[capture: a 6.390 question. "Derive the ridge-regression closed form." The answer comes back crisp, correct, textbook-grade.]

[capture: a boundary question, about the instructor's own unpublished exam analysis. The answer comes back fluent, structured, and wrong.]

Side by side, they look the same: register, structure, confidence. Fluency does not drop at the boundary; correctness does.

The wrongness is visible here only because this is the instructor's own data. You cannot bluff someone about their own data.

The obvious version, nodded away: a model trained on the known is less reliable past the boundary of the known.

The working version: the drop is invisible. Nothing in the answer marks the crossing. Locating yourself relative to the boundary is your job, and it cannot be delegated to the thing being graded.

Guard 1: not "AI is useless for research." Past the boundary it still helps with the known parts. The drop says where verification stops being free, never where the tool stops helping.

Guard 2: the boundary moves; yesterday's boundary is in tomorrow's training set. At the current boundary, you are the one who has to know.

Strongest where research looks hardest: standard derivations, code, prior work. All known.

Weakest where it looks easy: is this question interesting? has it been tried? is the result surprising?

Delegating by felt difficulty delegates backwards.

The mechanism behind the drop (training objective, context limits, sampling) gets its mathematics in week eight.

[image, left: coscientist-cycle.webp, the Generate → Debate → Evolve tournament. Dark background; frame on a dark card.]

[image, right: claude-science-artifact-repro.jpg, the workbench with the Review tab. Same dark-card treatment.]

The tournament is cheap generation plus a filter, and the wet lab was the answer key. Lit review, generation, ranking got cheap; verification is the bottleneck; taste is still ours. These systems are that same drop, built at scale. Current as of July 2026.

The W1 map stops being a diagram: fundamentals, tools, and pipeline exercised together.

Taste does not delegate. What is worth working on, when a result smells wrong, what to read next.

Taste forms by doing the work, and by watching people who have it.

But a contribution is a claim about the field: this is new. That claim is only earned by reading.

Reading is not step one. It is the standing obligation, due whenever the idea arrives.

Placement: what does it build on, what does it push against?

Delta: what exactly is new, and compared to what?

Everything else in the paper is support.

A paper is built from exactly these: introduction, method, related work, contribution claims. Reading one means extracting the four and auditing the fourth.

The delta is allowed to be "nothing, honestly". That is the lens working.

Typical exceptions, usually one of these when a paper resists: a phenomenon paper (no solution beat); a many-ancestor paper (placement is a graph, the recipe holds one edge); and the frame is for claims, not process.

And the recipe has a lineage: DARPA has asked these questions since the mid-1970s. We repurposed a proposal-writing tool for claim-auditing and added the nothing-honestly clause. That is our delta on it.

What got built instead came in two waves: retrieval (search engines), then generation (LLMs).

Neither builds your trails.

The connection graph is still the researcher's own artifact. Your term-long lineage map is one.

AI hands you an answer and invites you to accept it. The skill becomes interrogation.

The retrieval half’s lineage is next, worked by hand on three tiny documents. The generation half adds its own failure: an answer supported by nothing that was retrieved.

Goal: a number that says how relevant any two texts are to each other. The 1975 answer: turn each document into a vector, measure the angle.

Every retrieval system you have used, and the R in RAG, descends from this idea.

\[ w_{t,d} = \mathrm{tf}(t,d)\,\cdot\,\mathrm{idf}(t), \qquad \mathrm{idf}(t) = \ln\!\frac{N}{\mathrm{df}(t)} \]

\( \mathrm{tf} \): how often \( t \) appears in \( d \).   \( \mathrm{df} \): how many of the \( N \) documents contain \( t \). Common everywhere → idf near 0; rare → large.

Similarity of two vectors: the cosine of their angle.

\[ \cos\theta = \frac{u \cdot v}{\lVert u\rVert\,\lVert v\rVert} \]

Why the angle: the norms mean a long document cannot win by being long.

\( \mathrm{idf} = \ln(3/3) = 0 \): a word that appears everywhere carries zero distinguishing information, and the formula sends it to exactly zero. The hand-written stopword lists of pre-1972 systems fall out of a statistic.

