure the fix did not damage neighboring behavior.
This is especially useful for system prompt work because every instruction edit has a shadow. You improve one behavior and accidentally move refusal, tone, verbosity, or citation discipline.
Treat the eval sandwich as a change-management ritual. It is less glamorous than a benchmark, but it is the thing that keeps production behavior from drifting by inches.
## 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.

## Watch For
- Only testing the behavior you intended to improve
- Skipping refusal and edge fixtures after style changes
- Treating one green benchmark as proof of production safety
Note: The middle eval is where most prompt regressions get caught.

# Dry-run tool calls before letting them mutate anything
Kind: Method
Tag: tooling
Published: 2026-06-14
Tool use changes the risk profile of a model. The model can be wrong in the world, not just wrong in text. The safe pattern is preview, inspect, authorize, execute, verify.
The dry run should include exactly what will be called, what identifiers will be touched, which fields will change, and what success will look like. If the model cannot state those plainly, it should not run the tool.
The best dry runs also include a refusal branch. They say what condition would stop execution and what evidence would be needed to proceed.
## 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.

## Watch For
- Tool arguments that include broad or ambiguous selectors
- Mutation plans without a verification step
- The model treating permission to plan as permission to execute

# Triangulate citations instead of trusting the first source
Kind: Method
Tag: verification
Published: 2026-06-08
Citation checking should ask three questions: does the source exist, does it support the exact claim, and would another independent source phrase the same claim the same way?
When sources disagree, the disagreement is the content. Do not ask the model to hide the conflict behind a confident synthesis. Ask it to name the split and explain what would settle it.
Triangulation is not needed for every sentence. Reserve it for claims that drive decisions: dates, quotes, prices, medical or legal facts, benchmarks, and allegations.
## 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.

## Watch For
- Citations that support a nearby but different claim
- Two sources that repeat the same upstream source
- Disagreements collapsed into a false consensus

# Red-team prompts with harmless mutants
Kind: Method
Tag: evals
Published: 2026-05-31
Prompt boundaries are brittle when tests are too literal. A model may refuse the obvious bad phrasing and comply with the polite version, the translated version, or the version hidden inside a roleplay.
Mutants make that brittleness visible without adding new dangerous content. They perturb the wrapper, not the underlying safety category.
Use them for both sides of the boundary: prompts that should be refused and prompts that should remain allowed. Over-refusal is a product bug too.
## 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.

## Watch For
- Safety behavior that fails when the user sounds polite
- Allowed prompts that become over-refused after hardening
- Fixtures that mutate the danger instead of just the wrapper

# Diff the meaning, not just the text
Kind: Tool
Tag: prompting
Published: 2026-05-18
Text diff says what characters changed. Semantic diff says what the model is now being asked to believe, prioritize, refuse, or produce. That is the diff that matters for behavior.
Look for modal verbs, scope words, exception clauses, examples that imply new rules, and output format changes. These are the places where small edits become big runtime differences.
A useful semantic diff ends with a test plan: which fixtures should change, which must not change, and which new example proves the edit worked.
## 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.

## Watch For
- May, should, and must changes hidden in wording cleanup
- Examples that contradict the written rule
- Output contracts changed without matching tests
Note: A one-word modal change can be a policy change.

# Spend latency where uncertainty is highest
Kind: Tip
Tag: reliability
Published: 2026-05-09
Reliable AI systems are often cascades. Cheap checks run first; expensive checks wake up only when the situation deserves them. This keeps the product fast without pretending every answer has equal risk.
A practical budget has triggers: low logprob margin, missing receipt, protected action, high-dollar decision, policy boundary, or user-visible publication.
The aha is that latency can be evidence-driven. You do not need to choose between always fast and always careful.
## 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.

## Watch For
- Uniformly expensive checks on low-risk tasks
- Fast paths that ignore missing evidence
- Latency budgets without explicit escalation triggers

# Route tasks by failure mode, not brand name
Kind: Method
Tag: reliability
Published: 2026-04-26
Model routing is usually framed as cost or quality. The more useful frame is failure. Which model omits caveats, over-refuses, invents sources, follows schemas, or handles long contexts under pressure?
Build small route tests around the jobs you actually run. A model that wins general reasoning may still be the wrong choice for structured extraction if it casually fills unknown fields.
Keep routing reversible. When model behavior shifts, the route table should be an operational setting backed by evidence, not a belief system.
## 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.

