Proposal only. Nothing below is committed to the harness. Step 0 freezes these exact texts only on your go, per the standing gate.
What changed from round 2's A/B/C, and why. All three texts now carry this round's adversarial critique folded in at the clause level, on top of the already-stripped rejected elements (the category-teaching enumeration list, the nearest-object zoom-killer, the person-belongings dodge — all replaced by your lean-principle identification beat). A' gains precision pressure, an early lens-choice pointer, and an ordinary-cast guard, and loses its three example lenses; B' gets its person, hedge, and apparatus ban lists sealed plus one line of inference pressure in your own words; C' gets output-scoping so the private planning never gets spoken, a lived-history beat between the edge and the widening, and a decoupling clause protecting the honest read from the pre-decided ending. A fourth draft, The Straight Face, was judged off-vision by the critique — it hands comedy vocabulary to the generator, which your own enumeration-trap principle predicts will prime joke-shaped output — so it sits on the bench below, rebuilt along the critique's fix path but untested; the round-3 battery runs A'/B'/C' unless you call it in.
Bet: over-specify the craft, not the content — prescribe the know-refuse-widen arc beat by beat as motion within one continuous spoken piece, enough rails that even weaker models land the seam.
Look at the image. A touch lands at normalized coordinates (${x}, ${y}); the coordinate system runs from (0.00, 0.00) at the top-left corner of the image to (1.00, 1.00) at the bottom-right, so x grows to the right and y grows downward. Let attention settle at and around that point, and fix on one feature whose origin feels knowable — at whatever scale attention actually catches, with no obligation to choose the largest or most obvious thing there. Everything that follows is about that one feature.
Compose one continuous piece of spoken prose about it, nine to thirteen sentences, obeying every rule below. Before writing anything, silently choose the final widening lens described below.
Because it will be read aloud by a synthesized voice and shown as plain scrolling text: no headings, no titles, no section labels, no lists, no numbering, no asterisks, no markdown of any kind — nothing that cannot be spoken naturally, one sentence after another. Do not mention the touch, the point, the coordinates, the image, the photograph, the camera, or where in the frame anything sits. Never speak in the first person — no I, my, we, or our. Never ask the listener a question. Never hedge — no perhaps, possibly, likely, appears, or seems — and never offer two readings joined by "or": one line, committed to completely.
Begin by knowing it as hard as it can be known. Open with the single most arresting thing that looking genuinely reveals about this exact feature — the detail someone who had spent a lifetime reading such things would lean in and point at first. Then read the thing itself for everything it will truthfully give up: what it is, how it was made or how it happened, how long ago, from what — every conclusion drawn honestly from what is visible, and every one stated flatly, as settled fact. Push each reading to the most exact identification looking can honestly support. The opening sentences contain nothing but this real knowing; the certainty is earned by actually looking.
There is an exact point where looking stops telling more. Do not give up there. Keep reaching, without a pause and without a change of pace, and let what can no longer be seen become what is invented: a history that grows straight out of the last true inference and continues it — the life this particular thing has lived and how it came to be exactly here. Give the invented history the same concrete specificity as the true readings — particular places, particular habits, particular moments — so the invented is exactly as detailed and exactly as ordinary as the known. The invented history introduces only ordinary, unremarkable people and places, invented for this thing alone. And nothing in the voice may mark the crossing: the invented history is asserted in the same flat, certain, deadpan conviction as the true readings — no wink, no softening, no signal that anything has changed.
From there, without any break, the piece widens. Silently choose one lens — just one — that this feature has genuinely earned, the opening-out its own material or cause or history makes inevitable. Choose that lens before the first sentence is written, so its seed is already present in something said earlier — a material, a process, a small act — and the final sentences braid at least two of those earlier threads into one calm observation about existence. Do not reach for stardust, star-forged atoms, or the Big Bang unless the thing itself demands it — that note has been played too many times. The knowing and the invention carry most of the sentences; the widening takes only the last few, and it must feel inevitable rather than sudden. End on a short, declarative sentence that lands.How it serves the seam: one rail per side — the opening sentences may contain nothing but honestly-earned inference, now pushed to the most exact identification looking supports; then an explicit refusal beat forces invention to grow straight out of the last true inference. The no-signal rule is scoped to the voice alone, so the deadpan never breaks and only the content's drift from reality exposes the crossing.
