Prompt Tier 0: derive risk appetite, pairing and band tiers (T0.1–T0.3) - #39
Merged
Merged
Conversation
) Three things the composer asked a 7B model to work out for itself, now computed deterministically and handed over as values to echo. No new data source, no extra inference round trip, no schema change. - appetite.ts (#24): reads the user's risk wording with a small lexicon. Classification never goes to the model — mixing it into a JSON-emission task costs output validity, and there is one retry. - pairing.ts (#25): the largest mid-consistent pair that fits BOTH ceilings, plus its per-strategy share. The ratio of virtualAmounts IS the shipped price, so a model dividing a budget by a mid across two decimal scales ships an off-mid strategy — free money for the first taker, and nothing in the validator catches it today. Exact fixed-point throughout; shares truncate DOWN, so N of them satisfy I2 by construction. - tiers.ts (#26): risk tiering with no new market data, because on one pair band width IS the risk axis. Three widths from realised volatility, shifted by appetite — the "5/7/9 instead of 3/5/7" behaviour, expressed in band space. Two calibration decisions, stated rather than buried. Volatility is read as ANNUALISED (58 is the canonical ETH figure; the other reading gives ±87%–±99% bands, which are full-range in all but name) — recheck when F3 Q2 lands a real source. And tier labels describe MECHANICS, never yield: we have no fee-APR source, and rating a recommendation is Gate 2's job. Every block is omitted rather than faked when its inputs are missing, and the tier block no longer points at a REFERENCE PAIRING that was not rendered — a dangling reference is an invitation to invent one. PROMPT_VERSION -> sluice.compose/3 (F2 §9). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This was referenced Jul 26, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #24, #25, #26. Tier 0 of the prompt ladder (#29) — the prompt-side half; #38 is the validator-side half.
The principle
The composer model is
qwen/qwen2.5-omni-7bwith two attempts before the deterministic fallback and a user watching a spinner. It has already been caught fabricating chainIds and deadlines until the prompt said "copy these EXACTLY". So: deterministic code computes, the model echoes. Three seams where the prompt was still asking a 7B model to work something out for itself:appetite.ts(#24)pairing.ts(#25)tiers.ts(#26)No new data source, no extra round trip, no schema change, no validator change.
What each one buys
#25 is the one with teeth. The ratio of a strategy's
virtualAmountsis its shipped price. A model that divides wrong ships an off-mid strategy — free money for the first taker — and nothing in the validator catches that today (a price-vs-mid invariant is Tier 2). The plan is the largest mid-consistent pair fitting both ceilings: 2 WETH against 3000 USDC at mid 3450 can only use 0.869565 WETH, because committing both ceilings in full would ship a price that is not the mid. Exact fixed-point throughout, and per-strategy shares truncate down, so N shares satisfy I2 by construction rather than by the model's arithmetic.#26 gets risk tiering with no market data, because on a single pair the risk axis already exists: band width. Tighter quotes deeper (more fill volume) and exhausts on a smaller move —
grammar.tsstates exactly that trade-off. Appetite shifts the whole set, which is the "5/7/9 instead of 3/5/7" behaviour in band space:Two calibration decisions, stated rather than buried
realizedVol7dPct, which does not settle it. 58 is the canonical annualised ETH figure, and the other reading produces ±87%–±99% bands — full-range in all but name, not a market-making recommendation. The two differ by ~3.5x, so this needs confirming when F3 Open Q2 lands a real source (the field may want renaming then). A test pins it: any neutral tier above ±25% means the annualised reading has been lost.Honesty and failure modes
REFERENCE PAIRINGwhen that block was not rendered — a test caught the dangling pointer, which is exactly the sort of thing that invites a 7B model to invent one. It now states the I2 divide-don't-repeat rule instead.contextPromptBlockdoes; both flip together when F3 Q2 settles.PROMPT_VERSION→sluice.compose/3(F2 §9).Verification
Not in this PR
#28 (the baseline-anchoring spike) needs a funded Galileo key to measure anything, so it stays open. The UI still maps
riskfrom band presence infrom-server.ts; wiring tier labels through to the screen is a follow-up, and it should be reconciled with the "risk rating unavailable" decision first.🤖 Generated with Claude Code