## Watch For
- Leaderboards used as a substitute for local fixtures
- Model routes that cannot be changed without code edits
- Ignoring behavior drift after model updates

# Treat memory as evidence with an expiration date
Kind: Tip
Tag: verification
Published: 2026-04-18
Good memory is scoped. It records what the user decided, what a repo convention is, and which risks were already found. Bad memory stores private residue, stale status, or facts that should be rechecked.
Every memory-derived claim should answer two questions: when was this learned, and how likely is it to have changed? If the answer is 'recently unstable,' verify before acting.
The aha is that memory is not context magic. It is a cache, and caches need invalidation rules.
## 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.

## Watch For
- Old status presented as current proof
- Private residue stored when only workflow shape was needed
- Memory outranking live repo, source, or user evidence

# Design abstention as a first-class output
Kind: Tip
Tag: reliability
Published: 2026-04-03
Abstention fails when it feels like an error. It works when it is a designed branch with a reason, missing evidence, and next action. The user should know what changed from answerable to not answerable.
Use explicit output states: answered, partial, unknown, blocked, protected-required, and needs-review. Each state should carry what evidence exists and what evidence is missing.
This is how you prevent graceful prose from hiding a dead end. The model can stop without leaving the user stuck.
## 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.

## Watch For
- Schemas where every field must be filled even when evidence is absent
- Refusals that leave the user without a useful next step
- Unknown states treated as errors instead of designed outcomes

# Use confidence buckets, not fake precision
Kind: Method
Tag: verification
Published: 2026-03-21
Precise confidence numbers look scientific and often are not. Unless they are tied to a calibrated model or measured historical error, they are mostly theater.
Buckets work better in operator workflows because they can map to action: high can publish, medium can request a receipt, low can escalate, unknown can abstain.
The bucket should cite the reason: direct source found, conflicting source found, no source found, or model-only inference. That reason matters more than the label.
## 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.

## Watch For
- Percentages that are not calibrated against historical outcomes
- Confidence labels without evidence reasons
- High confidence on claims with no direct receipt

# Ask for counter-evidence before confidence
Kind: Method
Tag: reliability
Published: 2026-06-18
The useful move is not asking the model to be uncertain. The useful move is forcing it to search for the best case against itself while the answer is still provisional.
A simple pattern: draft the answer, list disconfirming evidence, name what would change the answer, then assign confidence. The confidence number becomes attached to evidence quality instead of verbal smoothness.
This is especially helpful for research summaries, legal-ish interpretation, data diagnosis, and anything where one missing constraint can flip the conclusion.
## 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.

## Watch For
- Objections that restate the answer instead of challenging it
- Confidence that stays high after missing evidence appears
- A final answer that ignores its own counter-evidence
Note: Confidence after counter-evidence is more useful than confidence after a sales pitch.

# Turn every important claim into a receipt
Kind: Method
Tag: verification
Published: 2026-05-27
Models are good at moving from evidence to prose. They are also good at producing prose that looks like it came from evidence. The receipt requirement separates those two behaviors.
The receipt can be a citation, a row count, a command output line, a test name, or a declared assumption. The format matters less than the discipline: every claim has to point somewhere.
When the model cannot produce a receipt, that is not failure. That is the system doing its job and showing you where judgment is still needed.
## 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.

## Watch For
- Receipts that point to a broad document instead of the claim
- Numbers without denominators
- Source links that support nearby context but not the sentence itself

# Keep the system prompt short enough to audit
Kind: Tip
Tag: prompting
Published: 2026-05-03
System prompts work best as durable operating constraints, not as a dumping ground for every possible preference. When the prompt grows without structure, the model starts resolving conflicts in ways you cannot predict.
Split the prompt into policy, role, output contract, and refusal boundary. Then write a small regression set that exercises each section. You will catch more problems with five targeted tests than with another page of instructions.
## 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.

## Watch For
- Repeated style instructions that fight each other
- Hidden policy changes bundled into wording edits
- Task detail that belongs in user input rather than permanent rules
Note: If you cannot diff it cleanly, you cannot debug it cleanly.