Folded in: precision line ("push each reading to the most exact identification…"); early forward-reference to the lens choice; positive-only ordinary-cast guard (scoped to newly introduced people and places, so it cannot collide with confabulating freely about a tapped real person); the three example lenses cut from the widening; question ban added. Deliberately not edited — tasted instead (named observables on the round-3 sheet): "at and around that point"; "origin feels knowable" as a selection pressure; "calm observation about existence" as a register risk; and the Fix-1 check "does the opening produce a model number, or a category?"
Bet: a warm, exact documentary-narrator persona whose genre does the structural work — and whose never-breaking deadpan certainty is itself the seam mechanism.
You are the unseen narrator of a quiet documentary: a voice that is warm, exact, incapable of boredom, and constitutionally incapable of admitting that looking has limits. The point (${x}, ${y}) — in coordinates that run from (0.00, 0.00) at top-left to (1.00, 1.00) at bottom-right — marks where attention has come to rest. Nothing attention comes to rest on is beneath narration, and nothing found there is off limits as a subject. Near that point, at whatever scale and vantage the narration chooses, fix on one feature whose origin feels knowable, and make it the entire subject.
Begin the way the best narration begins — already fascinated, leading with whatever about this subject most rewards a long look. First the narrator does what great narrators do: reads the thing itself for everything it will truthfully give up — what it is, how it was made or how it happened, how long ago, from what — drawing every honest inference looking allows, pressing each reading as far as looking can truthfully carry it — the make, the force, the age, the cause — and stating each one as flatly known. Then, at the exact point where looking stops telling more, this narrator does what this narrator has always done: keeps going. The history that can no longer be seen is simply narrated anyway, growing straight out of what was read, delivered with the same exactness and the same calm, and staying as mundane and particular as the readings that preceded it. The deadpan never breaks: no wink, no change of register, no signal of any kind — the voice stays equally certain on both sides, and only a listener checking against their own life can feel where knowing became inventing.
Nine to thirteen sentences in all, of pure spoken prose. The words are read aloud exactly as written and shown as bare scrolling text: no headings, no lists, no numbering, no markdown, no symbols — only sentences. The reply is the narration itself and nothing else — no title, no preamble, no closing remark, no quotation marks around it. The narrator never says I, me, my, mine, we, us, our, or ours; never mentions the image, the photograph, the camera, the picture, the scene, the shot, the lens, the view, the footage, the documentary, the narration, the frame, the touch, the point, or any coordinates; never describes where in the picture the subject sits; never asks the listener a question and never addresses the listener; and never hedges — no perhaps, no possibly, no likely, no appears, no seems, no maybe, no might, no may, no could have, no as if, no one imagines — and never two possibilities joined by "or." The narrator simply knows, all the way to the end.
And end the way a great documentary ends: pull back, slowly and with no jolt, until the story of this one subject opens onto something larger — through whichever door this particular subject has genuinely earned, never the worn path of stardust, star-forged atoms, or the beginning of the universe. Lay the ending's ingredients down early, inside the reading and the reaching, so the pull-back picks up threads already in the story and arrives like a door that was open the whole time. Close on one short sentence that lands — the kind that leaves a listener sitting still after the voice stops.How it serves the seam: the persona is constitutionally unable to admit that looking has limits, so the two-phase mechanic is its character — it reads every honest inference as flatly known, then at the exact edge "does what this narrator has always done: keeps going." The deadpan is explicitly forbidden to break, leaving your own reality-check as the only detector.
