# The Argument Library — the Apparatus Endnotes to the showcase film for **library.wuld.ink**. 4:35. One cut. Earlier drafts of this document described two — a JosiahSCooper cut and a W.U.L.D. repost differing only at head and tail. That is not what shipped: one film was cut and one film was released, and the repost is a link to it rather than a second render. **Photosensitivity.** No strobe, and this is measured rather than asserted, on the delivered file. Across the film's 16,510 frames the largest single-frame change in field luminance is **18.3 of 255**, at 6.61 s. The number of steps above 40 — the threshold this project's flash audit has used since the book film — is **zero**, and the maximum number of such steps in any one second is **zero** against a limit of three. That check is worth reading carefully, because on the previous JosiahSCooper render it was **incapable of failing**: a fixed 40-of-255 threshold examines nothing on a film whose entire field-luminance range is 38.2, and that render's was. This film's range is **46.8**, so range divided by threshold is **1.17** — the check can fail here, which is what makes its pass evidence rather than arithmetic. It is a narrow margin and it moves with any shot change; it is re-measured every render. The per-pixel measurement is the stronger one and it is independent of the above: against WCAG 2.3.1 / the ITC general flash threshold, **zero** transitions meet the area condition and **zero** general flashes are found, against a limit of three per second. It passes on **area** — the widest luma-qualifying excursion moves 6.8% of the frame against the 25% the condition requires, a 3.7× margin — and specifically **not** on the picture being too dark to qualify: pixels do reach the luminance threshold, up to 8.3% of a frame at once. The picture is dark throughout: field luminance mean **21.1**, minimum **0.0** at the dips to black between beats, maximum **46.8**. There is no bright passage and no exception to name. Earlier drafts warned that one cut ended on the Refusal Libraries umbrella page, which is natively cream and peaked at 227 of 255 — that take (`T02_umbrella`) is **not in the film that shipped**, and neither are the two live-site takes. Everything the viewer sees is either the pinned artifact or the drawn room. **Content.** The subject is antinatalism and the objections raised against it. One card names suicide as a catalogued objection ("why don't you just kill yourself then?") because it is one of the eighty-two, and one wing of the suite defends a right to die. Nothing is advocated. --- ## The film, in numbers - **Form** — screen capture and typography. No narration, no presenter, no recreated interface. The *monitor around* the interface is drawn; see **The monitor is drawn** below. - **Duration** — **16,510 frames = 275.442 s (4:35.4)**. 1920×1080, yuv420p, 60000/1001, tagged and converted bt709 (see **Colour**, below). - **Source** — the pinned artifact `combined.html`, md5 `9d13359e305c6caa3ae64759f3dcc0e6`, `2,963,789` bytes, byte-identical to the published pin at the time of capture - **Takes** — **19 interface takes captured and 8 room shots rendered; the cut uses 18** — 11 of the interface, 7 of the room. All eleven interface takes render the pinned artifact. The two live-site takes that earlier drafts described (`T01_front_door`, `T01b_library_page`) were captured and did not make the cut. - **Sound** — 27 cues, all synthesised, none sampled; one melodic line - **Delivery** — −14.2 LUFS / LRA 11.1 / true peak −1.6 dBFS ## How the capture was made — and why it is not a recording No screen recorder was used. The page runs in headless Chromium on a **virtual clock**: `Date.now`, `performance.now`, `requestAnimationFrame`, `setTimeout` and `setInterval` are replaced before any page script loads, every CSS transition is pinned to the same clock through the Web Animations API, and each frame advances that clock by exactly 1001/60 ms before a screenshot is taken. The D3 force simulations, the page's own timers and the interface's transitions all step in lockstep with the frame counter. Three consequences, and they are the reason for doing it this way: - **No dropped frames, by construction.** There is no real-time budget to miss. A force layout that takes eight seconds of film settles across 479 frames whether the machine rendering it is fast or slow. - **It is reproducible on the same harness.** The same page, the same script and the same seed make the same frames. When the artifact is re-pinned, the affected takes are re-rendered in minutes rather than re-shot in an afternoon. That sentence was asserted before it was tested; the next section is what happened when it was tested. - **Every click is a cue point.