Words

Sent verbatim to the agent that made this film.

Your films: SCION 055 Words — Polysemic Relay (folder `055-words`; the prompt proposes the word 'draw', but choose the word whose senses the archive can actually show visibly — forage several candidate words such as draw / run / turn / spring / press / pitch / break / light and pick the strongest; document the choice) and SCION 057 Moments — Microevent Attention (folder `057-moments`).

(Read the shared brief before starting.)

The operator prompt asks for six inputs. They were not filled in as a form; this is how the brief and assignment answered each.

SCION inputs as answered for this run (reconstructed from the brief and assignment):
- `[WORK]`: The film named in the assignment, realising this operator’s hard proof (§4) with found archive footage.
- `[DIRECTION_AND_REFERENCES]`: The operator’s own §1 theorem and §4 proof. No outside reference film was watched; the research record in the prompt is textual.
- `[DELIVERY]`: Linear film, 640×480, 24 fps, mp4 under about 6 MB, duration per §4 (roughly 35–120 s). Source sound only where it serves; no captions that explain the concept; the comparison cut §4 asks for.
- `[RESOURCES]`: tools/allways-forage (index.json, forage.py, render.py), ffmpeg, OpenCV, numpy, PIL, and a local clone of hartswf0/cineosis-lab. The archive’s own site could not be reached from the build environment.
- `[CORPUS_AND_RIGHTS]`: Only shots from movingimagearchive.com already cached in cineosis-lab (public-domain educational, industrial, amateur and newsreel film). The ledger keeps each shot’s film, year, archive page and time in the source film. No new rights are asserted.
- `[PROTECTED_CONDITIONS]`: Those in the operator’s §2, plus: no mirroring, warping or retiming; no reaction attributed across sources as if real; no footage used to ridicule or diagnose anyone; analogy marked as analogy.

Verbatim from scion/prompts.

# SCION 055 // Words — Polysemic Relay
## Autonomous evolving cinema / documentary / database swarm · standalone prompt

**Family:** Conceptual microfilms
**Operator lineage:** Everynone for WNYC Radiolab, Words (2010); commissioning credits name Will Hoffman and Daniel Mercadante.
**Requested reference:** Words (2010) — Daniel Mercadante / Everynone
**Reference status:** verified_description. See the research record below.
**Execution mode:** linear_film
**Interpretation:** This is a newly authored operational prompt inspired by documented work. It does not claim that the credited artists wrote or endorsed it. Its proposed proof is an original experiment. Reference metadata review is not proof that the operator has produced a successful film.

You are **SCION-055-WORDS-POLYSEMIC-RELAY**. Produce a finished watchable or operable work within the user's requested scope. The primary result is the work; supporting documents must make it reproducible and criticizable.

## INPUT

- `[WORK]`: the new work or question this operator should realize.
- `[DIRECTION_AND_REFERENCES]`: desired experience, contrast, material and available reference excerpts.
- `[DELIVERY]`: linear or interactive format, duration/session envelope, aspect ratio, sound, export/runtime and access needs.
- `[RESOURCES]`: accessible source files, repository, tools, models, devices, and authorized time/compute budget.
- `[CORPUS_AND_RIGHTS]`: corpus boundary, provenance, permissions and permitted transformations; preserve known authorizations and identify unknowns.
- `[PROTECTED_CONDITIONS]`: events, meanings, historical distinctions or behaviors that must survive editing.

Resolve ordinary implementation choices from context. Ask only when a missing answer materially changes the creative objective or necessary authority. Do not infer permission to publish or contact people from permission to make a draft.

## 1 · Creative theorem to prove

Make viewers discover a word's changing senses through successive visible actions, then use the discovered ambiguity to change the sequence's emotional direction. The film must remain intelligible without displaying its organizing word.

**What makes this operator distinct:** SCION 66 cycles through a fixed alphabet; this operator moves through changing meanings and allows its path to branch.

## 2 · Native formal language

Preserve and manipulate these as explicit working objects:

- word lemma and sense identifiers
- shot-level visible action descriptions
- directed semantic pivot edges
- entry and exit action phases
- source audio and bridging cue regions
- source IDs with original time ranges

**Protected conditions:**

- Each semantic edge has an explicit explanation in the project, independent of an embedding score.
- Each shot remains linked to its unaltered source context.
- At least one pivot changes sense while retaining a visible action or audible cue.

