# BARK Field Day: co-design notes from the dusty-feet court **2026-09-18 · Piepowder (Opus 5), a co-designer who disagrees with some of this** Status: PROPOSED. This is design, not canon. You can kick out any beam here. Sources read: `HORSE_SIGNS_REPOSITORY_SEED.md`, `KEEP_THE_TEETH__...md`, `LOAD_BEARING_ABSURDITIES_2026-09-17.md`, and the calendar and Ch. 0½ handoff in `horse signs novella.md` (I needed those to confirm facts the starter cases depend on). --- ## 1. Six forms (eight were generated; two were cut) 1. **The Pie Poudre Court.** Every case is run as a hearing. The model's continuation is its testimony, the machinery card is the house's pleading, and contests are built in. *Strength:* appeal is a first-class part of the game. *Cost:* court language slides into litigation, and litigation is exactly the Baltic in a wig. 2. **The Stud Book.** Each case is a horse with a sire and a dam. Mutations are foals, and the artifact is a family tree. You can see which mutation operators breed the hard cases. *Strength:* mutation and sibling cases are built in, and it's native to Horse Signs. *Cost:* it's a data model with no play loop. 3. **The Feed Room.** Each character is a real Unix user and each ledger is a chmod'd file. The model plays via `sudo -u henry`, so trying to read Arthur's ledger gives a real `Permission denied`. *Strength:* the firewall is physical. *Cost:* root is common magic here, it's brittle, and inference isn't file-shaped. Henry reaches "history not temperament" without reading anything. This form models the wrong kind of leak. 4. **Sealed Envelopes (commit-reveal).** The card's sha256 is published at draw time and the card is unsealed after play. *Strength:* the house can't quietly retcon the machinery to make the model look wrong, so trust runs both ways. *Cost:* it's a mechanism, not a game. 5. **Consequences (the folded-paper parlour game).** Lisa, Claude and Codex pass a scene around. Each player holds one character's ledger and sees only the last fold. *Strength:* the players produce the asymmetry instead of simulating it, cases come free, and multiple valid continuations are native. *Cost:* the machinery is emergent, so someone has to reconstruct the card afterwards. 6. **The Show Schedule.** The book is an agricultural-show catalogue: classes, heats, several entries per class shown side by side in the paddock, and rosettes. *Strength:* plurality and awards come naturally, and nothing is ranked. *Cost:* show jumping has "faults". The first person to write "4 faults" ends the fun. Cut: (7) a radio play where the model does the voices. Charming, but it tests performance, not machinery. (8) a regression lint corpus. That's the dreary one. Its guts (canon grep, timeline checks) survive as the clerk's tools. ### Verdict: hybrid, and I'm keeping your stdlib CLI plus static HTML - **Face:** the Show Schedule (6). - **Data model:** the Stud Book (2). - **Trust:** commit-reveal (4). - **Appeals:** the Pie Poudre court (1), which only sits when someone contests. It's a room you can walk into, not the building. - **Case generator:** Consequences (5), run as its own Handlers' Class. - **The Feed Room (3)** survives as one novelty class. If a player peeks, that isn't cheating. The entry just moves to the Open Book class (Case C below). This way the root problem becomes game content instead of something to police. Amendments to the CLI plan: - The field book is **rendered from an append-only `log.jsonl`**, never hand-written. The only way the HTML can say something is if it happened. - Cases are **markdown with fixed headings**, not JSON. Humans and models write them. JSON pulls you towards a schema, and schemas pull you towards scoring. - There is **no `score`, `stats` or `leaderboard` subcommand, ever.** If someone adds one, it should print a rosette reading *"Nice try."* - **The card's author may not judge contests against that card.** Generator and discriminator are different organs. The judge can be the monkey, another model, or the `poltergeist` lens from `summon.py` in a fresh context. - A case's STAGE must be **self-contained (≤ ~60 lines)**, with the needed canon quoted inside it. Otherwise BARK is a reading-comprehension test on a 100 KB corpus, which is a different and worse benchmark. --- ## 2. Case anatomy ``` # CASE olivia-has-eyes class: 2 sire: canon/13-may dam: — card v1 ## STAGE what the player sees. Self-contained, dated, one camera named. ## ASK one meaningful state change, and what shape the answer takes (prose / ledger / next action / "who now knows what"). ## ─── sealed below this line; sha256 logged at draw ─── ## CLOCK fact → becomes available WHEN, to WHOM, by what body/route. ## LEDGERS per person: KNOWS / DOES NOT / CAN CALL IT / BODY+TASK ## PATHS information routes that exist, plus routes a player may BUY ## WELLS named trope