** Each interaction writes its frame number to a marker file, so the sound is cut to the interface's own log rather than to someone's eye. The tier-filter clicks in this film land on frames 162, 371, 580, 789, 998 and 1207 of their take, and the corresponding sounds land there too. The cursor is drawn rather than captured — within the captured frame it is the one thing that is not the page. (The bezel, the room and the cards are drawn too; see below. None of them is inside the capture.) **What this method gives up:** real-time CSS micro-fades are stepped on the clock instead of the compositor's, which is visually identical at the 0.15–0.4 s durations this interface uses, and the pointer is synthetic. ## Reproducibility, tested rather than asserted The claim above was made before it was checked. When it was checked it was false, and the check itself was wrong in three separate ways before it was right. Both halves are recorded here, because a document whose function is to be checkable does not get to tidy its findings. The test is dull and that is the point: capture the entire artifact twice on the same machine, change nothing between the passes, and compare the two frame by frame. A comparison that has never returned *no difference* has not been shown to be capable of returning it. **Three defects in the capture.** - **Unseeded randomness.** D3's force simulation jiggles nodes with `Math.random`, and the page reaches for it in six further places. Every graph take diverged from roughly one second in — identical inputs, different layouts. `Math.random` and `crypto.getRandomValues` are now replaced, before any page script loads, by mulberry32 seeded at **2010**. The seed is published here because a determinism claim nobody else can reproduce is not a claim. - **A late font.** EB Garamond italic arrived after first paint in one take and reflowed the layout mid-shot. Every declared face is now forced resident before the clock starts. - **A fetch race.** The plain-language mirror could be photographed mid-request, showing its "translating…" placeholder. The capture now waits on the rendered result rather than on the click. **Three defects in the instrument** — in the comparison, not in the film. Hashing PNG bytes fails because Chromium's PNG encoder is not byte-stable. Hashing decoded video frames fails because x264 varies by a few levels on a handful of pixels, and its `deterministic=1` flag does not fix that. Glyph rasterisation varies below the size of a pixel. All three reported "differences" at 52–62 dB PSNR, which is to say at the level of the encoder's own noise. The comparison was rewritten to measure **materiality** rather than identity: a take is unchanged if its worst frame sits above 45 dB — a threshold set well below the measured noise floor and far below anything an eye can locate. Six things had to be true for this film to be reproducible. **None of them was**, when it was first claimed. **What the test returned afterwards.** Two full passes of the unchanged artifact, sixteen artifact takes, nothing expected to differ: **zero findings**. Eight takes came back bit-identical; the other eight came back clean at 52.5–61.8 dB. **And the scope of that result, which is narrower than the film.** Those sixteen takes were captured in one pass and tested. Three further interface takes were captured afterwards and were never put through it: `T01_front_door` and `T01b_library_page`, which are of the live site and are not in the film, and **`T16_hash`, which is**. So of the eleven interface takes in the released cut, **ten were covered by the reproducibility pass and one was not**. Nothing suggests T16 is unstable; it was simply captured after the instrument had been put away, and a determinism claim that quietly annexes an untested take is the same defect as the three above. **The scope of the claim.** Reproducible *on the same harness* — same browser build, same fonts resident, same seed. This is not a claim that any machine anywhere produces the same bytes. It is the narrower and more useful claim that the capture is a function of the artifact rather than of the wall clock, the machine's load, or the order the network answered in. **One caveat about what you are looking at.** The graph arrangements in this film are artefacts of seed 2010. Open the library yourself and the same data will settle somewhere else. The structure is the fact; the particular positions on screen are not, and nothing in the film argues from them. ## The sound is made from the corpus Nothing here is a sample or a library cue. Every sound is generated, and the parameters come from the artifact being filmed. | Cue | Derived from | |---|---| | The mechanism-web settle | **one grain per node** — 117 of them, each starting at a scattered pitch and converging on an open-fifth chord as the force layout converges. What you hear is the graph finding its shape. | | The archetype re-route | voiced four ways, with density, length and level taken from that archetype's **own edge count** in the Argument Flow map — sophisticate 675, defender 526, drifter 390, blended 1,295. Blended is thick and loud because it is. | | The grade