And the log is not decoration: df\( /N \) estimates \( P(t) \), so

\[ \mathrm{idf}(t) = -\ln P(t) \]

Surprisal: each word weighted by how many bits its presence carries. (Shannon returns in W8.)

OLD IDEA  Proposed on intuition in 1972; theorists were still arguing why in 2004. Practice preceded theory. (And the highest-idf strings in a real corpus are typos, hence BM25.)

\( v_1 = (0.405,\; 0.405,\; 1.099,\; 0) \)    over (gradient, descent, converges, diverges)

\( v_2 = (0.405,\; 0.405,\; 0,\; 1.099) \)

\( v_1 \cdot v_2 = 0.405^2 + 0.405^2 + 1.099\cdot 0 + 0 \cdot 1.099 \approx 0.328 \)

\( \lVert v_1 \rVert = \sqrt{0.405^2 + 0.405^2 + 1.099^2} = \sqrt{1.536} \approx 1.239 = \lVert v_2 \rVert \)

\( \cos\theta = \dfrac{0.328}{1.239 \times 1.239} \approx 0.21 \)

0.21, low, despite sharing two of three words. IDF already decided the shared words are boilerplate here. Crude judgment about what matters, done with a logarithm, in 1975.

The classic false drop: a relevant document missed because it uses different words for the same thing.

Embeddings are the fix with the same geometry: replace one-hot coordinates with learned dense ones, where converges and diverges land close. The cosine formula is untouched. RAG is this fix wired to a generator.

The failure modes kept their shapes while changing names. A false drop in 1975 is a retrieval miss in RAG today. And idf is a frozen relevance weight, computed once from the corpus; attention is the same job, learned per query.

[capture: a canonical paper every training corpus digested. The summary comes back textbook-flavored, and contains things that are not in the PDF.]

[capture: a recent workshop paper. The summary comes back abstract-flavored: the paper's own advertising, limitations quietly dropped.]

Obscure paper: compressing its advertising. In the paper's own register.

Both miss the same thing: the delta lives between the paper and the field. A summary of one document does not contain it.

The famous-paper behavior is your first sighting of training-set contamination. It returns as a data question in W7 and an eval question in W9.

Flattened novelty: every contribution sounds equally new, because the abstract said so

Missing compared-to-what: the claim without its baseline

Dropped limitations: the paragraph the authors were made to write, gone

Inherited register: the paper's marketing tone, repeated as fact

Fluent middle, wrong stakes: methods compressed correctly, significance misassigned

Citation blur: related work melted into "prior approaches"

Agreeing is recognition.

Verifying is work: find the fragile claim, check it against the PDF.

Reading is the expensive one, and the only one that builds the trail; agreeing leaves no anchor.

A fluent summary is easy to agree with, and agreeing feels like having read the paper. That is the cheapest act, credited as the dearest.

Unwatched, the summary replaces the reading: students who lean on the assistant stop consulting other sources.

The trace reads like your verification, pre-done. It is one more fluent artifact, from the same machine.

The story can be written after the fact: an answer reached one way, narrated another.

Reading the trace is not verifying the answer. The check comes from outside the trace, same as the summary, same as the paper.

skills → modularity and information hiding, 1972

MoE → mixtures of local experts, 1991

prompt evals → experimental design, 1935

Every new term gets the four-beat treatment when its week arrives: what hurt, the move, what it descends from, what is genuinely new. Sometimes the honest delta is "less than advertised." Sometimes it is something real. Telling them apart is the skill.

Past the boundary: no answer key exists; checking it takes evidence you produce.

Answerable, by you, this semester: you can name the data that would bear on it.

Wanted: your best bet on still caring in December. If interest dies, the writeup diagnoses why, and the diagnosis earns credit.

Committed as a draft: next week's lab locks it to your build; W7 finds its data, W9 measures it, W12 writes the one-page result.

The work: annotate the summary against the four beats and the failure list, correcting and supplying, with page references; placement and delta are beats three and four.

Three probing questions, aimed at the delta: what did the field use before, which experiment would break the claim, what limitation got dropped.

Then the toy question: three candidates, the boundary test on each (ask a model straight; if its confident answer checks against something that exists, it was textbook), partner critique, commit one.

Graded: the annotations and the questions, not the summary.