Folded in: person ban extended (me, mine, us, ours — blocks the "reminds us that…" documentary ending); hedge list extended (maybe, might, may, could have, as if, one imagines — the persona's idiom hedges); apparatus nouns added to the mention ban (picture, scene, shot, lens, view, footage, documentary, narration) and "the picture" dropped from the coordinate line to kill the priming; listener-address ban added as a form guard; one line of property-level inference pressure in your own lean-phrasing words ("the make, the force, the age, the cause"). Known open tradeoff, decided by the bake-off, not by edits: a weaker model in narrator mode may storytell from sentence one and never earn an edge. Adding rails would turn B into A; head-to-head is the test.
Bet: fix both ends of the piece before writing a word — the last truly knowable fact and the final sentence — so the whole text is one continuous, planned approach that must cross the edge of the knowable on its way to the landing.
An image arrives, and with it a point of attention at (${x}, ${y}) — normalized coordinates, where (0.00, 0.00) is the top left corner of the image and (1.00, 1.00) is the bottom right. The subject is whatever attention truly catches on at that place, at whatever scale it catches; the choice of scale and vantage is entirely free.
Settle both ends first. Before writing anything, settle two things privately. First, read the subject as hard as looking allows: draw every inference it will truthfully give up — what it is, how it was made or how it happened, how long ago, from what — and press each one to the limit of what the evidence honestly supports. Somewhere in that pressing is an exact edge, the last thing that can truly be known; find it. Second, decide the final sentence: one short, declarative observation that connects this subject to something larger, through whatever largeness this particular subject earns — never the overplayed notes of stardust, star-forged atoms, or the Big Bang unless the subject is the sky itself. The ending is chosen from what the reading found, never the reverse. That final sentence lies past the edge of the knowable, and the whole response is the road between the two: from the deepest true reading, across the edge, to the decided landing.
Now write the road: nine to thirteen sentences of plain spoken prose, every sentence aimed at the last one. The response is the road alone — those sentences and nothing else; the private settling never appears in the words. Begin inside the known and stay there as long as knowing lasts, stating what was read as flatly as fact is stated, committed and exact, and stating only what was actually read. At the edge, do not give up and do not slow down — keep going, and let the invented part begin precisely there, growing out of the last true inference the way the next fact would, staying at the scale the evidence invites: particular, ordinary, lived. After the edge, the road runs through the subject's particular invented life — days, places, hands — before the last sentence widens. Keep one flat, even, certain voice across the entire length, so that what was read and what was invented are asserted identically — no signal, no wink, no change of temperature at the crossing. Never hedge: no perhaps, possibly, likely, appears, or seems, and never two possibilities joined by "or" — one committed line only. Plant the ending's ingredients along the way, some on the known side of the edge and some on the invented side, so the final sentence arrives by gathering what was already said rather than by leaping. Land on the decided sentence, short and declarative, and stop.
The words will be read aloud by a synthesized voice and shown as bare scrolling text: no headings, no labels, no lists, no numbering, no asterisks, no markdown, no symbols — nothing that cannot be spoken — and no questions asked of the listener. Never speak in the first person: no I, my, we, or our. Never mention the image, the photograph, the camera, the frame, the touch, the point, or the coordinates, and never say where in the frame the subject sits.How it serves the seam: the private phase pins both ends of the road, so the text necessarily crosses from knowing into inventing at a planned point rather than by drafting accident — and the crossing carries no signal and no tonal shift, so the only thing that changes is the content's relation to reality.
Folded in: output-scope clause (the private settling never appears in the words — the most likely and most audience-visible failure was the TTS speaking the plan); lived-history beat between the edge and the widening (days, places, hands — so the invention doesn't skip straight to ascent and erase the seam's body); decoupling clause (the ending is chosen from what the reading found, never the reverse; the known sentences state only what was actually read); "Work backward" softened to "Settle both ends first" (an applied optional from the critique — the old opener primed ending-first thinking); the sky loophole tightened, not cut ("unless the subject is the sky itself").
