Discussions¶

Nine categories, eight title prefixes, one rule underneath all of them:

Evidence beats adjectives. A post with a number in it gets answered in one round trip. A post without one takes three, and usually the first two are people asking what you measured.

Open Discussions →

Where does this go?¶

Start at the top and stop at the first row that matches.

I want to…	Post in	Prefix
ask about a specific exercise I am stuck on	Q&A	`[clinic · EX-NN]`
ask about a specific lab	Q&A	`[lab · TN]`
say the material is wrong or overconfident	Q&A	`[errata]`
follow a derivation I cannot get through	Q&A	`[maths]`
ask anything else that has an answer	Q&A	—
submit a finished exercise, with numbers	Show and Tell	`[solution · EX-NN]`
submit a lab and have it run in a sandbox	Show and Tell	`[submit · LNN]`
do a 5–15 minute challenge and have it run in a sandbox	Show and Tell	`[arena · CNN]`
report something I measured and rejected	Show and Tell	`[negative result]`
show a finished capstone	Show and Tell	`[capstone]`
get a design torn apart before I build it	Design Reviews	—
practise an interview answer	Interview Prep	—
run or read a full simulated interview loop	Interview Prep	`[round · shape]`
argue about a paper	Reading Club	—
propose something that does not exist yet	Ideas	—
ask people to predict before a reveal	Polls	`[poll]`
announce a release, a cohort, an office hours	Announcements	`[office hours · date]`
introduce myself, ask about tooling or careers	General	—

When two rows fit, take the higher one. A question about a paper you are stuck on is Q&A; Reading Club is for arguing about a paper you have read.

Why prefixes exist at all¶

GitHub has no API for creating discussion categories — there is no category mutation in the GraphQL schema, and creating one is a manual settings action. This repository has nine.

Rather than leave the structure unbuilt, some kinds of thread carry a title prefix and a matching label, which together do what a tenth category would have done: make a set findable, filterable, and distinguishable at a glance.

Prefix	Category	Label	Filter
`[clinic · EX-NN]`	Q&A	`clinic`	→
`[lab · TN]`	Q&A	`lab`	→
`[maths]`	Q&A	`maths`	→
`[errata]`	Q&A	`errata`	→
`[solution · EX-NN]`	Show and Tell	`solution`	→
`[submit · LNN]`	Show and Tell	`lab`	→
`[arena · CNN]`	Show and Tell	`arena`	→
`[negative result]`	Show and Tell	`negative-result`	→
`[capstone]`	Show and Tell	`capstone`	→
`[round · shape]`	Interview Prep	`interview-round`	→
`[office hours · date]`	Announcements	`office-hours`	→
`[poll]`	Polls	`poll`	→

The label is the reliable filter. A title can be edited by accident; a label cannot.

A prefix earns its place only when it distinguishes threads within a category. "Design review:" inside Design Reviews is noise. [negative result] inside Show and Tell is not, because that category also holds capstones and solutions.

The nine categories¶

Q&A · answerable¶

The workhorse, and the only category where the Mark as answer mechanic applies.

Belongs here: anything with a determinate answer. Stuck on an exercise or lab, a derivation you cannot follow, a number that disagrees with the docs, a mechanism you cannot explain.

Does not belong here: opinions, proposals, and "what do people think about X" — those are General or Ideas. A question with no possible answer clutters the answered/unanswered filter, which is the only triage signal this category has.

A good post carries: the numbers you got, the config that produced them, and what you already tried. See asking well below.

The clinic threads are pinned per exercise. Ask there rather than opening a new thread — the answers accumulate where the next person will look. That is the difference between a support queue and a knowledge base.

Show and Tell · open¶

Finished work, with numbers. Three shapes, distinguished by prefix.

[solution · EX-NN] — an exercise submission. Must include your numbers, your interval, and what you tried that did not work. The last is not decoration: a submission that reports only the winning path teaches nobody the search.

[negative result] — something you measured and rejected. Full credit, and often the most useful thing in the category. State the mechanism you tested, not the technique you were inspired by: "contextual chunking does not work" is unfalsifiable; "prepending already-present metadata to chunks that already carry a heading path did not improve recall on this corpus, at 2.4× storage" is a claim someone can check and reuse.

[capstone] — an end-to-end build with a decision record.

[submit · LNN] and [arena · CNN] — a lab or a challenge solution. One thread per item, and a workflow runs your code and replies within about a minute: which checks passed and which did not with the check's own message, what you can now do, one link to go deeper, and the next item with the reason it comes next.

It is safe to post code in a public thread because of how the run is built, not because anyone trusts the poster. Three jobs, three permission sets: the one that runs your code holds no credentials at all and runs it in a container with no network, a read-only filesystem, no capabilities, an unprivileged user and a 30-second kill; the job that posts the reply never sees your code. Your comment never reaches a shell command line either. A hostile submission finds nothing to reach. For a predict challenge, post the bare word or number — no code block needed.

Reviewers reply with a measurement, not an opinion. "I'd have done it differently" is not a review. "Here is the number that would change your conclusion" is.

Design Reviews · answerable¶

Constraints in, design out, critique in the middle. An RFC that has not been built yet.

A good post carries: the constraints that make it hard — latency, cost ceiling, residency, ACL cardinality, corpus size, team size, timeline — before the design. A design posted without its constraints cannot be reviewed, only admired.

Reviewers: run the design review checklist. Say explicitly which parts survive review; a review that lists only problems reads as a verdict on the person rather than on the design.

The answer gets marked when the thread reaches a decision, and the marked answer should contain the synthesis — what was accepted, what was rejected, and what is still open.

Interview Prep · answerable¶

Two shapes.