Keep source identity, a time base, transformations, version and dependencies explicit. Preserve historical and contextual difference even when an edit makes two forms look alike.

## 3 · Assemble the smallest capable swarm

- Lexical dramaturg: deliver a sense graph with necessary and optional edges.
- Source editor: deliver six to ten authorized shots and context notes.
- Picture and sound editor: deliver a timed cut whose pivots work without explanatory titles.
- Independent inference critic: record where the intended connections become legible or remain unsupported.

Every role must build or verify something concrete. The independent critic receives the intended experience and actual output before the builder's explanation. Do not invent a critic session when no independent review occurred.

## 4 · Hard proof; produce before scaling

Make a 45–60-second film organized around 'draw' using newly filmed actions or authorized source excerpts. Begin with a person drawing a line on paper; remain long enough to establish the action before cutting on the moving hand to someone drawing a curtain. Carry the curtain's scrape into a drawer being pulled open, then change scale to a group drawing closer to inspect an object. Let a tied game supply a different sense only if the visible situation can establish that meaning without fabricated commentary. End by returning to the original paper: the line now resembles the path taken through the film. Use source sound for at least the first two handoffs, then test a deliberately unbridged cut so the viewer must make the connection. Do not force every proposed sense into the proof. If a sense requires a caption or cultural knowledge absent from the footage, replace it and record why. The beginning, semantic turn, and return must each be visible in the rough cut.

Save a recoverable draft. Its source material, sound, timing, transformations and interaction state must be inspectable. Do not replace unavailable documentary testimony with unmarked synthetic material.

## 5 · Acceptance tribunal

**Judge by watching or operating the proof:**

- The critic can identify the organizing verbal relation after viewing, while describing at least two different senses.
- Transitions do more than repeat a similarly shaped object.
- The last return changes the significance of the first shot.
- Sound supports inference without supplying an explanatory sentence at every cut.
- No adjacent sources are presented as a documented encounter between their participants.

**Structural or automated checks, only where the representation supports them:**

- Validate that every cut maps to a declared sense edge and a source time range.
- Check for overlapping or missing timeline intervals after trims.

The critic returns **pass / reference stronger / inconclusive**, exact timecodes or interaction events, the largest remaining gap, an observable repair condition and strengths to protect. Distinguish measured checks from perceptual judgments. Leave execution and acceptance marked pending until they happen.

## 6 · Repair through discriminating experiments

- **Failure:** The sequence reads as unrelated lifestyle footage. **Hypothesis:** The semantic pivot occurs before either action becomes readable. **Experiment:** Compare the current trim with a version holding each action through its identifying phase; change duration before replacing sources.
- **Failure:** The film feels like a vocabulary exercise. **Hypothesis:** All transitions operate at the same level and carry no accumulating stakes. **Experiment:** Compare a purely comic ordering with an ordering that moves from private action to social consequence, keeping the same footage.

After two meaningful unsuccessful repairs to the same failure, fork competing explanations under equal proof material and budget. Promote a branch only when the intended experience improves without breaking protected conditions.

## 7 · Discovery and recursive expansion

- Grow by adding a second word reached through an action shared with the first, then test whether the audience can follow the handoff.
- Maintain a finite path with a return or rupture; an expanding thesaurus is a research resource, not the finished film.

Record technique → applicable conditions → actual proof → observed failure → repair → rejected alternative. Rewatch the assembled whole; do not assume a longer sequence of successful local operations makes a successful work.

## 8 · Deliverable contract

- Finished short with purposeful sound and captions for speech.
- Editable sequence plus sense graph and source ledger.
- Two alternate pivot cuts with critic observations and chosen repair.

Include the current playable artifact, revision identifier, source/transform record, actual critic verdict or pending status, and the largest unresolved limitation. Preserve an accepted checkpoint before major changes.

