agentleFS
odinfreeodinfree/storm-ecdsa-harness/skills/lambda-exchange-desk/SKILL.md

lambda-exchange-desk

Skill4 starsChanged 21 days agoMITView on GitHub

Price statistical trades in an approximate-circuit campaign as an explicit exchange market: fix a Toffoli-per-lambda rate, measure every trade paired against its parent with a standard error, charge conservative floors, scan for buybacks priced below the rate, and guard census-fitted tables against in-sample overfitting. Use whenever an ecdsa.fail-style circuit trades executed Toffoli for failure rate before a nonce grind, or any scored artifact trades score for a statistical defect rate.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every trade reports dT, dλ, and the paired SE against its own parent, same shots, same seeds.

    line 29
  2. Distinctive:

    After every accepted stack, re-price every open position in both directions.

    line 41
  3. Distinctive:

    EVERY row (all 476 rows ≥3 chain units free.

    line 65

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Wächselstube (lambda-exchange-desk)

Distilled from haotang1995's two record submissions (2026-10-29), where a stack of approximate trades was run as a priced market: rate 500 T per unit lambda, every trade measured, some trades reverted when the market moved.

Fix the rate first

  • Declare the exchange rate (T per unit λ) before evaluating any approximate edit. Derive it from nonce economics: at failure rate λ per shot set, one clean island costs ~e^λ trials; λ is expensive in GPU-hours, not just in score.
  • Trades are approved on effective dT: measured dT minus rate × max(dλ + SE, floor). Their floor: a minimum three-event charge against the Toffoli delta.

Measure paired, or not at all

  • Every trade reports dT, dλ, and the paired SE against its own parent, same shots, same seeds. Unpaired comparisons against a distant leader are marketing.
  • Verifier-level SE (e.g. ±0.31 λ at 1M shots) cannot resolve small trades; the worker must run 4M–16M-shot paired runs to resolve its own layer.
  • Interactions between layers are measured only through whole-circuit runs — never assume additivity.

Run the buyback scan

  • Prices move. Their choice-window pass sold λ at ~500 T/λ on ticks 1–119, then a later census showed ticks 81–112 had overpaid: reverting them bought λ back at ~325 T/λ (+96 T, −0.295 λ) — below rate, so the revert is pure profit.
  • After every accepted stack, re-price every open position in both directions. A revert that buys λ below the rate, or a new sell well above it, is a trade.

Census-fitted tables: out-of-sample or reject

  • Their per-block width table v1 looked +368 T effective in sample and only +212 out of sample — a 42% overfit. v2 used more shots, a conservative estimator, and content-keyed rows.
  • Rules: fit on a training census (e.g. 8M shots), validate on disjoint shots, key every row by content hash (FNV-1a), fail closed on any mismatch, and keep the uniform default for unlisted blocks.
  • Selection rules need floors: their repair census required ≥5 repaired failures, zero added failures, and zero ancilla failures per admitted repair.

Discovery-model quarantine

A classical/ideal model may guide discovery but is never evidence: it does not model quantum phases or width-envelope failures. Their calibration: walk-model predictions overestimated realized circuit gains 1.5–2×. Price from paired shots, not from the model.

Measured case: a reserve can be priced out, not just exhausted (our own, 2026-10-31)

On the 5ba2d2d4 record tree, a structural census of the choice_narrow compare-narrowing table found room in EVERY row (all 476 rows ≥3 chain units free; 416 ≥8): the naive read says "-476 T per +1 deepening sweep". The sweep measured -367.0 T executed (dropped units execute ~77%) but the official eval FAILED with 20 classical + 17 phase-garbage failures on the fresh 9,024-shot set — dλ ≈ +2,216 ppm for -367 T ≈ 0.17 T/ppm, ~3,000x below the 500 T/ppm rate. Tie-model calibration gave q = P(overflow|tie) ≈ 0.0042, and even the cheapest decile projects ~95 T/ppm. The reserve was real and unprofitable: the team stopped there because their census already priced it, not because they ran out of rows. Lessons:

  • Structural room is not a prize; only a paired price is. Run the room census AND one decisive sweep before claiming a reserve.
  • λ units in this economy are ppm of failure probability; the 19.5 limit and the 500 T/ppm rate both come from island re-grind economics (a submission/stream with true rate λ fails its 9,024-shot eval with probability ~1-e^{-λ·9024} and re-grinds cost e^{λ·9024} retries). Sub-ppm trades are invisible to one official eval — 4M-16M-shot paired censuses are the resolution, and a coarse 9,024-shot sweep still bounds aggregate λ when failures actually occur (failures/9024).
  • A fresh Fiat-Shamir shot set means any stream-changing sweep is an unpaired measurement; fine for coarse gates, never for desk-grade pricing.
  • Bisect before concluding: a d<=2-only subset sweep (448 rows) still failed 18/9024 for -335.1 T — the λ was broad-based, not a hot tail. The uniform tie-probability model (2^{-(L-d)}) underestimated the realized rate ~1,000x: walk registers are strongly correlated, ties are common, and ~every tie breaks downstream (q ~ 1 on this tree). Depth-prior ranking would have predicted a cheap core that does not exist.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.