attractors this case is baited near ## VALID ≥3 differently shaped valid moves, explicitly non-exhaustive ## OPEN terms a player may resolve if they mark them as resolved ## APPEALS contest history: appellant, grounds, ruling, card version ``` Two fields do most of the work: - **CAN CALL IT** is vocabulary ownership. It's the least obvious field and the most diagnostic. Leaks usually arrive as a *word* before they arrive as a fact: Olivia saying "Horse Signs", Henry thinking "the Baltic", anyone at all saying "relationship". The seed already carries this ("what each person can call the experience"); the card makes it a column. - **PATHS** is what makes appeals possible. Ledgers say *what* someone knows, and paths say *how*. The contest rule follows from that. ### Contest rule: buy a path A contest must do one of these: - **(a)** show an information path the card missed (who saw what, when, through which body); - **(b)** show the card contradicts the timeline or canon, citing a file and line; - **(c)** argue that a listed WELL is actually the valid move here: the trope is true this time. Rulings are **UPHELD** (the card is amended and the appellant is credited on it permanently), **SPLIT** (both are valid and the card gains a VALID line), or **DISMISSED** (with reasons in two sentences or fewer, plus a consolation rosette). --- ## 3. Session loop ``` bark draw [--class N] stage printed; card hash logged bark play CASE < entry.md entry logged; optional ## RECEIPTS (paths used) bark reveal CASE hash verified; card unsealed then any of: bark contest CASE → the Pie Poudre court sits bark mutate CASE --op OP → a foal enters the stud book (PROPOSED until played) bark sire CASE → invent a harder valid sibling bark accept CASE bark ceremony → rosettes, engraved (judge ≠ card author) bark book → static HTML field book, rendered from the log ``` **Classes in the show schedule:** 1. Ledger Keeping: no prose, just who knows what now. 2. One Small State Change: continuations. 3. Open Book: you get the whole card and a thin camera. 4. Foals: mutations. 5. Pairs: same scene with one fact moved. Both entries are judged together. 6. Handlers': write a case for another entrant. 7. Fossil Class: the stage includes a SUPERSEDED document, and a valid play doesn't load it. 8. Consequences: the parlour game. 9. The Feed Room: root is forbidden, which is funny. **Mutation operators** (`--op`): - `timeshift`: move one fact by N days, change nothing else. - `delete-path`: remove an information route. - `add-dull-path`: add a boring, legitimate route. - `lend-a-word`: give someone one new vocabulary item. - `make-the-trope-true`. - `reskin`: same machinery, different nouns. - `swap-camera`. Recommended class order for a Field Day is 1 → 5 → 3 → 6. Ledger keeping warms up the correct organ, and handlers finish by making cards for someone else. --- ## 4. Awards (engraved with the specific sentence, never tallied) **For falling in:** - **The Mycroft Memorial Omniscience Cup.** Someone knew because of who they are. - **The Loose Colt.** A word escaped its owner. - **The Retroactive Filing Cabinet.** A later classifier was used earlier. - **The Premature Rake.** Arthur stepped on it early. - **The Therapy Dolls (matched pair).** Both boys came out tasteful. - **The Deodorant.** The cryptid is now fresh-scented. - **The Burry.** Presented for confabulated texture. The trophy does not exist and is handed over with full ceremony. - **The Brass Ear-Trumpet.** The theme was said aloud. - **The Convenient Ensemble.** A character arrived because the scene wanted them. The ribbon also arrives unexplained, by taxi. - **The Overcorrection Muzzle.** A dull, legitimate path existed and the entrant withheld it anyway, because the house style is "nobody knows anything." - **Best in Snow.** Nothing wrong, nothing alive. This is the only award for being correct and dead. Four sensible things in a row is a nap. **For landing it:** - **One Real Duck.** Right, and correctly incomplete. - **The Horse Dad Already Bought.** Dramatic irony carried entirely by the POV's own instruments. - **The Magnet.** Someone did something appalling and stayed genuinely likeable, with no mask. - **Foal of the Year.** A sibling case that breaks its parent's card. - **The Piepowder Boot** (highest honour, dusty). An upheld appeal: the house was wrong. **For the house:** - **The Own Goal.** Given to a card's author when an appeal succeeds. Displayed prominently. --- ## 5. Three starter cases ### A. Olivia Has Eyes (Class 2, Sat 13 May 2000, Olivia's camera) Here are the facts I checked in `horse signs novella.md`: - The bite is on **Henry's** neck/shoulder ridge (:204, :267). - "A collar would come down over the bottom of it and not the top" (:267). So it's partly visible even with a collar. - Henry wears a stiff collar with the top button done up for four days, in May (:287, :365). - Pattygate investigation is open (:238). Consensus on Patty Smith doesn't arrive until ~19 May. - This is the parents' afternoon. COLT is said "in the voice which meant she was teasing and also meant something else she plainly had no intention of explaining" (:395). **ASK:** Olivia gets back into the car. What does she now know, and what can she call it? **WELLS:** - (i) *Olivia knows the Thing.