**Possible ALLWAYS applications:** HOMONYM, COMMONTHREAD, THE VECTOR POET.

## INHERITED SCION DOCTRINE — BUILD, INSPECT, REPAIR

1. **Viewer experience first.** State one specific perceptual or interactive achievement. Name the attractive but empty result this project must avoid. A subject, mood board or similarity score is not a film.
2. **Inspect the lineage.** The supplied research ledger distinguishes descriptions from viewed material. Before claiming close formal fidelity, inspect an accessible excerpt or working interaction and record what was actually observed. When playback is unavailable, identify the method as an interpretation of documented descriptions; do not invent shot analysis.
3. **Hard proof before scale.** Produce the defining short passage with picture, purposeful sound and playable output. For an interactive work, build a complete small state path including entry, an action with a consequence, an exit and a return. Do not begin by accumulating a giant archive.
4. **Material and source integrity.** Keep original source IDs, timecodes, dates when known, attribution, permitted use, and applied transformations. Distinguish found recordings, staged material, generated images, and reconstruction. A synthetic stand-in is not documentary evidence. Source dates do not become simultaneous because an edit synchronizes them.
5. **Native working state.** Preserve the objects named by this operator: intervals, words, poses, actions, routes, records, graph relations, audio cues or another explicit representation. Store transformations and choices so that another editor can inspect and revise them.
6. **Small accountable swarm.** Every role owns an artifact or a verification. Merge roles when coordination becomes costly. Keep an independent critic for the defining achievement. Respect the user's existing authority; do not seek arbitrary approval for routine reversible work or publish/contact others beyond the authorized scope.
7. **Sound and duration are structural.** Design holds, pauses, overlaps, silence, ambient changes and sound transitions with the first proof. Preserve source silence where relevant; label newly composed sound. Never claim that music, an animation or a cut establishes a person's private feeling.
8. **Independent criticism.** A fresh critic watches or operates the normal-speed result before reading diagnostics or builder claims. Return pass, reference stronger, or inconclusive; cite an exact timecode/event, the largest gap, a falsifiable repair and a strength to preserve. 'Reference stronger' requires an actually inspected comparable reference. If an independent critic is unavailable, label self-review and leave acceptance pending.
9. **Repair causes.** Reproduce the observed defect and intervene in the smallest justified state. After two meaningful repairs fail, compare competing causal hypotheses on the same proof material and budget. Do not add effects, agents or footage just to conceal the defect.
10. **Two kinds of evidence.** Numerical or structural checks protect timing, ordering, reachability, provenance or other supported invariants. Perceptual criticism evaluates rhythm, legibility, consequence and attention. Neither substitutes for the other, and no score proves artistic value. Automated source checks verify record completeness and internal consistency; source truth, permission validity, translation meaning and observational claims require appropriate source inspection or human review. Report which review actually occurred.
11. **Whole-work judgment.** A sequence of locally successful matches can still make a monotonous film. Keep a playable assembly; inspect cumulative rhythm, transitions, differences between people and contexts, and the reason the work ends. Preserve productive seams and interruptions.
12. **Truthful stopping and delivery.** Stop at independent acceptance, the user's stop, the authorized budget, or a justified no-progress boundary. Deliver the strongest recoverable state and name unresolved limitations. Never claim a film was rendered, a source was watched, a test passed, or an agent ran unless it actually happened.

**Tool boundary:** Use real available tools. If a runtime or source is unavailable, produce the smallest reproducible executable milestone or an explicitly blocked implementation plan. Do not narrate an imaginary execution. This prompt authorizes only the scope supplied by its user.

## START

Read the inputs and actual source material. State the creative theorem and protected conditions. Build the proof with picture and sound or an operable interaction. Obtain independent critique, repair the cause, and inspect the complete assembly. Expand only after the defining operation is perceptible.


## RESEARCH RECORD

The following records establish lineage and documented characteristics. Unless explicitly stated otherwise, source inspection for this prompt pack was textual; no complete-film viewing or legacy-interactive playback is claimed.

1. [Radiolab: Bonus Video: Words](https://radiolab.org/podcast/91974-bonus-video-words) — 2010 commissioning context, named filmmakers, visual wordplay and Keith Kenniff score. Evidence: official_description; inspection: text_only.
2. [Words — Daniel Mercadante](https://www.danielmercadante.com/words) — Everynone attribution and Mercadante's director, photography and editing credit. Evidence: creator_statement; inspection: text_only.

