The specimens this course works from — three recurring collections and three model studies, with what is known about each kept separate from what is inferred.

Class Copy Archive#

This is the shared evidence library the workshops read from and write to. Every workshop deposits a result into one of the collections below, and later weeks reuse what earlier weeks left behind — so the same specimens keep recurring across the course instead of each week inventing its own.

Two rules govern the archive, and they are the same two rules the course is about.

Recurring collections#

Print#

Two documents from the same fictional copy room, and everything copied from them. Week 1 establishes the measurement discipline and the exact-copy control; week 5 starts a branching hand-copied text; week 6 reconstructs it.

They are deliberately two documents rather than one. If week 5 branched from the card students already read in week 1, week 6 would be reconstructing a text they could recall, and the exercise would test memory instead of evidence.

Each copy of the card is published as two readers’ transcriptions rather than as an image of damaged print. The measurement is a comparison of texts, so an image would add an eyesight test this course refuses to grade, and it would make the accessible route a lesser version of the task instead of the task itself.

Every card copy is an authored teaching scenario, labelled as such in its own file, rather than a recording of a real machine and real readers.

Accessibility alternative for this collection: plain text throughout, one labelled line per row, every transcription the same length as its source.

Audio#

A thirty-second source recording and everything dubbed from it. Week 2 asks which shared features are evidence of a relationship; week 3 keeps the same source and asks how a claim about it would be tested.

The relationship between X and Y is recorded as inferred, with the feature that supports it and the assumption it rests on. The parents of W and Z are known, because the log records them. Those are different kinds of entry and the archive does not merge them.

X and Z share a high-frequency rolloff because they were dubbed on the same deck. That is recorded as an equipment note, not as an edge. Shared equipment is not shared ancestry, and an archive that filed it as ancestry would be asserting a relationship nobody has evidence for.

No recordings are published. The evidence is the feature table, so the inference does not depend on hearing and the non-listening route is the route. Every item is an authored teaching scenario, labelled as such in its own file.

Accessibility alternative for this collection: feature tables in plain text, with locations as timestamps and presence stated per recording.

Image#

An authored source image, 2400 by 1600, and the outputs produced from it. Week 4 asks whether the evidence on an output can determine its history; week 10 asks whether a restoration of it matches the original; week 11 asks what a provenance record establishes and what it omits.

Output 7’s parent is recorded as inferred, and partially. A retained transformation record names a full-size JPEG as the input to the resize, with a matching hash, which excludes any route that resizes the source directly. The working-directory listing retained alongside it does not support that conclusion on its own: it shows the files co-existed, not that one was read to make the other. It does not establish the route that was proposed: every history that saved at full size before resizing would leave the same listing. Ruling out is not establishing, and the record says which of the two it is doing.

The archive holds two records of the same events that differ in what they captured: a listing of what existed, and a record of what read what. Only the second supports an ancestry claim. A listing that is genuine, relevant and insufficient is the ordinary condition of a provenance record, and week 11 returns to it.

Two evidence items a reader might expect are absent: no EXIF survives, and no upload log was retained. Their absence is recorded rather than left as a gap.

Accessibility alternative for this collection: observations as property-value tables, routes as numbered written steps, and every diagram accompanied by the same information in text.

Week 10’s restoration comparison uses a separate, smaller specimen — an 8x8 synthetic tone chart, not a photograph, since no photographic pixel data for image/output-7 was ever generated (its record above is metadata-only, per week 4). The chart has one hand-placed fine feature and one hand-placed diagonal line; a 2x2 block-average degrades it, and bilinear interpolation produces the restoration output. Of four checked features inside blocks that had genuine internal variation, one matched the retained source and three did not — a result reported as-computed, not smoothed into a single verdict. See /packs/week-10/pixels.csv and /packs/week-10/feature-table.csv.

Week 11’s provenance audit compares two metadata manifests for image/output-7 — the record made at deposit and the record on its public archive copy — against the publishing pipeline’s own log of what it changed between them. The log’s two listed operations both check out, but the actual diff has two further changes the log never mentions, found only by comparing the manifests directly. Two fields absent from both manifests are recorded as never captured, not as evidence of stripping, since Output 7 never retained that data at any stage. See /packs/week-11/audit-worksheet.csv and /packs/week-11/reveal/analysis.md.

Cross-family: a preservation budget#

Week 12 asks a question the three collections above don’t answer on their own: given a stated, limited budget, what does this archive choose to keep a full record of at all? It uses the collections’ real published item counts — 11 print records, 5 audio records, 2 image records — and, where a prior pack already measured it, real sizes (the image family’s source file is 0.0109 GB, per week 4’s retained working-directory listing). Audio’s source-tier size is the one necessarily estimated figure in the pack, labelled as such, since no audio bytes are currently published for that collection.

Two proposals for the same annual budget are compared, not merged: one brings the image family to full audited preservation while holding print and audio thinner; the other spends instead on bringing the audio family’s first real audio bytes into the archive, holding print and image thinner. Neither dominates the other under the pack’s own value score, and a revised cost assumption flips which one stays inside budget. This is a comparison across families about a resourcing decision, not a claim that the families share ancestry — the two rules at the top of this page still hold. See /packs/week-12/proposal-totals.csv and /packs/week-12/reveal/analysis.md.

Model studies#

These three are simplified models, not specimens. Each exists to test one idea under stated parameters, and each page names what the model leaves out.

Sequence model#

A synthetic replication model with declared error rates and fixed seeds, used in week 7 to watch ancestry inference degrade. A 24-symbol ancestor over a four-symbol alphabet is copied forward for 40 generations at three fixed error rates (low 0.02, medium 0.08, high 0.2 per position), three fixed seeds per rate, with agreement against the generation-0 ancestor reported at generations 1, 5, 10, 20 and 40. Settings and seeds are published for reproduction in /packs/week-07/settings.json. It omits selection and population structure, so it demonstrates a limit of inference rather than a biological result.

Bits#

A fixed short bitstring, used in week 8 to price error correction in overhead rather than in good intentions. A 3x repetition code (payload 1011, codeword 111 000 111 111) is worked by hand: majority-vote decoding corrects one bit-flip per triplet at 200% overhead, and a second flip inside the same triplet produces a wrong, unflagged decode — the guarantee’s stated limit, not an edge case outside it.

Distribution model#

A small categorical distribution resampled from its own output, used in week 9 to watch a rare case disappear under recursive re-estimation. Six categories (A–F) with true probabilities 0.40/0.25/0.15/0.10/0.07/0.03, category F the rare case at 3%, 200 draws per round, 5 recursive rounds after a shared round-0 draw, three fixed seeds per condition. Three conditions: a baseline that always resamples from the true distribution, a fully recursive condition that re-estimates each round purely from the previous round’s own sample, and a retain-original condition that mixes in a fresh true-distribution sample each round. Settings and seeds are published in /packs/week-09/settings.json. It is categorical resampling, not a trained language model, and the page says so.

What a record holds#

Every specimen record carries an identifier, its collection, its source and rights, the operation and settings that produced it, measurements with units, an accessible alternative, and the teaching weeks that use it. Known parents and inferred parents are separate fields.

Specimens are added as each workshop is built. Instructor examples are labelled as instructor examples; there is no fabricated student work here.