06-reference/research

quiver art pipeline cost reduction

2026-10-19·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
scribble-worksquiver-arrow-2unit-economicsspending-authorityart-pipeline

Prompt and caching are the two weakest levers available: output tokens are ~96% of a Quiver call, Arrow's API has no cache parameter, and the gap already closed itself through local asset reuse

The question

Verbatim: "What prompt or caching changes to the Quiver art pipeline would bring Scribble Works' AI-art spend back under the $20/mo autonomous-spend ceiling?"

Context: the 2026-10-09 unit-economics review projected studio art at about $71/month against the $20/month autonomous line in [[project_rdco_spending_authority]], and asked whether batching, caching or prompt-shortening could close the gap. That projection has since been superseded by measurement, which changes the answer.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The question asks about the two worst levers on the board, and the arithmetic says so before taste does. Prompt shortening touches input tokens only. Input is a flat ~722 tokens per call, $0.0029 at the published $4/M, against a measured production mean of about $0.080 per call: 3.6% of the bill. Deleting the entire prompt would save about $1.06 on a $29.51 month. Even allowing that production prompts run three times longer than the bake-off's, input tops out near 11%, so halving prompt text buys 2-5%. Caching is worse than weak, it is absent: the Arrow generations endpoint accepts seven body fields and rejects unknown ones with a 400, so there is no cache parameter to set, and if there were it would act on the same 3.6%. The levers that matter all act on output tokens, which are the SVG's own path data plus reasoning billed as output, and none of them is a prompt or a cache.

What actually closed the gap was local asset reuse, and it closed it harder than any prompt change could. Per-round art went $2.37 (10-01 to 10-08) to $0.89 (10-09 to 10-16) to $0.00 on both 10-17 and 10-18, with 4/4 gate PASS on each of those two nights. That is a 100% reduction on new-game art, achieved by the round brief's standing order to search content/art and existing game SVGs before generating, plus ruling 20's code-geometry default for grids, mazes and tables. Two structural notes keep this honest. First, the mechanism is not Quiver's search_library, which returns empty for this account because the REST path persists no Artifacts; it is grepping 136 inlined game.html files. Second, the catalog that would make it systematic, PR #428's 113-drawing catalog with a duplicate gate and art:find, is still an OPEN DRAFT, so today's reuse depends on byte-copying from unmerged pull-request heads, which is what generates the merge-order coupling and the R53 rework list. The lever is proven; its plumbing is not merged.

The residual driver is refresh rework, which is a merge-throughput problem wearing a cost problem's clothes. Refreshes measured $2.28 (#368) and about $1.62 (#383), five to seven times a new game, and queue.md:796 attaches one to every game pull request older than #425. The 10-16 review's own pace arithmetic is the tell: about $43 by 10-31 on new games alone, about $53-57 once six refreshes land. Those refreshes exist because 0 of 21 game pull requests merged and main last moved on 10-11, blocked on the production D1 catalog token open since 10-11. So the cheapest dollar in this whole analysis is not a prompt edit; it is the founder's decision on the merge-as-hidden-pending-review proposal and the D1 token, which retires the refresh rule's trigger entirely.

One structural finding deserves separate weight, because it could make the metering question moot rather than optimised. QuiverAI runs two disjoint pools: subscription plans carry included weekly app credits (Go $5, Plus $15, Pro $40, resetting Mondays UTC with no rollover), while "API calls use the API Platform prepaid balance, not weekly or prepaid App credits." Studio builders fetch QUIVERAI_API_KEY through op-service-account.sh and hit the REST generations endpoint, so Scribble Works art is drawing the API prepaid balance, and the subscription's weekly app allowance is being paid for and left unused. Go at $8/month includes roughly $21.70/month of app usage and is comfortably autonomous; Plus at $20/month includes roughly $65/month but sits exactly at the founder-gated threshold. If the MCP send_prompt surface debits app credits rather than the API balance, then a flat $8/month plan could absorb the studio's entire measured art volume and the variable-cost problem becomes a fixed-cost line. The supporting evidence is suggestive rather than conclusive: MCP returned insufficient_credits on 09-19 while REST worked, which is what an exhausted weekly no-rollover pool looks like, and today's list_models shows generate with allowOnDemandCredits: false and state: "ok". I am labelling this an inference, not a finding, because no page reached documents which pool MCP debits and the billing endpoint 403s.

Recommended configuration, with the arithmetic

Baseline, imputed: about $29.51 October-to-10-16; measured $0.34 per new game since 10-09; measured $0.080 per call over the 54 calls logged across #434 and #437-#446; about 2.6 new games per round, roughly 79 per month.