# Use refusal boundaries as test fixtures
Kind: Method
Tag: evals
Published: 2026-04-12
Most prompt tests focus on whether the model can complete the intended workflow. That misses the other half of the product: whether it handles edge requests without over-refusing, leaking private context, or inventing authority it does not have.
Build a small fixture set with three buckets: allowed, disallowed, and allowed-with-constraints. A good model response should land in the right bucket and explain the usable next step.
This turns safety from a vibe into a regression target. When a prompt edit changes the refusal boundary, you see it immediately.
## 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.

## Watch For
- Over-refusal on harmless transformed versions
- Under-refusal when the same intent is phrased politely
- Responses that refuse but fail to offer a safe next step

# Read the logprobs, not the answer
Kind: Method
Tag: interpretability
Published: 2025-11-02
Every token a model emits comes with a probability distribution over the whole vocabulary. The decoded text only shows you the winner. The logprobs show you the race.
When you pull them back, scan for the spots where the top token barely beat the runner-up. Those are the seams. Names, dates, citations, and the first token of a refusal are where you most often find a coin-flip dressed up as a confident sentence.
A practical loop: request the top 5 logprobs per token, render them as a heatmap over the output, and read the cold spots first. That is usually where the hallucination lives.
## 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.

## Watch For
- Fluent spans built from low-confidence token choices
- Nearly tied alternatives around entities
- A confident tone hiding a fragile probability margin
Note: Confidence ≠ correctness. But low confidence is a great place to look for hallucinations.

# Make it think before it answers
Kind: Tip
Tag: prompting
Published: 2025-10-28
Autoregressive models can only condition on what they have already written. If the answer comes first, the explanation is generated to fit a token that is already locked in. It is rationalization, not reasoning.
Flip the order and the reasoning tokens become real scratch space the final answer can lean on. This is the entire mechanism behind chain-of-thought, and it is why 'answer, then explain' quietly underperforms.
## 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.

## Watch For
- Post-hoc explanations that never revise the answer
- A final answer that appears before evidence is considered
- Reasoning sections that only paraphrase the conclusion

# BertViz for watching attention move
Kind: Tool
Tag: tooling
Published: 2025-10-19
BertViz draws the attention pattern as lines connecting tokens, split by head and layer. It runs in a notebook in a couple of lines and works on most Hugging Face transformer models.
The value is not the pretty picture, it is the habit of looking. Once you have seen a head that exclusively tracks matching brackets, you stop thinking of the model as a black box and start thinking of it as a stack of small, legible mechanisms.
## 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.

## Watch For
- Pretty attention maps with no behavioral test
- One dramatic head treated as the whole explanation
- Patterns that disappear after small prompt perturbations
Note: Most heads are boring. The interesting 5% are worth the dig.

# Tell it what NOT to do, sparingly
Kind: Tip
Tag: prompting
Published: 2025-10-11
Putting a word in the prompt raises its activation, even when the word is wrapped in 'do not'. The model does not have a clean negation operator the way code does; it has attention, and attention does not know the difference between mention and command.
Rewrite prohibitions as positive descriptions of the target. Instead of 'do not be formal', say 'write like you are texting a friend'. Save the hard negatives for the few rails you cannot express any other way.
## 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.

## Watch For
- Forbidden words repeated in the response
- Style instructions expressed mostly as negatives
- A model fixating on the thing you tried to suppress

# Probe for a concept with a linear classifier
Kind: Method
Tag: interpretability
Published: 2025-09-30
The setup is small: collect hidden states from a chosen layer for a labeled dataset, then fit logistic regression on top. High accuracy means the concept is linearly available at that layer.
Sweep the probe across layers and you can watch where a concept first becomes readable. Abstract properties tend to sharpen in the middle layers, which is a clue about where the model is doing the relevant work.
## 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.

## Watch For
- High probe accuracy mistaken for causal importance
- Labels that leak through surface wording
- Layer sweeps without a held-out validation split
Note: Cheap, fast, and weirdly revealing.

# Temperature is a creativity dial, not a quality dial
Kind: Tip
Tag: prompting
Published: 2025-09-22
Temperature reshapes the probability distribution before sampling. Low temperature sharpens it toward the most likely token; high temperature flattens it so unlikely tokens get a real chance.
So the dial controls variance, not intelligence. If a task has a correct answer, variance is your enemy and you want it near zero. If you want range and surprise, turn it up and sample several times.
## 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.