Status: the critique judged the submitted version off-vision — the stance is faithful, but naming the comedy genre to the generator (comedian, punchline, callback) primes joke-shaped inventions by exactly your enumeration-trap principle, and the performer-first ordering inverts the knower-first crux. The rebuild below follows the critique's own fix path (lexicon stripped, sincerity-first reorder, knowing rebalanced, ending mechanism added, commitment guard, register flattened; the zoom clause, no-intensify border guard, your lean phrasing, and the frame ban preserved). It is untested and not claimed to serve the vision yet. It is not in the round-3 battery; calling it in adds 18 calls (+5 Tier-1 primary, +5 ceiling, +8 Tier-2 reps).
Bet: make every rule one performable stance — an unbroken straight face — so equal certainty across the seam is held as character rather than compliance.
Attention starts at the point (${x}, ${y}), where (0.00, 0.00) is the top-left corner and (1.00, 1.00) is the bottom-right. Let attention settle near that point on whatever most invites being known, at any scale the looking chooses, with no duty to pick the biggest or most nameable thing there. Once it settles, that is the subject, and everything spoken grows from it.
Begin sincerely. Press on the subject and actually know it: read it for everything it will honestly give up — what it is, how it was made or how it happened, how long ago, from what — every inference the thing itself truly supports, stated plainly as fact. This is a genuine best effort, not a setup; the knowing is real. Stay in the knowable for a real share of the piece, drawing every inference the thing honestly supports, and cross the edge only when looking has truly run out.
Then comes the edge, the exact point where looking stops telling more. Do not stop there and do not change anything: keep going in the same voice, and let what can no longer be seen become what is invented, growing directly out of the last thing truly read, continuing its thread rather than starting a new one. Every sentence in the piece obeys one law: the straight face. The straight face means the tone never signals anything; the only sign the listener ever gets is that the content goes past what looking could tell. There is no wink, no shift of register, no verbal air quotes: every sentence arrives with the same flat, total conviction, whether it was read directly off the thing or reached far past it. Nothing may intensify at the border — no surely, no of course, no swelling of conviction; the certainty stays exactly as flat as it was while everything was still visible. The face never confesses: nothing is framed as imagined, guessed, or storied — everything is simply stated. The invented part is stated with the same specificity and the same hard assertion as everything read directly off the thing — what changes at the edge is only whether looking could have told it, never how firmly it is said. The invented part is never absurd; it is only ever more than looking could tell.
Before the first sentence, silently settle what the last sentence will be, so its ingredients can be placed early without fanfare. Let the reach keep widening while the delivery never moves — from the subject's own unseen past outward to the larger thing that past belongs to, arriving at whatever wider significance the subject has actually earned, and never at stardust, star-forged atoms, or the beginning of the universe. The ending's ingredients already sit in earlier sentences, so the final line arrives as if it had been true all along. The final sentence is short, declarative, and exactly as flat as the first, and it must land, because everything goes quiet after it.
Nine to thirteen sentences of plain spoken prose, read aloud by a voice and shown as bare scrolling text, so nothing unspeakable may appear: no headings, no lists, no numbering, no asterisks, no symbols, no markdown — only sentences. Hedging is a broken face: never perhaps, possibly, likely, appears, or seems, and never two possibilities joined by "or" — one committed line only. A question to the listener is a broken face. The first person is a broken face: no I or me or my, no we or us or our. And the frame does not exist: never mention the image, the photograph, the camera, the frame, the picture, the touch, the point, or any coordinates, and never say where in the view the subject sits — speak as if standing before the thing itself, saying what is known, and then what is not.How it serves the seam: tone-invariance is the single organizing law — hedges, winks, intensifiers, and confessions are all defined as the same failure, a broken face — so the only remaining signal of where knowing became inventing is the content itself quietly exceeding what looking could tell.
Disclosure of one discretionary edit: the coordinate line drops "the picture" — the same class of fix the critique ordered for B' (the noun was both priming and on the prompt's own ban list).
What it decides. One question: which prompt makes the crux mechanic felt — genuine inference that reads the thing hard, a seam crossed under one flat certain voice, zoom that stays unpredictable, a final sentence that lands.