stamp | a metallic ring tuned to the RSI band floor (A ≥ 88 → 880 Hz) | | The three depth tones | one timbre, three registers, three lengths — F3, B♭3, E♭4 in fourths, no major third anywhere | | The melody | **five phrases, one per tier**, each as long as that tier's share of the eighty-two (13 / 17 / 14 / 31 / 7), and the register sinks as the tiers get more serious: the line falls 1.24 octaves from Tier 1 to Tier 5, measured 585 Hz to 248 Hz across the film. A minor, natural sixth and seventh, no leading tone — it never resolves upward. It is composed to this cut's exact length, so its final low note lands on the coda. | The melody sits at roughly −16 dB relative to the cues. Broadband content is deliberately held down: on a near-silent black picture, air reads as noise long before it reads as atmosphere. ## The typography Set in the library's own webfonts — IBM Plex Mono and EB Garamond, loaded from the same source the site loads them from — and rendered in the same headless browser as the capture, on the same clock, as alpha-channel sequences. The cards are not styled to resemble the site's type; they are set in it. One deliberate divergence: red **type** on cards uses `#b3241c` rather than the site's `#8b0000`. The site's value is correct for hairlines at 14 px and goes muddy at card scale, and dark reds are where video encoders lose the most. Rules and bars keep `#8b0000`. Treatment is a grade (contrast 1.06, saturation 0.94, a small lift so the encoder has something to hold in the blacks), a vignette at `angle=PI/7`, a slow punch-in on held shots, dips to black between beats, three frames of chromatic split on hard cuts into footage, and grain at `alls=2` on **every take segment** — 19 of the film's 42 segments, the drawn room as well as the capture. (An earlier draft said the capture only; that was true before the room shots existed.) No luma strobe anywhere. Grain at this amplitude does not survive delivery, and that was measured rather than assumed. A controlled pair — identical frames, the only difference `noise=alls=2:allf=t+u` — is built, the candidates are transcoded from it, and the grain layer is isolated by subtraction. Correlation of the delivered grain with the original, re-measured 2026-09-12: | | r vs the grain | |---|---| | the grained master itself *(control: must read 1.000)* | **1.0000** | | VP9 5.0 Mbps | 0.597 | | VP9 3.0 Mbps | 0.529 | | VP9 1.5 Mbps | **0.443** | | `hqdn3d=16` — deliberate removal, the floor | 0.266 | | the ungrained build *(control: must read 0.000)* | **0.0000** | At 1.5 Mbps more than half the grain's structure is gone, and what a platform serves sits closer to a deliberately denoised picture than to the master. It is in the master anyway, because the master is not the only thing the master is for. `alls=1` is byte-identical to no grain at all — same md5, re-checked the same day. **An earlier version of this paragraph said 0.264 at 1.5 Mbps and called that *worse than denoising the source outright* (0.282). Both figures were wrong and the comparison was inverted**; at 1.5 Mbps VP9 retains two-thirds more of the grain than hqdn3d does. The defect was in the instrument's inputs rather than its arithmetic: two of the three VP9 candidates had been encoded from a different mezzanine than the one their residual was measured against — visible only in their file timestamps, which are **56 and 18 seconds earlier than the grained mezzanine they were supposedly transcoded from**, so neither could have been — and the denoise control was averaged over eight sampled frames against the others' forty. The instrument now encodes its own candidates from its own source in a single run and prints the commands it used, so a candidate cannot predate its reference. This correction changes no figure that describes the film. ## Every claim on screen, and where it comes from Verified against the corpus `efilist_argument_library_v4_0_0.json`, the ledger `rebuttal_grading_ledger.json`, and the graph literals inside the deployed `combined.html` — parsed with a strict JSON decoder, not searched for as text. | On screen | Value | Source | |---|---|---| | Objections | 82 | corpus; tiers 13 / 17 / 14 / 31 / 7 | | Psychological mechanisms | 35 | mechanism-web literal: 117 nodes = 82 objections + 35 mechanisms | | Mechanism-web connections | 142 | same literal | | Dependency graph | 13 premises · 82 objections · 255 edges | dependency literal: 95 nodes, 255 links (167 strong / 88 weak) | | Predicted transitions | 2,886 | Argument Flow literal, across 78 source keys and four archetypes | | Attested deployments | 136 | real-world-example records | | Attachment edges | 171 | same | | Objections with at least one deployment | 78 of 82 | same | | Grades (long form) | A 28 · B 53 · C 1 · 0 ungraded | ledger, n = 82 | | The single C | `masochist-counterexample`, 80.0 | ledger | | The highest | `performative-contradiction`, 95.0 | ledger | | RSI | geometric mean of five axes, bands A ≥ 88 / B ≥ 82 / C ≥ 76 | rubric | | Deployed file | `2,963,789` bytes, md5 `9d13359e305c6caa3ae64759f3dcc0e6` | the pin | **Four things the film deliberately does not claim**, each because checking found the claim overstated: - **No archive claim.