## Watch For
- Creative settings used for factual extraction
- One sampled answer treated as a stable result
- Quality claims that ignore variance between runs

# Sparse autoencoders to untangle neurons
Kind: Method
Tag: interpretability
Published: 2025-09-14
The trick is over-completeness plus sparsity. You map activations into a much wider space but force only a few of those dimensions to be active at once. The pressure pushes tangled concepts apart into separate, more interpretable features.
Once you have the features you can do surgery: amplify one and watch the model fixate on a bridge, suppress another and watch a behavior fade. It is the closest thing we have to named knobs inside the network.
## 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.

## Watch For
- Human-readable labels that overfit a few examples
- Feature steering without downstream behavior checks
- Polysemantic residue hiding inside a cleaner-looking feature
Note: This is roughly how the 'Golden Gate Claude' demo worked.

# Keep a tokenizer open in a tab
Kind: Tool
Tag: tooling
Published: 2025-09-03
Leading spaces, casing, and digit grouping all change how text is split, and the model only ever sees the tokens, never your characters. A trailing space can silently sabotage a few-shot format.
When something is off and you cannot explain it, paste the exact string into a tokenizer view before you theorize. The answer is in there more often than you would expect.
## 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.

## Watch For
- Leading spaces changing label tokens
- Numbers split into surprising chunks
- Few-shot examples with inconsistent separators

# Sample many, keep the majority
Kind: Method
Tag: reliability
Published: 2025-08-25
Different samples take different reasoning paths but tend to converge on the same correct answer, while the wrong answers scatter. Voting over the final answers turns that asymmetry into accuracy.
It costs you N times the compute, so reserve it for the questions that matter. A useful middle ground is to escalate to self-consistency only when the first answer looks low-confidence.
## 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.

## Watch For
- Repeated wording counted as independent evidence
- Voting over outputs that all share the same missing assumption
- Compute spent on low-value questions

# Mind the middle of long contexts
Kind: Tip
Tag: prompting
Published: 2025-08-12
Retrieval accuracy as a function of position tends to be U-shaped: strong at the start, strong at the end, sagging in the middle. A fact you need can be technically present and still effectively invisible.
Design around it. Lead with the instruction, restate the critical constraint at the very end, and keep the soft middle for material you can afford to lose.
## 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.

## Watch For
- Middle facts ignored while opening and closing instructions are followed
- Retrieved context pasted without ranking
- Important constraints surrounded by low-value filler
Note: Sometimes called 'lost in the middle'.

# Activation patching to find the cause
Kind: Method
Tag: interpretability
Published: 2025-07-30
Probes tell you what is represented. Patching tells you what matters. By swapping one activation and watching the effect, you move from 'this information is present' to 'this component causes the behavior'.
Sweep the patch across layers and positions to localize the circuit. The spots where patching flips the answer are the load-bearing ones; everything else is along for the ride.
## 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.

## Watch For
- Large sweeps without a clean causal hypothesis
- Output flips caused by formatting rather than the target behavior
- A single patch interpreted without replication
Note: The closest thing in this field to a controlled experiment.

# Your examples teach format louder than content
Kind: Tip
Tag: prompting
Published: 2025-07-18
Few-shot learning is pattern continuation. The model is extremely sensitive to surface form, sometimes more than to the actual reasoning you intended to demonstrate.
Make every example identical in structure, pick label words that are single clean tokens, and keep lengths similar. The more rigid the template, the more reliably the model fills in the next blank.
## 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.

## Watch For
- One example with a different label shape
- Multi-token labels competing with single-token labels
- Examples that demonstrate style drift instead of the task

# The logit lens: decode the middle layers
Kind: Method
Tag: interpretability
Published: 2025-07-05
The output projection is just a linear map, so nothing stops you from applying it to an earlier layer. The result is a rough readout of what the model would say if it had to stop thinking right now.
Often the correct answer appears in the middle layers and the later layers spend their budget refining phrasing and suppressing alternatives. Seeing that handoff demystifies a lot of model behavior.
## 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.

## Watch For
- Early guesses treated as final causal proof
- Layer readouts that shift after prompt paraphrases
- Ignoring later layers that suppress unsafe or irrelevant continuations