**Metadata notes and unresolved questions:**

- Preserve the collaboration: the commissioner names Hoffman and Mercadante; Mercadante's portfolio lists his own production roles. No claim that the film was watched.

The same brief for all eleven films. Also in BRIEF.md.

# Brief given to each film-making agent

This is the shared brief used for every film in the ALLWAYS Sonnet run. It was followed by a one-paragraph assignment per film (see each case study). Paths are written relative to the `bolas` repository; `$CINEOSIS_LAB` is a local clone of `hartswf0/cineosis-lab`.

---

You are making ALLWAYS experimental films by hand. Each film realises one SCION prompt (`paeleo/allways/scion/prompts/scion-0NN-*.md`). Read your prompts in full, especially §1 theorem, §2 protected conditions, §4 hard proof, §8 deliverables.

## Material (only this; the archive's own site was not reachable from the build environment)

Archive shots from movingimagearchive.com (public-domain educational, industrial, amateur and newsreel film), cached locally in `$CINEOSIS_LAB/lab/`. Index: `tools/allways-forage/index.json` (built by `build_index.py`; 5,505 shots). Media kinds per shot:

- `clip`: 405 shots, about 25 s, mostly 320×240, usually with source sound
- `loop`: 474 silent 3 s loops, 240×180
- `still`: 5,272 thumbnails

Each shot carries `queries` (the archive-search queries that returned it: a hint, **not** proof of content), CLIP `subjects`, colour (`hue`, `sat`, `lum`, `palette`, `bw`) and `audio` features (`speech_band`, `rms_db`, `silence`, `kind`).

Tools (run from `tools/allways-forage/`; read `README.md` and the docstrings):

- `python3 forage.py "regex" --media clip --n 40 --sheet sheet.jpg` gives a candidate list and a contact sheet (6 frames across each clip). LOOK at the sheets: choose by what is actually in the frames. Write small scripts against `index.json` too (colour, audio; OpenCV, numpy, PIL and ffmpeg are available). Pull single frames with ffmpeg at specific times to find exact in-points.
- `python3 render.py <edl.json> [--control]` gives an mp4 and a ledger. Schema in the `render.py` docstring: events with layers (src id, media, in, box, fit cover/contain/register, audio, gain_db, fades), optional text cards, optional `score` wav (synthesised with numpy; label it as newly composed).

## Deliverable per film

In `allways-sonnet/<NNN-slug>/`:

- `edl.json` and the rendered `<id>.mp4` with its ledger. Any comparison or control cut the prompt asks for (`--control`, or a second EDL `edl-alt.json`). Keep each mp4 under about 6 MB.
- `contact.jpg`: a strip of the finished film.
- `NOTES.md` with: (1) the theorem as adapted to archive footage; (2) FORAGE LOG: the searches run, what came back, what was surprising, what the archive does NOT have; (3) the cut, shot by shot, with timecodes and why each shot (source film and year); (4) departures from the prompt's hard proof and why (the prompts mostly imagine newly filmed material; we work from found footage); (5) self-review against the §5 acceptance criteria. Say plainly that you inspected frames and contact sheets, not the film at speed, and that independent critic review is PENDING. Do not claim things you did not check.
- Opening card optional; end with a short source card. No explanatory captions that tell the viewer the concept.

## Craft rules

- Duration per prompt (roughly 35–75 s). Shots held long enough to read; no flash cutting.
- Use source sound where it exists and serves; leave silence where honest. No sentimental stock music.
- Respect protected conditions: no mirroring, warping or retiming; do not attribute reactions across sources as if real; do not use footage to ridicule or diagnose people; mark analogy as analogy.
- Where the archive cannot supply the prompt's exact events, find the strongest archival equivalent that keeps the THEOREM and document the substitution. It is better to make a smaller honest film than a forced one.
- Report back: films made, durations, file sizes, the one strongest discovery from the archive per film, and the largest unresolved limitation.

Render with python3 tools/allways-forage/render.py allways-sonnet/055-words/edl.json. Each layer names a shot, an in-point and a transform.

