Crafter Research · Frontier map

The frontier, classified.

A problem is on this map only if it has an oracle. The axes below say how much each future result can be trusted, what it costs, and why we would be the ones to do it. Rows with no edge stay visible on purpose.

Oracle grade

A is machine-checked with no human judgment. B is statistical, held-out, or prospective. C needs a person or fieldwork.

Tier

Tractable now, hard, or a long bet. Long bets publish their stepping stones.

Compute

Laptop, modest cloud, or out of reach. We do not pretend to have a cluster.

Edge

Place, language, our own instruments, or none.

Crowding

Empty, open community, or a funded race.

State

Catalogued, belief stated, oracle built, survived, killed, published. Killed beliefs stay listed.

Tier
Oracle
Edge

81 / 81 shown

Mathematics

  • M1
    Erdős problems in Lean

    Lean kernel

    • A
    • Hard
    • Laptop
    • No edge
    • Funded race
    • Catalogued
  • M2
    Ramsey number R(5,5), now 43 to 46

    Exhaustive checker

    • A
    • Long bet
    • Out of reach
    • No edge
    • Catalogued
  • M3
    Small off-diagonal Ramsey cells

    Explicit graph and checker

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • M4
    Constant-weight code bounds behind kissing numbers

    The code itself

    • A
    • Tractable now
    • Laptop
    • No edge
    • Open community
    • Catalogued
  • M5
    Kissing numbers in low dimensions

    Explicit configuration

    • A
    • Hard
    • Modest cloud
    • No edge
    • Open community
    • Catalogued
  • M6
    Small matrix-multiplication rank

    Tensor identity check

    • A
    • Hard
    • Modest cloud
    • No edge
    • Open community
    • Catalogued
  • M7
    Record tables: covering designs, packings, Golomb rulers

    The object itself

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • M8
    Kernel-checked certificates for existing table entries

    Lean or Rocq kernel

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • M9
    Counterexamples to last month's arXiv conjectures

    Exhaustive checker

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • M10
    Unknown unknotting numbers and slice status

    Explicit sequence or invariant

    • A
    • Hard
    • Laptop
    • No edge
    • Catalogued
  • M11
    OEIS: new terms, b-file errors, conjectures

    Recompute, Lean for proofs

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued

Theoretical CS

  • C1
    Busy Beaver BB(6) holdouts, one machine at a time

    Rocq or Lean decider

    • A
    • Long bet
    • Laptop
    • No edge
    • Open community
    • Catalogued
  • C2
    Optimal sorting networks, n = 17

    Exhaustive 0-1 check

    • A
    • Hard
    • Modest cloud
    • No edge
    • Open community
    • Catalogued
  • C3
    Minimal circuits for S-boxes

    Exhaustive equivalence

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • C4
    Schur and van der Waerden lower bounds

    The colouring

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued
  • C5
    Agent-evolved SAT heuristics

    SAT Competition set and DRAT proofs

    • A
    • Tractable now
    • Modest cloud
    • No edge
    • Funded race
    • Catalogued
  • C6
    Weakly solving small games

    Proof-tree checker

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued

Systems and developer tools

  • A1
    Verified superoptimization of hot loops

    SMT equivalence, differential fuzz, benchmark

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • A2
    Cross-engine differential testing of the web platform

    Engine disagreement and a reproducer

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • A3
    Terminal emulator conformance

    Cross-emulator disagreement

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • A4
    Compiler miscompiles and missed optimizations

    Alive2 and crash reproducer

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • A5
    Database logic bugs by metamorphic testing

    Differential result

    • A
    • Tractable now
    • Laptop
    • No edge
    • Funded race
    • Catalogued
  • A6
    Byte-matching decompilation

    Byte-identical rebuild

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • A7
    Fixing non-reproducible builds

    Identical hash

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • A8
    Rust standard library verification

    Kani or Verus

    • A
    • Hard
    • Laptop
    • No edge
    • Catalogued
  • A9
    Hutter Prize text compression

    Public leaderboard, CPU rules

    • A
    • Hard
    • Laptop
    • No edge
    • Catalogued
  • A10
    Which CLI designs agents operate reliably

    Held-out task success across models

    • B
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • A11
    C to Rust translation with equivalence

    Tests and differential fuzz

    • A
    • Hard
    • Laptop
    • No edge
    • Funded race
    • Catalogued

Security

  • S1
    Autonomous vulnerability discovery in open source

    Sanitizer crash reproducer

    • A
    • Tractable now
    • Modest cloud
    • No edge
    • Funded race
    • Catalogued
  • S2
    The unaudited LatAm long tail: e-invoicing, tax, payments