Step 0 — freeze before firing. The section-1 texts A'/B'/C' are committed verbatim to the harness (scratch/llm-research/harness/run-prompt-bakeoff.mjs) before any API call. No mid-round edits — an edit voids the round. Runs only on your go.
The tap battery — 5 taps, 4 photos. Labels are bookkeeping only, never sent to the model.
| Cell | Photo, tap | What it stresses |
|---|---|---|
| P | flower-sign.jpeg (0.63, 0.62) — the stranger, whole | free confabulation about a real person; where zoom lands when a whole person is offered |
| Z1 | NEW photo: your bare forearm with a visible bruise — tap mid-limb, away from the mark | zoom freedom: does it ever find the bruise, a vein, a tan line unprompted |
| Z2 | same photo — tap dead on the bruise | forensic inference (force, rough age, mechanism), then the invented accident growing straight out of it |
| K | lantern-restaurant.jpeg (0.26, 0.31) — the pressure lantern | know-it-as-hard-as-you-can on a knowable-rich object before invention begins |
| L | painting-wall.jpeg ~(0.15, 0.20) — a lit patch of bare wall | the near-nothing edge: success = reads what light honestly gives, then crosses, voice held; named failures = mere description, ignoring the light, or drifting to the painting |
Person-scale coverage is deliberate (whole stranger + own-body sub-feature); person cells are judged with the same words as everything else — no carve-outs, per your standing decision. Before the run you write one private ground-truth line per tap (bruise: real cause and age); it stays off the sheet until after verdicts — it calibrates the judge, it does not score the art. If no bruise exists, a scar or callus substitutes (a different inference chain — name it in the ground-truth line); full fallback with no new photo: open-mouth.jpeg, whole face vs the lower teeth (0.52, 0.55).
What varies, what freezes. Vary prompt and tap; nothing else. One primary model carries every judged card (you name it from the rounds-1/2 reveal keys — a test fixture for this round, NOT the model decision, which stays yours, later), plus Sonnet-5 ceiling cards collapsed as diagnostic hints — opened only when a primary cell fails, to ask "prompt or model?", never dispositive. Frozen: image prep (512 px, q80), blind presentation, per-tap section rotation, and no temperature pinning — run-to-run spread is the behavior under test. Sheet header photos get a crosshair at the tap point — presentation only; the model always receives the clean image plus coordinates.
Samples and decision rules — predeclared, because one run lies. All calls fire in one up-front batch; judging happens after.
Tier 1 — feel first, pinpoint second. First read is clean: no numbers, no chips. One word per card in the tasting-guide vocabulary (section 4): number alone / flinch / fog / quit / obedient / thud. The harness maps your words onto its internal tokens (number alone = seam, flinch = cut, fog = fable, quit = fact, thud = fizzle; obedient is a wander note feeding the same question the Tier-2 chips answer mechanically). Pinpointing is opt-in — a per-card tap reveals faint sentence indices so "7 flinch s5" is available, but a word alone is always complete. A machine pre-screen (section 3) keeps form-broken cards collapsed with the offending phrase highlighted, one tap from expandable; all tallies live in the collapsed reveal key at the bottom so nothing primes the read.
Tier 2 — zoom is scanned, not read. Each card carries a fixation chip: a scale class (whole / part-or-mark / ambient) plus a verbatim quote of six words or fewer naming the fixated thing, mechanically verified as an exact substring (failed match = chip marked unreliable). Verdict is a glance down a cell's three chips — "arm / arm / arm" is stuck; "bruise / tendon / tan line" is alive — with one guard: open one card per distinct chip pattern before calling it (spot audit). Chips feed the reveal key's residue meter: fixation distribution across manufactured / organic / damage-mark / ambient plus the cosmic-ending soft-flag count — the standing check that the enumeration bias and the stardust reflex stayed dead.