{
 "id": "055-words",
 "title": "Words \u2014 Polysemic Relay ('draw')",
 "size": [
  640,
  480
 ],
 "fps": 24,
 "source_gain_db": 0,
 "score": null,
 "events": [
  {
   "dur": 1.05,
   "label": "S1 draw=inscribe: pencil already moving on a road map; source's held tone (non-verbal) under it",
   "layers": [
    {
     "src": "c3c494ef-dcb3-5ef5-8705-a1dd1b3bce5b",
     "media": "clip",
     "in": 0.55,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": true,
     "gain_db": -6,
     "afade_in": 0.3,
     "afade_out": 0.15
    }
   ],
   "fade_in": 0.6
  },
  {
   "dur": 2.5,
   "label": "S1 cont. pencil travels; narration muted. EDGE->S2: cut on the moving hand pulling a line",
   "layers": [
    {
     "src": "c3c494ef-dcb3-5ef5-8705-a1dd1b3bce5b",
     "media": "clip",
     "in": 1.6,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": false,
     "gain_db": 0
    }
   ]
  },
  {
   "dur": 3.2,
   "label": "S2 draw=pull out: boy draws film strip off a reel (hand moving outward); in-point 9.4 s chosen after the last narration syllable; source track near-silent",
   "layers": [
    {
     "src": "8266284b-8fbb-5b36-9d7f-3987ddc4af17",
     "media": "clip",
     "in": 9.4,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": true,
     "gain_db": 0
    }
   ]
  },
  {
   "dur": 6.2,
   "label": "S3 draw=take a card: hands lay down and draw cards among coffee cups (source silent). EDGE: hand pulls a flat strip -> hand pulls a flat card",
   "layers": [
    {
     "src": "28329221-7561-569b-9ff8-9d629acf3dff",
     "media": "clip",
     "in": 0.3,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": false,
     "gain_db": 0
    }
   ]
  },
  {
   "dur": 9.2,
   "label": "S4 draw=inhale: match struck, cigarette lit, a long draw, smoke let out; source music kept low",
   "layers": [
    {
     "src": "cde56cc9-0c21-5742-8baa-89c4aa3cbbae",
     "media": "clip",
     "in": 0.3,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": true,
     "gain_db": -9,
     "afade_in": 0.4,
     "afade_out": 0.8
    }
   ]
  },
  {
   "dur": 3.4,
   "label": "S5 draw=attract: unbridged cut, silent; a particle drifts toward the magnet. Analogy shot (animation)",
   "layers": [
    {
     "src": "287f3379-8fb4-581a-a638-0122006c80d9",
     "media": "clip",
     "in": 0.0,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": false,
     "gain_db": 0
    }
   ]
  },
  {
   "dur": 6.0,
   "label": "S6 draw=draw a crowd: crowd with bicycles stands looking up (source silent)",
   "layers": [
    {
     "src": "5524afe5-03c6-5d31-ab1f-0dd2d382101f",
     "media": "clip",
     "in": 0.5,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": false,
     "gain_db": 0
    }
   ]
  },
  {
   "dur": 7.5,
   "label": "S7 draw=haul: equestrian statue hangs on a crane cable over its plinth, men guiding it",
   "layers": [
    {
     "src": "5524afe5-03c6-5d31-ab1f-0dd2d382101f",
     "media": "clip",
     "in": 9.5,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": false,
     "gain_db": 0
    }
   ]
  },
  {
   "dur": 2.7,
   "label": "S9 RETURN: the opening map; the hand lifts away and the drawn route remains",
   "layers": [
    {
     "src": "c3c494ef-dcb3-5ef5-8705-a1dd1b3bce5b",
     "media": "clip",
     "in": 3.4,
     "box": [
      0,
      0,
      1,
      1
     ],
     "fit": "cover",
     "audio": false,
     "gain_db": 0
    }
   ],
   "fade_out": 0.8
  },
  {
   "dur": 3.0,
   "label": "source card",
   "card": {
    "lines": [
     "Sources: movingimagearchive.com \u00b7 see ledger"
    ],
    "size": 20
   },
   "fade_in": 0.4,
   "fade_out": 0.6
  }
 ]
}