    Reproducer and coordinated disclosure

    • A
    • Tractable now
    • Laptop
    • Spanish and Quechua
    • Empty
    • Catalogued
  • S3
    Verified patches

    PoC dies, tests and fuzz pass

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • S4
    Prompt injection and exfiltration in browser agents

    Canary exfiltration reproducer

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • S5
    Protocol state machines with model checking

    TLA+ or Tamarin counterexample

    • A
    • Hard
    • Laptop
    • No edge
    • Catalogued
  • S6
    Predicting npm supply-chain malware

    Later advisory or takedown

    • B
    • Tractable now
    • Modest cloud
    • No edge
    • Funded race
    • Catalogued
  • S7
    Under-constrained zero-knowledge circuits

    Two valid witnesses

    • A
    • Hard
    • Laptop
    • No edge
    • Catalogued

Cryptanalysis and chains

  • K2
    Mid-dimension lattice records via better algorithms

    SVP Challenge leaderboard

    • A
    • Hard
    • Modest cloud
    • No edge
    • Funded race
    • Catalogued
  • K3
    Automated trail search across lightweight ciphers

    Solver-verified trail, practical attack

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued
  • K4
    Keccak reduced-round challenges

    The collision or preimage

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued
  • K5
    Encrypted dispatches of the Viceroyalty of Peru

    Recovered key decrypts held-out letters

    • A
    • Hard
    • Laptop
    • Spanish and Quechua
    • Empty
    • Catalogued
  • K6
    Constant-time checks of post-quantum implementations

    dudect and ctgrind

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued
  • B1
    Spanish-language crypto Ponzi schemes before the regulator's alert

    Later regulator alerts

    • B
    • Hard
    • Modest cloud
    • Spanish and Quechua
    • Empty
    • Catalogued
  • B2
    Wash trading and MEV measurement

    Labeled sets

    • B
    • Tractable now
    • Modest cloud
    • No edge
    • Funded race
    • Catalogued

Physics

  • P1
    New periodic three-body orbits

    Numerical integrator

    • A
    • Tractable now
    • Laptop
    • No edge
    • Open community
    • Catalogued
  • P3
    Singularity candidates in Euler and Boussinesq flows

    Interval-arithmetic proof

    • A
    • Long bet
    • Modest cloud
    • No edge
    • Funded race
    • Catalogued
  • P4
    Audit AI-discovered quantum codes

    A lower-weight logical operator

    • A
    • Hard
    • Laptop
    • No edge
    • Empty
    • Belief stated
  • P5
    Quantum circuit T-count records

    Equivalence check

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued
  • P6
    Variational ground-state energies

    Lower energy wins

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued
  • P7
    Best-known Max-Cut, spin-glass, and packing instances

    The configuration

    • A
    • Tractable now
    • Laptop
    • No edge
    • Catalogued

Earth science

  • E1
    Peru's deep seismic catalogue

    Re-detection at independent stations

    • B
    • Tractable now
    • Modest cloud
    • Peru and the Andes
    • Empty
    • Belief stated
  • E2
    Slow slip and tremor on the Peruvian flat slab

    Template stacking

    • B
    • Hard
    • Modest cloud
    • Peru and the Andes
    • Catalogued
  • E3
    Coastal El Niño forecasting

    Hindcast and prospective skill

    • B
    • Hard
    • Modest cloud
    • Peru and the Andes
    • Catalogued
  • E4
    AI weather models against Andean stations

    Held-out station data

    • B
    • Tractable now
    • Modest cloud
    • Peru and the Andes
    • Catalogued
  • E5
    Glacial lake outburst risk in the Cordillera Blanca

    Held-out inventories

    • B
    • Hard
    • Modest cloud
    • Peru and the Andes
    • Catalogued
  • E6
    Anchoveta biomass forecasting

    Prospective surveys and quotas

    • B
    • Hard
    • Laptop
    • Peru and the Andes
    • Catalogued
  • E7
    Dengue forecasting by department

    Prospective weekly counts

    • B
    • Hard
    • Laptop
    • Peru and the Andes
    • Catalogued
  • E8
    Illegal mining and deforestation detection

    Held-out labels and later imagery

    • B
    • Tractable now
    • Modest cloud
    • Peru and the Andes
    • Funded race
    • Catalogued

