A transfer matrix scaling technique is developed for randomly diluted systems, and applied to the site-diluted Ising model on a square lattice in two dimensions. For each allowed disorder configuration between two adjacent columns, the contribution of the respective transfer matrix to the decay of correlations is considered only as far as the ratio of its two largest eigenvalues, allowing an economical calculation of a configuration-averaged correlation length. Standard phenomenological-renormalisation procedures are then used to analyse aspects of the phase boundary which are difficult to assess accurately by alternative methods. For magnetic site concentration p close to p c , the extent of exponential behaviour of the T c × p curve is clearly seen for over two decades of variation of p − p c . Close to the pure-system limit, the exactly-known reduced slope is reproduced to a very good approximation, though with non-monotonic convergence. The averaged correlation lengths are inserted into the exponent-amplitude relationship predicted by conformal invariance to hold at criticality. The resulting exponent η remains near the pure value (1/4) for all intermediate concentrations until it crosses over to the percolation value at the threshold.PACS numbers: 75. 10H, 75.40c, 05.50 Typeset using REVT E X 1 The transfer matrix is well known to provide exact solutions for a number of lowdimensional pure systems such as spin models etc [1]. Applied to finite-width strips and combined with finite size scaling [2], it has also proved to be extremely powerful and accurate for non-solvable pure cases, ranging from magnetic systems [3] to walks [4] and quantum Hamiltonians [5]. The application of such "strip-scaling" techniques to random systems has however been extremely limited, despite the enormous interest in such problems as the spin glass, random field and dilute magnets, ceramic superconductors etc. This is because the form so far utilized [6][7][8][9] applies the strip scaling to particular realisations of the random system which need not be representative unless appropriate averaging (or self-averaging) is employed, resulting in very large scale computing on extremely long strips and/or many realisations.Here, we provide a strip scaling approach for random systems, in which the disorder averaging is carried out as one proceeds along the strip, and which therefore does not have the deficiencies noted above. Results reported here extend, and give details of, those contained in a previous Rapid Communication [10]. Our scheme does rely on certain assumptions regarding the dominant contributions to the decay of correlations, which must be spelt out clearly and supported by numerical evidence whenever possible. This is one of our purposes in what follows.The main numerical application of the technique here is to the two-dimensional sitediluted Ising model [11]. The transfer-matrix descriptions of the special limiting cases of percolation [12] and pure Ising system [13] are well-understood, and are reproduced by t...