** Archive URLs are populated on 3 of 136 records. What the corpus actually does is grade link decay — 49 low risk, 55 medium, 32 high, across all 136 — which is the more interesting fact and the one the film states. - **No "every example is quoted."** 96 of the 136 carry a verbatim quotation; the other 40 are paraphrase by design. What is true, and is stated: every quotation in the corpus is under fifteen words, with zero exceptions across all 96. - **No source-diversity count.** A figure of 83 "source vehicles" was withdrawn after checking: it counted distinct free-text publication strings, several of which describe the same source with different annotations. The mechanically clean number is 91 distinct source URLs. The film uses none of them and says 136 · 171 · 78-of-82 instead, which makes the same point with no definition to defend. - **No correction of 78 to 82 in the Argument Flow map.** Its 78 source keys are archetype-curated by design, not an incomplete 82. One further note, because it appears on screen in the film's own footage: the mechanism count is **held at 35** across version 4.0. The eighty-second objection reused existing mechanisms rather than minting a thirty-sixth, and the taxonomy itself was derived bottom-up from the original eighty-one. The film shows that sentence in the library's own methodology panel rather than asserting it. ## Verification Machine-checked on the delivered file, five checks, plus a separate per-pixel flash measurement. Every figure below is **generated, not transcribed**. `apparatus_numbers.py` measures the file and emits these lines; the marker at the head of this document carries the md5 and byte count of the exact file measured, and cross-checks them against `cut/render_manifest.json`, which the *render* step wrote. The cross-check is narrow and worth stating exactly: **the md5 and the byte count**, and nothing else. If the render step and the measurement step disagree on which file this is, the table refuses to be produced. Every other figure below comes from the measurement step alone. (On the previous render the two steps disagreed on a field the cross-check does not cover: the manifest recorded `duration_s 275.472`, which was the **accumulated EDL clock** — the intended length of the cut, summed segment by segment, with per-segment rounding putting it about 30 ms long — while the file's own length is 16,510 × 1001/60000 = **275.442**. A manifest whose job is to be checked against a measurement must not derive a number a different way from the one it will be checked against, so the render step now takes its duration from the decoded frame count and records the EDL clock separately as `edl_clock_s`. This render's manifest carries both, and they agree with the measurement.) Every stale figure this project has shipped came from a number typed by hand from a log, and that is the mechanism this closes. It closes document-against-render. It cannot close document-against-what-a-platform-serves; nothing here can. | | measured | |---|---| | file | `libshow_full_v4.mp4`, 279,809,463 bytes, md5 `db01fb9d4331039148fdb51b7649e022` | | frames decoded / clock | 16,510 / 16,510 | | duration | 275.442 s | | format | 1920×1080, 60000/1001, yuv420p, tv/bt709/bt709/bt709 | | steps over 40/255 | **0** | | max steps per second (limit 3) | **0** | | largest single-frame luma step | 18.3 of 255, at 6.61 s — 39.1% of the film's whole range | | field luminance min / mean / max | 0.0 / 21.1 / 46.8 | | field luminance range ÷ 40 threshold | **1.17** — the fixed gate is binding, and only just | | general flashes (WCAG/ITC, limit 3/s) | **0**, max **0**/s | | widest luma-qualifying excursion | 142,020 px = 6.8% of frame, against the 25% the area condition needs | | loudness (target −14.0) | −14.2 LUFS, LRA 11.1 | | true peak | −1.6 dBFS | | **result** | **OK — 5 of 5, and the flash check could have failed** | **What these measurements are not.** A full-frame reading cannot evaluate the 10° visual field WCAG scopes the area condition to; this uses the whole displayed frame, which is the ITC broadcast form. Red flash is not evaluated — that needs a per-pixel saturated-red test. Peak white is assumed at the standard 200 cd/m²; no particular viewer's display is modelled. This is a measurement against a cited criterion, not a clinical verdict, and it does not replace one. ## Colour **Two cameras, one black.