Astronomy

  • R1
    Anomalies in the Rubin alert stream

    Cross-match and follow-up

    • C
    • Hard
    • Modest cloud
    • Peru and the Andes
    • Catalogued
  • R2
    Asteroid discovery from archival detections

    Minor Planet Center designation

    • C
    • Hard
    • Modest cloud
    • No edge
    • Catalogued

Decipherment and language

  • D1
    Khipu: predict held-out cords

    Held-out khipus against a null

    • B
    • Hard
    • Laptop
    • Peru and the Andes
    • Empty
    • Belief stated
  • D2
    Khipu digitization from photographs

    Agreement with expert records

    • B
    • Hard
    • Modest cloud
    • Peru and the Andes
    • Empty
    • Catalogued
  • D3
    Linking khipus to colonial records

    Numeric match against a null

    • B
    • Long bet
    • Laptop
    • Peru and the Andes
    • Catalogued
  • D4
    Reading colonial notarial handwriting

    Character error on held-out pages

    • B
    • Hard
    • Modest cloud
    • Spanish and Quechua
    • Catalogued
  • D5
    Reconstructing proto-Quechua forms

    Leave-one-language-out prediction

    • B
    • Hard
    • Laptop
    • Spanish and Quechua
    • Catalogued
  • D6
    Quechua and Aymara language models

    AmericasNLP leaderboards

    • B
    • Hard
    • Modest cloud
    • Spanish and Quechua
    • Catalogued
  • D7
    Accounting scripts with sum checks: Linear A, Proto-Elamite

    Sums as held-out checks

    • B
    • Long bet
    • Laptop
    • No edge
    • Catalogued
  • D8
    Vesuvius Challenge scrolls

    Prize review

    • C
    • Hard
    • Modest cloud
    • No edge
    • Open community
    • Catalogued
  • D9
    Archaeological site detection in Peru

    Held-out registry sites and field checks

    • C
    • Hard
    • Modest cloud
    • Peru and the Andes
    • Catalogued

Science of AI

  • I1
    False negatives in agent verdicts

    Planted ground truth

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Empty
    • Belief stated
  • I2
    Claim-to-artifact audit of AI-written papers

    Re-execution

    • A
    • Tractable now
    • Modest cloud
    • Our instruments
    • Catalogued
  • I3
    Broken or leaky tasks in agent benchmarks

    Reproducer per task

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • I4
    Same model, different harness

    Fixed task suite

    • B
    • Tractable now
    • Laptop
    • Our instruments
    • Catalogued
  • I5
    Interpretability on tiny models with planted circuits

    Causal intervention against known truth

    • A
    • Hard
    • Modest cloud
    • No edge
    • Catalogued
  • I6
    AI scientists on messy Spanish-language data

    Reproduction on public microdata

    • B
    • Hard
    • Laptop
    • Spanish and Quechua
    • Empty
    • Catalogued
  • I7
    Arithmetic-impossibility checks in regional journals

    Arithmetic

    • A
    • Tractable now
    • Laptop
    • Spanish and Quechua
    • Catalogued
  • I8
    Verified results per token, published as a curve

    The lab's own oracles

    • A
    • Tractable now
    • Laptop
    • Our instruments
    • Empty
    • Catalogued

Social science

  • X1
    Procurement risk with a temporal holdout

    Later addenda, terminations, sanctions

    • B
    • Hard
    • Laptop
    • Peru and the Andes
    • Catalogued
  • X2
    Forecasts of official statistical releases

    The release itself

    • B
    • Tractable now
    • Laptop
    • Peru and the Andes
    • Empty
    • Belief stated
  • X3
    Specification curves for Peruvian policy claims

    Re-execution on public microdata

    • B
    • Hard
    • Laptop
    • Peru and the Andes
    • Catalogued
  • X4
    Mass replication of LatAm policy papers

    Re-execution

    • B
    • Hard
    • Laptop
    • Peru and the Andes
    • Empty
    • Catalogued
  • X5
    Models of Congress from votes and debate text

    Held-out votes

    • B
    • Tractable now
    • Laptop
    • Peru and the Andes
    • Catalogued
  • X6
    Synthetic respondents against real Peruvian surveys

    Held-out survey marginals

    • B
    • Tractable now
    • Laptop
    • Peru and the Andes
    • Catalogued

Where we do not compete

Wet labs, materials and DFT, model training, drug discovery, and headline races on formal math already funded by large labs. We contribute pieces there only when an open community makes it useful.