Skip to content

Latest commit

 

History

History
44 lines (37 loc) · 4.62 KB

File metadata and controls

44 lines (37 loc) · 4.62 KB

Obstruction Taxonomy

Use this when a proof is stuck. Name the obstruction before trying another route.

Common Obstructions

  • False claim: small counterexample exists or theorem contradicts known boundary case. Next move: refute or add assumptions.
  • Missing assumption: proof needs compactness, convexity, continuity, boundedness, independence, monotonicity, single crossing, or regularity not stated. Next move: state conditional theorem.
  • Wrong theorem family: route proves a related conclusion but not the target. Next move: reclassify with proof-router.md.
  • Quantifier mismatch: pointwise proof used for uniform claim, expectation proof used for high-probability claim, asymptotic proof used for finite-sample claim. Next move: restate target and choose matching concentration/compactness argument.
  • Boundary failure: FOC or derivative argument ignores corner, tie, zero denominator, support endpoint, inactive/active constraint, nonunique optimizer. Next move: KKT/subgradient/case split.
  • Measurability/selection gap: existence of pointwise argmax does not give measurable policy/selector. Next move: measurable maximum theorem or finite approximation.
  • Dependence gap: adaptive data treated as iid, or filtration not defined. Next move: martingale/self-normalized route.
  • Nonconvexity gap: KKT or local optimum used as global optimum. Next move: find convexity/concavity, use global certificate, or search counterexample.
  • Too-strong conclusion: uniqueness, strict monotonicity, gap-free bound, or exact optimality claimed where only weak/set-valued/approximate result holds. Next move: weaken conclusion or strengthen assumptions.
  • Missing summation lemma: regret proof has correct instantaneous bound but cannot sum. Next move: apply peeling, elliptical potential, harmonic/log sum, or Cauchy-Schwarz.
  • Lower-bound construction too easy: instances are too separated or not both feasible. Next move: make instances closer and compute KL explicitly.
  • Overfit pattern guess: formula, potential, active set, or construction fits the first toy cases but fails a holdout case or lacks a certificate. Next move: keep the failed guess as evidence, change the normal form, or search for the missing condition.
  • Opaque algebra: the route depends on a manipulation with no visible sign, telescope, dual, or equality case. Next move: change representation before continuing: gap form, add-subtract benchmark, conjugate/dual, Bellman residual, envelope term, KL/TV bridge, or potential.
  • Low decision value: the next proposed move would not prove/refute a kernel, expose an assumption, retrieve a theorem, or produce a certificate. Next move: choose a higher-value move or stop for theorem repair/user steering.
  • Unchanged proof state: a proof move only renames the same subgoal, missing lemma, or algebraic obstacle. Next move: record the failed state, retrieve a premise/paper pattern, tool-check the local claim, or try theorem repair.
  • Repeated construction: the same central object, parameterization, invariant, or failure witness has appeared under new notation. Next move: update the attempt fingerprint in WORKSTREAMS.md; retry only with a new assumption, invariant, certificate, verified trick, or theorem repair.
  • Repeated route without delta: the same theorem family attacks the same missing lemma with no new proof ingredient. Next move: block the retry in LEDGER.md, then switch route, retrieve a pattern, tool-check the obstruction, or repair the theorem.

Stuck Protocol

  1. Name the obstruction in the ledger.
  2. Write the smallest version where the obstruction appears.
  3. Check whether this is a repeated route or construction by comparing against WORKSTREAMS.md.
  4. Rank the next move by decision value: kernel proof/refutation, counterexample, certificate, retrieval, representation change, or theorem repair.
  5. Try to refute that small version.
  6. If true, promote it to a named lemma and prove it separately.
  7. If false, repair the theorem statement before continuing.

Red Flags

  • "Clearly", "obvious", "by standard arguments" without a named theorem.
  • Differentiating an argmax without envelope/regularity conditions.
  • Swapping limit and expectation without domination/uniform integrability/concentration.
  • Claiming monotone policy from monotone primitives without Topkis/single-crossing conditions.
  • Treating a random stopping/adaptive sample size as deterministic.
  • Proving only for finite support but stating a continuous-type theorem.
  • Renaming the same construction without changing the invariant, certificate, missing assumption, or failure witness.
  • Saying "now prove the remaining lemma" when the remaining lemma is identical to the previous failed subgoal.