#include<algorithm>
#include<cassert>
#include<cmath>
#include<iostream>
using namespace std;
typedef long long ll;
typedef float ld;

int n, r, m, T[52], d[52];
ld p[52];

// E[i][t] = expected time it takes to have performed tricks up to i (incl.)
//           with exactly t seconds lost due to delays.
ld E[52][5001];

ld traceback(ld ans) {
	for (int t=0; t<r; t++)
		E[m+1][t] = n - T[m];

	// O(m (r-n)) < 250'000
	for (int i = m; i >= 1; i--) {
		int dt = T[i] - T[i-1];
		for (int t = 0; t < r; t++) {
			if (t + d[i] < r)
				E[i][t] = dt + p[i] * E[i+1][t] + (1 - p[i]) * min(ans, d[i] + E[i+1][t + d[i]]);
			else
				E[i][t] = dt + p[i] * E[i+1][t] + (1 - p[i]) * ans;
		}
	}
	return E[1][0];
}

ld F(ld x) { return x - traceback(x); }

int main() {
	(cout << fixed).precision(7);

	cin >> n >> r >> m;
	for (int i=1; i <= m; i++) {
		cin >> T[i] >> p[i] >> d[i];
	}

	T[0] = 0;
	r -= n; // we are only interested in the max. allowed trick delay

	// uses Newton iterations, so we also don't need an upper bound on the answer :).

	const ld eps = 1e-6;
	ld x = 0;
	// assert(F(x) < 0);

	// This requires only 5(!) iterations (in fact, teams will probably set this higher):
	for (int rep = 5; rep--; ) {
		ld deriv_at_x = (F(x + eps) - F(x)) / eps;
		x = x - F(x) / deriv_at_x;
	}
	cout << x << endl;
	return 0;
}