ABSTRACT ALMA observations of z > 6 galaxies have revealed abnormally high [O iii]88 $\mu$m/[C ii]158 $\mu$m ratios and [C ii]158 $\mu$m deficits compared to local galaxies. The origin of this behaviour is unknown. Numerous solutions have been proposed including differences in C and O abundance ratios, observational bias, and differences in ISM properties, including ionization parameter, gas density, or photodissociation region (PDR) covering fraction. In order to elucidate the underlying physics that drives this high-redshift phenomenon, we employ sphinx20, a state-of-the-art, cosmological radiation–hydrodynamics simulation, that resolves detailed ISM properties of thousands of galaxies in the epoch of reionization which has been post-processed with cloudy to predict emission lines. We find that the observed z > 6 [O iii]88 $\mu$m–SFR and [C ii]158 $\mu$m–SFR relations can only be reproduced when the C/O abundance ratio is ∼8 × lower than Solar and the total metal production is ∼4 × higher than that of a Kroupa IMF. This implies that high-redshift galaxies are potentially primarily enriched by low-metallicity core–collapse supernovae with a more top-heavy IMF. As AGB stars and type-Ia supernova begin to contribute to the galaxy metallicity, both the [C ii]158 $\mu$m–SFR and [C ii]158 $\mu$m luminosity functions are predicted to converge to observed values at z ∼ 4.5. While we demonstrate that ionization parameter, LyC escape fraction, ISM gas density, and CMB attenuation all drive galaxies towards higher [O iii]88 $\mu$m/[C ii]158 $\mu$m, observed values at z > 6 can only be reproduced with substantially lower C/O abundances compared to Solar. The combination of [C ii]158 $\mu$m and [O iii]88 $\mu$m can be used to predict the values of ionization parameter, ISM gas density, and LyC escape fraction and we provide estimates of these quantities for nine observed z > 6 galaxies. Finally, we demonstrate that [O i]63 $\mu$m can be used as a replacement for [C ii]158 $\mu$ m in high-redshift galaxies where [C ii]158 $\mu$ m is unobserved and argue that more observation time should be used to target [O i]63 $\mu$m at z > 6. Future simulations will be needed to self-consistently address the numerous uncertainties surrounding a varying IMF at high redshift and the associated metal returns.