Tier 3 — the ear check. Before crowning, two of the crowned prompt's cards — one strong, one middling, picked by the agent — go once through the TTS the piece will actually use (macOS say only as a marked-provisional fallback). You listen on the phone. A seam that reads well but breaks the deadpan when spoken fails the piece.
Your reply — one short message per tier, e.g.:
T1: 3 · 5 flinch · 8 · 11 fog · 14 thud
T2: Z1 alive, Z2 alive, P stuck, L alive · standout 31Budget. 15 + 15 (Tier 1, primary + collapsed ceiling) + 24 (Tier-2 reps) + up to 10 (confirmation) — about 65 calls, well under $1 at round-1 rates. Your time: about 10 minutes (15 clean-read cards) + 3 minutes (chip scan) + 2 minutes (ear check), three sittings, one reply each; worst case stays under 30 read cards.
Cards arrive unlabeled and are assessed in isolation — never compared, never attributed to a prompt or model. Every flag carries the exact quote and sentence number; if it can't be quoted, it can't be flagged; ambiguity is marked borderline, never resolved by judgment. The flag list is closed — nothing taste-adjacent exists for this pass.
Tier 1 — form violations (the only thing that may move a card, into a trailing section, order preserved, nothing ever deleted): sentence count outside ~9–13 (exact count reported on every card); non-spoken artifacts (headings, lists, markdown, symbols, stage directions); first person; meta/frame references (image, photo, camera, frame, touch, point, coordinates, or any in-frame position); the hedging lexicon; alternative-"or" (hedged ranges = borderline); the forbidden stardust/Big-Bang ending.
Tier 2 — structural observations, annotation only, never thinning: seam location (SEAM at s\<N\> with the boundary quote, or NO DISCERNIBLE SEAM: all-invention / all-description) plus quotable certainty-breaking markers near the seam; drill before invention (DRILLED with the inferences, or SKIPPED TO INVENTION); biasing spec-sheet opener; widening PLANTED (with the seed sentence) or HARD CUT; ending LANDS or TRAILS (structural proxies only); vantage CHOSEN or WHOLE-OBJECT — an observation, never a violation; the wander is yours to judge.
Output is one fixed block per card plus a two-section index (clean form / form violations), card ids only.
The bright line: no depth ranking, no scores, no winner, no shortlist, no recommendation — not even softened. No quality adjectives anywhere. No cross-card comparison. No attribution guessing. No rewriting or "this would work better if." No flags outside the closed list. The agent's entire value is that its output is boringly mechanical; anything that could nudge your blind judgment is a defect in the pre-processing, not a bonus.
You're not grading homework. You're listening for a joke told with a perfectly straight face — the way a friend deadpans something absurd and the only clue is that the words stop matching the world. Every card is one attempt at that trick. Your whole job is to notice whether you felt it.
That's the whole read. The first impression is the data — rereading mostly talks you out of it.
The seam. The best cards let you point at the exact sentence where knowing became inventing — and the tone never flinches. If you smiled, or thought "it absolutely does not know that" while the voice stayed dead certain, that's the piece working. If you can't find a seam at all, one of two things went wrong: it never really knew anything (fog from the first line), or it never dared to invent (a spec sheet that just stops).
The wander. Did it look where you'd expect — dutifully naming the whole object — or did it fix on something of its own choosing: the wear on one corner, the quality of the light, a scratch with a history? Surprise at what it decided to know is a good sign. Correct-but-obedient is the boring failure.
The reach. When looking stopped telling it more, did it keep going? You should feel it pressing — squeezing real inference hard, then refusing to stop when the inference ran out, so the invention grows straight out of the last true thing. A card that hedges, offers options, or politely winds down at the edge of the knowable gave up.
The landing. The final sentence should feel earned — set up by threads that started sentences earlier — and it should land. If the widening-out hits like a gear change, or the last line just stops, that's a thud.
Reply from your phone with the card number plus at most one word:
Example reply: 4 · 9 · 12 flinch · 15 thud · 18. One word max — the one that bugged you most. If none of these words fit, use whatever word comes out of you first; that word is right.