Answer critique — post an answer you want torn apart. A self-critique is mandatory. An answer posted without one gets the critique you already knew about, which wastes everyone's turn.

[round · shape] — a full simulated loop: the prompt as it would be given, the clarifying questions that are scored, the trap, a model answer, and a rubric table. Not a single question. These are written to be read by someone preparing tomorrow.

Reading Club · answerable¶

One paper per thread. The house position:

Replicate, do not cite. A paper we can re-run against this corpus is worth far more argued about with numbers than without.

A good post carries: the claim in one paragraph, how we would test it here (which notebook, which cell, which metric), and two or three things to argue about. If it cannot be tested against this corpus, say so — that is a finding about the paper's scope, not a reason to skip the thread.

Ideas · open¶

Pre-issue. An idea that survives the thread becomes an issue with acceptance criteria; one that does not is still a useful record of something considered and rejected.

A good post carries: the problem first, not the technique. Ideas that begin with a technique usually do not survive contact with The Precondition Test. Then: which metric moves, and by how much would it have to move to be worth the cost.

Polls · poll format¶

Calibration. Ask people to predict before the reveal — the gap between the vote and the measurement is the teaching content, and it evaporates if you post the answer first.

The reveal protocol, because a poll with no reveal is a survey:

Post the scenario with the numbers people need to decide, and no more.
Leave it open at least a week.
Post the measurement as a reply, not by editing the body — editing destroys the record of what people were asked.
Say what the distribution of votes tells you, including when it says nothing.

Announcements · announcement format¶

Releases, cohort kickoffs, errata roll-ups, and office hours.

[office hours · YYYY-MM-DD] — one dated thread per session. Agenda before, notes after, and the thread stays open, so the answers are findable by whoever hits the same thing in three months.

An announcement that implies an action must name it. If the answer is "nothing to do", say nothing to do.

General · open¶

Everything that does not fit and does not have an answer. Introductions, tooling, careers, opinions, war stories, link round-ups.

This is not a dumping ground. If your post has an answer, it belongs in Q&A, where the answered filter can find it later.

Asking a question that gets answered once¶

The three things that turn three round trips into one:

1. THE NUMBERS you got        not "recall got worse" — 0.807 → 0.752
2. THE CONFIG that produced   k, N, alpha, chunking strategy, slice, encoder
   them                       print RagConfig next to the metric, always
3. WHAT YOU ALREADY TRIED     including the thing that seemed obvious and failed

And the thing most people omit: what you expected. "I expected widening N to move full-chain recall and it did not" tells a reader which mental model to correct. "Full-chain recall did not move" tells them a fact they now have to guess the significance of.

Before posting, check the clinic thread for your exercise. Roughly half of new Q&A threads are already answered in one.

Answering well¶

Answer the question that was asked, then the better question underneath it, in that order. Reversing them reads as a lecture.
Bring a measurement. "I have seen this fail" beats "I would not do that", and both beat seniority.
Say what is still open. An answer that resolves everything is usually hiding something.
Be willing to be wrong in public. The threads worth reading here are the ones where somebody changed their mind on evidence.

Thread lifecycle¶

flowchart LR
    A["opened"] --> B{"answerable<br/>category?"}
    B -->|yes| C["answered<br/><i>marked</i>"]
    B -->|no| D["runs its course"]
    C --> E["harvested weekly into<br/>90-reference/faq.md"]
    C --> F{"needs code?"}
    D --> F
    F -->|yes| G["issue with<br/>acceptance criteria"]
    G --> H["PR with a<br/>measurement"]
    H --> I["thread updated<br/>with what shipped"]

Answered threads are harvested weekly into 90-reference/faq.md by the FAQ workflow. It links rather than copies: the thread stays the source of truth, because it carries the argument that got there, which is usually the more useful half.

Closing the loop is not optional. If a thread produces an issue, say so in the thread with the issue number. If the issue produces a PR, say what shipped. A thread that quietly stops is indistinguishable from one nobody read.

Duplicates, supersession and disagreement¶

Duplicates are not deleted. They are linked to the canonical thread and closed with a comment naming it — someone searched those words once, and they will again.

Superseded threads get a comment at the top of the answer explaining what changed and linking forward. A thread whose answer is now wrong and says nothing is worse than no thread.

Disagreement resolves on evidence, never on authority. The maintainer does not win an argument by being the maintainer. If a thread ends with someone deferring rather than measuring, it is a bad thread regardless of who was right.

Moderation is light and stated: off-topic gets moved, not deleted. Anything that violates the Code of Conduct is removed and said so.

Casebook threads¶

Some threads are reconstructed conversations written for teaching, published by the maintainer with illustrative roles. They carry a banner and the casebook label.

Every number in one is reproducible from a named cell in this repository. Roles are illustrative; results are not. The full convention, including the fixed role vocabulary and why this is done openly rather than with additional accounts, is in casebook-convention.md.

What not to post¶

A question with no numbers, when you have numbers. It costs everyone two turns.
A technique with no problem. "Should we use X?" → what failure does it fix, and do we have it?
A benchmark result with no interval. A point estimate is not a result.
Anything confidential. Client corpora, real customer questions, internal metrics. Threads here are public and permanent.
A new thread when a clinic thread exists. Ask there.

Good first discussion¶

New and want to contribute something useful in fifteen minutes?

Run a lab, and post in its track thread what the check messages did not make obvious. Bad check messages are bugs and saying so is a real contribution.
Take a paper from the reading list you have actually read, and open a Reading Club thread with a replication plan.
Find a number in the docs you can reproduce, reproduce it, and post [errata] if it does not match.

The third is the highest-value thing a newcomer can do here, and it is how #85 happened.