symbolic-check

name	symbolic-check
description	Use SymPy to prove or refute a self-authored algebraic identity, derivative, limit, comparative-static sign, or closed form. Use when exact symbolic manipulation can settle the claim. For parameter sweeps or full theorem proving, use $numerical-check or $lean-check.
allowed-tools	["Read","Write","Edit","Bash","AskUserQuestion"]

Symbolic Check: Prove/Refute a Self-Authored Algebra Step with a CAS

Verify a symbolic manipulation you wrote — an identity, a derivative, a limit, a comparative-static sign, a closed form — using sympy. Unlike numerical falsification, this can positively verify the step: a CAS-confirmed identity is correct.

When to Use

• You wrote A = B, ∂f/∂x = g, lim = L, sign(∂f/∂x) = −, or "the closed form is …" and want it proven before it ships.
• symbolic-check, "verify this algebra / derivative / limit", "check the comparative-static sign", "does this closed form equal the original".
• Companion to mark-unverified for self-authored algebra (the derivative-sign / closed-form family the rule explicitly names).

When NOT to Use

Situation	Use instead
A full theorem/lemma you want machine-proven end-to-end	lean-check (R3)
A distributional / probabilistic claim over a parameter space	numerical-check (R1)
Re-verify a computed empirical result	cross-language-check
Conceptual / assumption review	domain-reviewer

Position in the verification spectrum

R2 — symbolic / CAS. Can VERIFY (prove) or FALSIFY a symbolic step; between numerical falsification (R1) and formal proof (R3) in strength. It proves algebra, not arbitrary theorems — reasoning beyond symbolic manipulation (measure theory, limits sympy can't evaluate) escalates to lean-check or domain-reviewer.

Procedure

1. Transcribe the claim precisely, with declared symbol domains

• Restate the exact claim: identity A == B, derivative diff(f,x) == g, limit limit(f,x,a) == L, sign sign(diff(f,x)) over a domain, or closed form expr == cf.
• Declare assumptions on the symbols — symbols('x', positive=True, real=True) etc. Comparative-static signs and simplifications are wrong without the right domain. State them explicitly (they are part of the claim).

2. Prove the core with .equals(), not simplify(...)==0

sympy's simplify is heuristic — a non-zero result does not mean the claim is false, only that simplify gave up. Use (A - B).equals(0), which combines symbolic + random-point numerical testing and returns:

• True → VERIFIED (identity holds)
• False → FALSIFIED (a witness point disproves it)
• None → INCONCLUSIVE (undecided) — go to step 3.

For derivatives: diff(f, x).equals(g). For limits: limit(f, x, a) and compare to L. For a closed form: expr.equals(cf).

3. On None, escalate — do NOT guess

• Try stronger simplifiers targeted at the form: factor, radsimp, trigsimp, powsimp, together, rewrite(...), assuming(...) with the domain.
• Numerically substitute several random in-domain points into A - B — if all ≈ 0, report INCONCLUSIVE (numerically consistent, symbolically undecided); if any is far from 0, that's a FALSIFIED witness.
• Never upgrade None to VERIFIED. Undecided is undecided.

4. Comparative-static / monotonicity signs

• Compute d = diff(f, x). Ask whether d has a definite sign under the assumptions.
• Try refine(d > 0, Q.positive(...)) / ask(Q.negative(d), assumptions); if sympy can't decide, sample the domain numerically to conjecture the sign, then report INCONCLUSIVE (sign consistent on N points) — a sign you can't prove symbolically is a candidate for numerical-check (falsify) or lean-check (prove).

5. Belt-and-suspenders

Even on a .equals() == True, do a quick numerical substitution at one random point as a sanity check against a symbol/transcription bug. A transcription error is the most common real failure.

6. Emit the verification report

Script skeleton (adapt)

# uv run --no-project --with sympy python <script>.py
import sympy as sp
mu, mumax, rho = sp.symbols('mu mumax rho', positive=True)   # DECLARE the domain
# --- identity / closed-form ---
A = mu / sp.sqrt(rho); B = mumax                              # claim: threshold solves A == B
rho_star = sp.solve(sp.Eq(A, B), rho)                        # -> [mu**2/mumax**2]
print("rho* =", rho_star, " expected (mu/mumax)**2:", (mu/mumax)**2)
print("identity holds:", (rho_star[0] - (mu/mumax)**2).equals(0))   # True / False / None
# --- derivative sign (monotonicity) ---
Phi = lambda z: (1 + sp.erf(z/sp.sqrt(2)))/2                 # standard normal CDF
Q = Phi(mu/sp.sqrt(rho)); d = sp.diff(Q, rho)
print("dQ/drho =", sp.simplify(d), " sign<0 on domain (mu>0):", sp.ask(sp.Q.negative(d), sp.Q.positive(mu)))

Anti-Patterns

• Don't read simplify(A - B) != 0 as FALSIFIED — that's simplify giving up. Use .equals().
• Don't upgrade a None (undecided) to VERIFIED. Report INCONCLUSIVE.
• Don't verify a sign/closed-form without declaring the symbol assumptions — the domain is part of the claim, and the wrong domain gives the wrong answer.
• Don't skip the numerical sanity substitution — it catches transcription bugs a symbolic pass can hide.
• Don't use bare python3 — use uv run --no-project --with sympy python.
• Don't force a hard theorem through the CAS — if it needs real reasoning (not symbolic manipulation), escalate to lean-check (R3) or domain-reviewer.

Output — Verification Report (shared *-check shape)

Write to reviews/<scope>/verify-symbolic/<YYYY-MM-DD-HHMM>.md:

claim:    <exact symbolic statement + declared symbol domains>
method:   R2 symbolic/CAS (sympy .equals() [+ numeric sanity at <k> points])
verdict:  VERIFIED | FALSIFIED | INCONCLUSIVE (numerically consistent, symbolically undecided) | ERROR
evidence: <simplified form / witness point that disproves / the derived closed form>
reproduce: uv run --no-project --with sympy python experiments/<script>.py

Verification (did this skill work?)

• Verdict is one of the four; VERIFIED is emitted only on .equals() == True (or an exact solve/limit match), never on None.
• A numerical sanity substitution accompanies every VERIFIED.
• If feeding a paper, the confirmed closed form / sign is the value that goes in (and any derived constant is emitted via a macro, no-hardcoded-results).

Worked example — 2026-07-04 (median-collapse paper)

• Closed-form threshold: claim ρ* = (μ_med/μ_max)² solves Φ(μ_med/√ρ) = Φ(μ_max). solve(μ_med/√ρ = μ_max, ρ) → μ_med²/μ_max²; .equals() → VERIFIED.
• Homogeneous monotonicity (Prop 3.1): d/dρ Φ(μ/√ρ) = φ(μ/√ρ)·μ·(−½ρ^{-3/2}), negative for μ>0 → sign VERIFIED under Q.positive(mu) (numeric-confirmed on the domain where sympy hesitates).
• These are the algebra rungs beneath the theorems numerical-check stress-tested and lean-check could formalize.