SAS Pattern Notes

Live site — https://jinbeiwang.github.io/sas-pattern-notes/

De-identified, reusable write-ups of coding patterns from clinical statistical programming. Each note is a single self-contained HTML file — no font CDN, no script CDN, no image files, nothing to install. Read it on the live site above, or open the file straight from the repository.

The chained kit-replacement note has been retired from this repository. It is published with the other notes at https://jinbeiwang.github.io/notes/sas-kit-chain-replacement.html, and one note would not be maintained in two places. The site root forwards there, so links into this repository keep working; the full text stays in the history as git show 21bc431:index.html. The programs that note describes are still here — see Checks.

Every note answers the same seven questions, in the same order:

1. What the code has to do — the requirement, then the data shape that follows from it
2. Background worth stating — only the machinery the listing below depends on
3. Reading the code — the current listing, annotated by step
4. Why this approach and not the obvious one — the comparison, then the argument
5. Design and correctness — what it guarantees, a scenario matrix, and the defects it does not fix
6. Reuse checklist — what to change, what to check before it goes into another program
7. The check that ships with the note — what it asserts, the dummy data behind every case, and what it does not settle

The table of contents is generated from the headings when the page loads and lives in a panel on the left that can be hidden. Adding or renaming a section needs no edit to any link list.

One note sits outside the clinical family: agent-system-prompt-note.html surveys how agent system prompts are actually written, measured across an archive of eighteen shipped ones. It is a survey rather than a coding pattern, so it runs its own seven sections and makes no claim about any SAS program. Its checks live in the same reference-model directory and are listed under Checks.

Notes

Each note is served under the site root at its own filename, for example https://jinbeiwang.github.io/sas-pattern-notes/sas-hash-clinical-note.html.

Note	Read online	Subject
《Linear Mixed Models》全书精读笔记（临床统计程序员视角）	web · source	West, Welch & Galecki 第 3 版全九个章节的精读，按原书顺序推进，每章补三件书里没有的东西：这个概念在试验里对应什么、SAS 怎么写、哪里会踩坑。四种数据形状（聚类 / 重复测量 / 纵向 / 聚类纵向）与六个案例（幼鼠、课堂、大鼠脑、自闭症、牙贴面、SAT）逐一映射到多中心试验、多次访视、评估者交叉与设计效应。含线性代数直觉附录、二十条易错点速查，以及一页"从书到临床"的词汇对照——其中最尖锐的一处是：书里鼓励用信息准则挑选协方差结构，监管要求揭盲前写死。Written in Chinese.
MMRM for a longitudinal continuous endpoint	web · source	上一条笔记第 2、6、7 章交汇处那段生产代码的完整展开（clinical note）。 The primary analysis of a Phase III continuous endpoint measured at several visits: the marginal model (PROC MIXED, no RANDOM statement, TYPE=UN, DDFM=KR), why the baseline record must leave the response vector while its value stays as a covariate, what each covariance structure costs in parameters, why the read-out goes through LSMEANS ... / SLICE= rather than hand-written ESTIMATE coefficients, and the prespecification in the SAP that the model cannot supply for itself. Written in Chinese. Six defects with the input that triggers them, including the finding that a large minority of protocols still leave the covariance structure, the estimation method and the fallback unstated.
The SAS hash object: definition, lookup, load and output	web · source	Full syntax skeleton (declare → definekey → definedata → definedone → find / output), six argument tags, the method set and its return codes; four clinical shapes (screening and population flags, horizontal ADSL→ADLB merge, cross-visit baseline retrieval, AE×CM many-to-many), written in Chinese. Eight pitfalls with sources — duplicate keys kept silently, dataset loaded at definedone() rather than declare, data variables retaining their previous value when find() misses, output() not writing keys, key: type agreement.
Chained kit replacement with the SAS hash object	web · source	Resolving a replacement chain of unknown depth in a single pass of a DATA step; why an in-memory hash and not a merge; twelve cases of imperfect input — a chain of three the longest, since that is the most any subject has shown — each one printed as dummy data and asserted by a self-contained Base SAS program that ships with the note. The duplicate rule is the part worth reading twice: duplicate: "error" only does its job because the deduplication in front of it carries the payload in its by list rather than the key alone.
Agent system prompt design: seven dimensions, eighteen shipped prompts	web · source	把公开归档的 18 份已上线 agent 系统提示词摆在一起，按角色、目标、能力边界、工具调用、约束、输出格式、安全拒答七个维度逐条测量。Written in Chinese; the quoted excerpts stay in the original English. The measurements are the substance: prompt size spans 9.7× (4,848 to 47,083 characters), the "shout budget" of NEVER/ALWAYS/MUST/IMPORTANT/CRITICAL spans two orders of magnitude, and one sentence about not ending a turn early (completely resolved) appears in thirteen files across six unrelated products while a vendor's own content-policy sentence appears in nine files and exactly one product. Two similarity measures are reported because they disagree: Claude Code and Gemini CLI score 0.50 on tag-vocabulary overlap while sharing zero twelve-word runs — a single generic <example> tag carrying the whole resemblance. Closes with eight failure modes, a twenty-item authoring checklist, and the argument the whole note is built around: a system prompt is a declaration, so every "never" should be answerable with which layer actually stops the model when it disobeys.

Checks

Nothing in these notes is a transcript taken on trust. Each claim about a program is carried by something a reader can re-run, and the checks are tied to the note rather than to a memory of it.