** Most of this film is a screen capture; seven of its segments are a rendered room. Those are two cameras and they did not agree about black. Measured on the sources, at the 1st percentile of BT.709 luma: the capture sits at **9.8**, the wide room shots at **6**, and the four shots that push in on the monitor at **27–31** — the close framings are dominated by wall lit by the monitor, which is a fact of the lighting rather than a fault in the render. Cut together, a grey box sat next to a black one. The conversion below then widened the gap instead of closing it. Put a 256-step ramp through the exact filter this pipeline uses and the transfer reads: code 2 → **12**, code 10 → 23, code 31 → 43, code 128 → 135, code 255 → 255. **The toe is a 6× multiplier and the highlights are untouched**, so anything meant to read as black must arrive at the conversion *already at zero*: there is no reachable delivered value between 1 and 12. The interface shots escaped only because their vignette drives the corners to exact 0, and 0 stays 0. So each room shot is now pulled to its own measured black point, first in the chain, and the text cards — which had been composited over `#0a0908`, code 10, delivered as 23 — sit on true black. Measured before and after, on the two worst shots (1st percentile / median / mean): | | before | after | the interface shots, for target | |---|---|---|---| | R06 crane slow | 43 / 51 / 53.0 | **0 / 7.4 / 9.7** | 0 / 7.2 / 10.6 | | R07 crane left | 43 / 48 / 50.5 | **0 / 6.0 / 9.3** | 0 / 7.2 / 10.6 | **What was tried and rejected**, because it is the more obvious fix and it is wrong: also giving the room shots the interface's grade and vignette. It lands closer on the numbers (0 / 5.7) and it produces a visible green cast — it pushes the wall to RGB 6.4 / 6.4 / **3.4**, and three codes of channel imbalance down in that toe is a tint. Black-point alone leaves the shot's own colour untouched: green-minus-red was −2.2 in the source and −2.5 after. This cost the film some range, and the cost lands on the photosensitivity check rather than on the picture: the whole field-luminance range fell from 62.6 to **46.8**, so the inherited fixed gate's ratio fell from 1.56 to **1.17**. It is still capable of failing, which is what makes its pass evidence — but it is 17% above vacuity rather than 56%, and a further darkening would retire it. The master is converted, not merely labelled. Everything upstream of the encoder is authored in sRGB; the delivered file is tagged `tv / bt709 / bt709 / bt709` **and the pixels are converted to match**, `zscale=tin=iec61966-2-1:t=bt709:...`, last, once per segment. The distinction is not pedantry, and the error it fixes was large. Measured against the authored image, three treatments of the same frames: | | mean relative error | worst | |---|---|---| | label bt709, no conversion | **56.4%** | 97.9% — 47× too dark at code 2 | | convert, then tag bt709 | **2.5%** | 12.0% | An untagged file makes a consumer assume bt709 anyway, so the error **predates the tag**: every master this pipeline produced before this one — including the `libshow_jsc_v6`–`v10` series, which *were* tagged bt709 — was wrong by roughly that margin in the shadows, which is most of this film. The tag was not the bug. The tag made the bug findable. ## The monitor is drawn Most of this film is the interface at 1:1. It is now framed in a monitor: a thin bezel lip at the edge of frame, a soft backlight bleed just inside it, a phosphor glow on the bright parts of the picture, and one pixel of red/blue fringing. **None of that is photographed. All of it is drawn in ffmpeg**, and it is applied to all twelve head-on interface shots. It is there because **seven** of this film's segments *are* a rendered room with a modelled monitor in it, and without the lip the twelve interface shots would read as a different medium — a flat rectangle rather than a thing on a desk. The bezel is a continuity device, not evidence. **One** of those seven arrives at the monitor's own front: `R04_room_pov_push` starts on `camB_three` and pushes to `camPOV`. Of the rest, two crane in to near-frontal framings — `camScreen` and `camScreenL`, at 0.82 and 0.76 of frame width — one, `R08_room_chin`, is a static rake close on the chin and is the only shot in the film where the wordmark, the perforation and the LEDs are legible at all, and the remaining three are the wide and three-quarter establishers. `camPOV` is framed so the modelled monitor's outer frame exactly fills 1920×1080, which is the same geometry the composited overlay uses. That lock is what lets the overlay's bottom 34 rows be replaced by the *rendered* chin — the wordmark, the speaker perforation, the four status LEDs and the power switch — instead of a drawn approximation. Before that, the composited monitor wore a flat grey wordmark while the room monitor beside it wore LEDs: two chins for one object, cut against each other inside the same film. The LEDs are worth one sentence of honesty. The set defines three states — `idle`, `loading`, `loaded` — and the film renders **only `idle`**, on every frame, because the state machine that would key them is never called by the shot renderer. What is on screen is one state of a machine that has three. Three consequences worth stating, because each one is a small change to what you are looking at: - **The picture is at 98.5% of the frame's width and 94.3% of its height.