# Change Measured effect Confidence
1 Keep arrow-2; never arrow-2-telos $0 saved, prevents a 1.5x regression High (vendor rates; Telos unused today)
2 Merge #428; make art:find a mandatory pre-step before any generate call $0.89/round to $0.00 on 10-17 and 10-18, gate 4/4 both nights High (measured, two rounds)
3 Cap new drawings at 3 per game Heavy games were #434 (11 calls, $0.74) and #438 (9, $0.81), both pre-pass driven; 3 calls = ~$0.24 Medium
4 In the blind legibility pre-pass, swap from catalog before regenerating #439 did exactly this (sun swapped #400 to #398): 1 call, $0.07 for the whole game Medium-high
5 Extend ruling 20 code geometry to every derivable element #448, #451, #452 shipped at $0 High (founder ruling + measured)
6 One-call icon sheet only above 12 subjects ~3x cheaper per icon, threshold almost never reached in studio games Medium
7 Cap refreshes at 3/month pending the merge decision $1.62-$2.28 each measured High on unit cost, medium on count
8 Lower reasoning_effort one step Unknown; untested in every run on record Low on magnitude

Path to under $20, stated as three different numbers because they are three different things:

The binding constraint is therefore refresh count, not prompt or cache design. Items 2 through 5 are already sufficient on new-game art; items 7 and the merge decision are what decide whether the month lands under the line.

Unverified assumptions in that arithmetic, named explicitly:

  1. Every dollar is imputed (provider-reported tokens x list price) because the balance endpoint 403s for want of billing_read. Nothing is reconciled to an invoice.
  2. 79 new games/month is extrapolated from eight days at 2.6/round; the slot count while merges stall is an open founder decision.
  3. $0.163/game assumes the catalog hit rate holds. New subject classes force generate calls, and bands 2_3, 7_8 and 9_10 are untouched, so the rate cannot be assumed stable.
  4. $0.080/call is the production mean for isolated icons. Scene-class or high-detail art measured $0.219-$0.426 in the bake-off, three to five times higher, so any style shift invalidates the line.
  5. Three refreshes/month is a proposed cap, not a measured rate. The live queue.md:796 rule implies more.
  6. Whether the subscription price itself counts against the $20 ceiling is undecided. If Go at $8/month is already billed, the API budget is $12/month and even the $12.88 line is over.

Quality risk, ranked by cost saved per unit of risk

  1. Catalog reuse (best ratio). Saves ~$0.89/round at near-zero risk, because the drawing already passed critic and founder review. The real risk is sameness and blast radius: the turtle shell that "reads as a soccer ball" now appears in #444, #445 and #447 (rounds.md:990, :1003), so one weak asset propagates. Cheap fix, redraw the one asset once.
  2. Code geometry for derivable elements. $0 marginal, deterministic, founder-ruled on-contract. Hard boundary: geometry only. Pushing it into illustration is the 2026-10-01 regression.
  3. Capping calls per game at 3. No measured link between drawing count and gate outcome; #443 used 8 drawings and still drew a "page a bit sparse" note. Count is not the quality driver.
  4. Swap before regenerate in the legibility pre-pass. Low-moderate risk: a swapped drawing fits the slot's composition less exactly than one drawn for it.
  5. Lowering reasoning_effort. Unknown saving, and the highest chance the founder notices. Output tokens are path count, and the bake-off's structural audit ties path count to style: 40 paths for the flat corgi against 246 for the doodle probe and 300 for the engraving probe. Cutting reasoning plausibly cuts exactly the density that reads as hand-drawn, on an axis where Arrow already under-delivers (the engraving probe scored 3/5 charm and "only gestures at engraving"). That is the founder's single strongest stated taste ([[feedback_design_taste_high_personality]]). Test blind before adopting.
  6. Simplifying art style to cut output tokens. Biggest per-call saving, three to five times, and the worst idea here: it is the flat/minimal direction the founder has explicitly rejected. Do not offer it as a cost measure.
  7. Prompt shortening (worst ratio). Bounded at 2-5%, and it carries real quality risk, because the 10-18 framework's law is that Arrow obeys semantic, content-level instructions and ignores global ones. The prompt is the only working control surface, so shortening it trades the one thing that steers the drawing for a rounding error on the bill.

Why this is in the vault

It answers decision 1 carried in escalations.md since 2026-10-09 and re-raised 10-16 ("Monthly Quiver ceiling?") with measured per-call and per-game numbers instead of the stale ~$71 projection, and it tells the founder that the ceiling decision he is being asked for is really two different decisions: a per-game art cap that items 2-5 already satisfy, and a refresh-count cap that depends on the unanswered D1 token and merge-as-hidden-pending-review proposals. It also redirects queue item #396/#428 from "art reuse would be nice" to "art reuse is the only lever that measurably worked," and it kills prompt-shortening and prompt-caching as queue candidates with arithmetic rather than opinion.

Open follow-ups

Related

Sources

Vault

Primary operational state (read directly 2026-10-19)

Vendor / MCP (retrieved 2026-10-19)