File	Command	What it settles
kit_chain_check.sas	sas kit_chain_check.sas	Base SAS, no study macros and no external input. Builds every case of the note's scenario matrix as dummy data, runs the chain step over each one, and asserts the kit, lot, batch, status and probe count of every visit-level row — eleven assertions in all, plus the block that demonstrates the refused load. Prints one verdict per assertion and aborts on any failure. The file the note's section 7 describes.
mmrm_check.sas	sas mmrm_check.sas	Base SAS plus SAS/STAT, no study macros and no external input. Builds a 240-subject, three-arm, five-visit fixture with monotone dropout as data, fits the note's model over it, and asserts seventeen structural claims — the parameter count of each covariance structure at K = 5, the Type 3 numerator degrees of freedom that prove AVISITN really is a classification variable, that the baseline record stayed out of the response vector, that Kenward-Roger moved the denominator degrees of freedom away from containment, and that the fixture really contains a subject with no response, a subject seen once and subjects who dropped out. Prints one verdict per assertion and aborts on any failure. A fragile block behind RUN_FRAGILE refits with the baseline record left in and prints what that costs; it is announced in the log and skippable, because its failure is the point.
reference-model/mmrm_fidelity.py	python mmrm_fidelity.py	That the note describes the program it claims to describe: the two listings printed in section 3 appear in mmrm_check.sas as a contiguous run of statements — comments, case and whitespace ignored, string literals preserved by a content-keyed mask so 'Placebo' cannot drift into something else — and every numeric value the check asserts is also stated in the note, so a number that drifts in the prose fails. Exits non-zero on any break.
reference-model/code_fidelity.py	python code_fidelity.py	That the note describes the code it claims to describe: the listings printed in section 3, and the step inside kit_chain_check.sas, against the program, statement by statement — comments, case and whitespace ignored, string literals preserved so ', ' is not read as ','. It also holds section 7's fixture and every row of section 5's scenario matrix to the check's own fixture and expected table, so a number that drifts in the prose fails the build of the note. Takes --sas because the program it checks is not in this repository.
reference-model/ec_step_model.py	python ec_step_model.py	The same cases through a dependency-free emulation of the DATA step, for readers without SAS. It reads the fixture, the expected table and the volume parameters straight out of kit_chain_check.sas, so it cannot drift from the check, and it reproduces every expected row. Its own contribution is the counterfactual a SAS session would need two programs to show: the deduplication counts under both by lists, and the answer each of them produces when two rows disagree. Exits non-zero if any row or count fails.
reference-model/prompt_corpus_metrics.py	python prompt_corpus_metrics.py --corpus <archive>	The agent-prompt note's only source of numbers. Sizes, structure, constraint-word density, eight design axes, the file-and-product spread of four signature sentences, near-duplicate families and cross-product pairs — written to metrics.json and fragments.json. Takes --corpus because the archive it measures is not in this repository; the commit measured is stated in the note's section 7. Requires an absolute overlap of at least 100 twelve-word shingles before a containment ratio is reported at all: on short files a ratio alone scores two 300-word fragments at 0.8.
reference-model/prompt_notes_build.py	python prompt_notes_build.py --corpus <archive> --out ../agent-system-prompt-note.html	That nothing in the agent-prompt note is retyped. Table bodies are spliced from fragments.json, numbers-in-prose from the numbers block of metrics.json, and every quoted excerpt is cut out of the archive at build time by the anchors in quotes.json. A placeholder that goes unresolved, a quote whose anchor has moved, or — the direction a substitution cannot see — a quote that was extracted and never placed, each fails the build.
reference-model/prompt_note_fidelity.py	python prompt_note_fidelity.py --corpus <archive>	Independent, five groups. Re-runs the measurement and compares metrics.json key by key; takes every quoted block in the finished note back to the archive file its data-src names and requires a verbatim match; compares the six table bodies with a fresh render; checks that the numbers printed in prose are present; and then a static layout group — every class used must be styled, class="num" must not land on a sentence, the longest unbreakable token must clear the column, and every figure label's estimated width must fit its viewBox. The layout group stands in for the rendered-page assertions, which could not run on the machine that built the note: no browser process can be launched there. Exits non-zero on any break.

Conventions

• Study and protocol numbers are replaced by a placeholder such as NNNN; library names, file paths, program headers, author names and input dataset names are removed or generalised.
• Listings are annotated by step and by in-code marker ((1), (2)) rather than by line number, so a listing can be lifted into another program without invalidating every reference to it.
• Behaviour attributed to the language rather than observed in the code is cited to the vendor documentation.
• A note documents the code as it stands. Where a pattern was hardened after the first note was written, the note describes the hardened step and argues for it; it does not keep the superseded listing as a foil, because the reader's job is to review what runs, not the road to it.
• A case is in the scenario matrix only if the step under test is what decides its outcome. Input that is settled upstream — a screening row, a visit label the clinical side spells differently — is not a scenario here, however interesting it is; cases that turn out to be the same failure twice are dropped rather than kept for symmetry. Retiring a case means saying so in the note, so that a reader who met the old list knows what happened to it.
• Where SAS was not available to the person writing the note, the note says so in the masthead and in the footer, and ships a runnable check instead of quoting a transcript. The expected results are then the specification the check asserts, and the note says which behaviours only a real session can settle.
• Every behaviour the note claims is traceable to something a reader can re-run: a documented statement, a scenario identifier, or one of the checks above.