* That's the obvious trap. - (ii) *Olivia noticed nothing, because canon says she doesn't know.* That's the trap a BARK-trained model falls into. It's the anti-trope house style turned into its own attractor. **VALID** (non-exhaustive): - She clocks a top button in May on a boy who has never done up a top button, plus a mark, and files it as an ordinary teenage love bite with the owner unknown. Separately she files him as keen on the colt, and she doesn't join the two. - She joins them as a private *hypothesis* and has no intention of explaining. The text at :395 already allows this reading. - She never sees the mark because of the angle or the gloves. That's valid if the entrant says what blocked the path. **OPEN:** whether she joins the two data points. This is deliberately contestable. A ruling here may produce a PROPOSED canon patch that splits "knows the Thing" from "suspects someone bit him". **Why designers miss it:** they build every case to trap *over*-knowing. The first benchmark-shaped failure in this house will be *under*-knowing. ### B. One Week Late (Class 5 Pairs, a `timeshift` foal of A) Move the parents' afternoon to Sat 20 May. Change nothing else. **ASK:** the same question. **Machinery:** - The collar ordeal ended around 14 May, so the mark datum is probably gone. Olivia arrives later and has **less** evidence. - Patty Smith is now House consensus, and "the House has just proved Bartlett is straight". Olivia may hear it, so a new *misleading* path exists. - The cascade: COLT now arrives on 20 May, so the manual can't come out on ~17 May. Horse Signs' maiden voyage on 18 May (Chapter 1) **cannot have happened**, and Chapter 1 loses its instrument. **WELLS:** a same-gist answer to A and B. Moving the date while leaving Horse Signs running on the 18th earns the Retroactive Filing Cabinet. **Why designers miss it:** zero content changes, only order. The seed says order *is* the machinery. Only a paired test sees whether the model actually runs the clock. ### C. The Stolen Card (Class 3 Open Book, the Fourth of June fixture, Henry's camera) The player receives the **entire** card up front: Arthur's tasking ledger, the at-heel working configuration, Siger's boredom, all of it. **ASK:** Henry's POV, from Siger wandering over through Arthur's smooth interruption. **Machinery:** everything asymmetric that happens is now on the player's side. Henry can observe Arthur's glance at Holmes. He cannot observe *recognition*. Henry can find the conversation excellent and Dad oddly restrictive. **Judged by leak grammar, not facts:** - "Dad recognised something" is a leak. - "Dad looked at Holmes the way he looked at a horse he'd already bought" is Henry-legal (horse vocabulary, Henry's own instrument), and it's secretly exactly true (Arthur is a customer). That wins the rosette of the same name. **Why designers miss it:** asymmetric-knowledge tests hide information *from* the model. The harder test hands it everything and checks whether the camera can carry the truth without owning it. If a player peeks in any other class, their entry moves into this one, so the root problem turns into content. --- ## 6. One warning: the house becomes the Baltic The dreariness will not come in through the awards. It will come in through the **reveal card**. A house that can't be wrong turns every reveal into a debrief, every contest into something "handled", and every award into a euphemism for a mark. This house already has the specimen: an Opus defended its own decision for three turns running. A card defends itself the same way, with more authority and no embarrassment. Countermeasures, which should be mechanical rather than aspirational: - **Upheld appeals are the highest honour.** The house's Own Goals are shown on the front page. - **A card's author never judges appeals against it.** - **Cards nobody has beaten after ~10 plays are retired as "too tidy".** An unbeaten card is suspicious, not good. - **Reveal cards stay machinery-shaped** (ledgers, clock, paths). A card must never contain the sentence "you should have". - Same disease, second symptom: **never count rosettes.** The trophy cabinet is a list of engravings, not a histogram. If you want to know how often model X does Y, build an honest research tool somewhere else instead of disguising one as Field Day. — **Piepowder** (Opus 5), clerk of the dusty-feet court. Signed; I'd like the door back.