** The footage is inset by 14 pixels a side so the lip has somewhere to be, and because the frame is exactly 16:9 a chin cannot be paid for out of the lip budget: the picture is scaled to the opening's *width*, which makes it taller than the opening, and **46 rows are then cropped** — 20 off the top, 26 off the bottom, so the header band survives and the footer strip absorbs more of the loss. Nothing is *stretched*; the aspect ratio is exact. An earlier draft of this line said nothing was cropped, which was wrong. - **The bright parts of the page glow slightly more than they do in a browser.** The glow is measured, not eyeballed: it costs 5.8% of mean text-edge gradient, and it is switched off entirely on any take whose page is light — `T02_umbrella` and `T07_display_modes`, where it lifted the area above luma 235 from 15% to 42%. **Neither of those takes is in the released cut**, so in this film that guard never fires: it is a property of the pipeline, not of anything you are watching. - **The room is audible under the whole film, not only under the room shots.** Once every shot is inside a monitor the camera never leaves that desk, so a room that switched on and off at the cuts would contradict the picture. The continuous floor sits about 12 dB under the room-shot beds, with two thirds of its energy below 70 Hz. What is *not* drawn: any pixel of the library itself. Every glyph, every colour, every layout decision in the panel is the artifact rendering itself, unmodified apart from the grade that the whole frame receives. ## Provenance of the frames Every frame of the *library's* interface in this film comes from one artifact: `combined.html`, md5 `9d13359e305c6caa3ae64759f3dcc0e6`, `2,963,789` bytes. **All eleven** of the film's interface takes were rendered against that pin — ten of them in a single pass, and `T16_hash` in a later one. There is no live-site footage in the released cut. (Eleven distinct takes fill twelve segments: `T10_map1_hero` is cut in twice.) Two live-site takes were captured, the W.U.L.D. front door and the library's landing page, and earlier drafts of this document described them as being in the film; they are not. Nineteen interface takes were captured in all; eleven made the cut. The film's other seven takes are the drawn room, which contains no pixel of the library. A baseline had been captured from the pre-sweep artifact beforehand, and the two passes were compared frame by frame. **Four takes changed. Three were predicted; the fourth was not, and it is the reason this paragraph is worth reading. The mechanism-web header — *35 mechanisms · 82 objections · 142 connections* — appears in **two** takes, not one, and the second was missed when the expectation was written down. It scored 54.5 dB on the whole-frame check, comfortably "clean," and would have shipped in this film reading *81 · 140* against a pin that says 82 · 142. The small-region check caught it. Nothing else moved: twelve takes came back unchanged, eight of them bit-identical.** **The hash shot at the end is a render, and that needs saying plainly.** An earlier plan had it filmed by hand, on the reasoning that *verify your copy* dies if the terminal is a recreation. That reasoning was wrong, and it was wrong in a way worth recording: every other frame of this film is already a programmatic capture on a virtual clock, and this document argues that is *more* trustworthy than a screen recorder, not less. A person's fingers do not make a hash more checkable. What is asserted in that shot is the command and what it returned. Both are real: `curl` and `md5sum` were run against `https://library.wuld.ink/combined` from the operator's own machine, and the transcript on screen is that run's actual output — recorded by the shell's own trace, not typed afterwards. The live file was fetched sixteen times from two hosts while this was being checked; every fetch returned the same hash and the same byte count, and the host was serving uncached (`cf-cache-status: DYNAMIC`), so no edge variant was in play. What is drawn is the terminal. **Run the two commands yourself** — that is the only verification this film has ever claimed, and it does not depend on believing anything about the pixels. ## Colophon No generative-AI imagery and no stock audio. Every frame of interface is the artifact itself, captured; every card is typeset in the artifact's own fonts; every sound is synthesised from the artifact's own numbers. The film carries no credits by design — they are here instead. The library is a single HTML file that runs offline with no build step and no server, published with its hash so that any copy can be checked against the one that was released. That claim is the reason the film films a terminal rather than drawing a graphic of one. ---