CAPERS-LRD-z9: A Gas-enshrouded Little Red Dot Hosting a Broad-line Active Galactic Nucleus at z = 9.288

CAPERS is a JWST Cycle 3 legacy program that uses NIRSpec/Micro-shutter Assembly (MSA)/PRISM to observe very high-redshift galaxies, AGN candidates, and other objects of community interest in three of the CANDELS legacy fields (N. A. Grogin et al. 2011; A. M. Koekemoer et al. 2011) as identified from the deep multiband NIRCam imaging provided by the PRIMER (J. S. Dunlop et al., in preparation) and Cosmic Evolution Early Release Science (S. L. Finkelstein et al. 2025) JWST surveys.

CAPERS is targeting seven MSA pointings in each of the PRIMER-UDS, PRIMER-COSMOS, and extended Groth strip fields, observing three MSA configurations at each pointing. This multiconfiguration approach allows CAPERS to observe high-value targets in multiple configurations to increase their exposure times while simultaneously observing a large number of other targets in each individual configuration to increase the overall sample yield. As a high-value target, CAPERS-LRD-z9 (aka CAPERS-COSMOS-119334) was observed in all three configurations executed in CAPERS COSMOS pointing P4. These observations targeting CAPERS-LRD-z9 were executed on 2025 April 15 using two iterations of a standard three-shutter nodding scheme per MSA configuration for a total of 18 exposures at a NIRSpec aperture position angle (east of north) of 246$\mathop{.}\limits^{\unicode{x000b0}}$585. Each of these 18 nods was observed for a single 13 group integration using the NRSIRS2 readout pattern for a combined effective exposure time of 17,069 s (4.74 hr).

We reduced the spectroscopic data using the JWST Calibration Pipeline³⁵ (H. Bushouse et al. 2025) version 1.17.1 and CRDS version 1350.pmap. We largely used the default configuration for the pipeline, with a few notable exceptions. First, we enabled the clean_flicker_noise step (using the “median” method³⁶ ) in the calwebb_detector1 stage to remove the effects of 1/f noise in the count rate maps. We also employed a modified flat-field file in the calwebb_spec2 stage (see P. Arrabal Haro et al. 2023 for details). Beyond these changes, we ran the calwebb_spec3 stage using the CRDS defaults to produce a final 2D spectrum (see Figure 1, bottom panel).

Zoom In Zoom Out Reset image size

We next used a custom optimal extraction (K. Horne 1986) to best detect and extract the signal from the source in the 2D spectrum to produce a 1D spectrum. Finally, we calibrated this 1D spectrum to the JWST/NIRCam photometry provided by the PRIMER survey (Section 2.2) by computing the total flux in the spectrum within each of the F150W, F200W, F277W, F356W, and F444W NIRCam passbands. Specifically, we compared these measured fluxes to the photometric measurements of the object from NIRCam and fitted a third-order Chebyshev polynomial to the ratios between the spectrum and the photometry (as in bagpipes; A. C. Carnall et al. 2018). This correction peaks at a factor of ∼0.9 in F150W and quickly flattens to ∼0.7 from F200W through F444W. We then multiplied the 1D spectrum by the resulting fitted correction curve to produce our final calibrated 1D spectrum (see Figure 1, bottom panel). This calibration is necessary to correct for the effects of NIRSpec path loss and automated corrections applied in the JWST Calibration Pipeline that are not tied to NIRCam data.

The analysis in this Letter makes use of JWST/NIRCam photometry in the PRIMER-COSMOS field. This imaging is an internal reduction from the PRIMER team (internal version 1.0). The PRIMER imaging data were reduced using the PRIMER Enhanced NIRCam Image Processing Library (J. S. Dunlop et al. 2025, in preparation; D. Magee et al. 2025, in preparation) software.

The photometry was measured via a similar process as in S. L. Finkelstein et al. (2024). Briefly, for each band of imaging, this methodology point-spread function (PSF) matches images (using empirical PSFs) with smaller PSFs than F277W to that band and derives correction factors for images with larger PSFs (via PSF-matching F277W to a given larger PSF). Kron apertures are used to measure colors to optimize the signal-to-noise for high-redshift galaxies. Total fluxes are estimated by deriving an aperture correction in the F277W band as the ratio between the flux in the larger Kron aperture and the custom smaller aperture, with a residual aperture correction (typically <10%) derived via source-injection simulations. We show cutouts of the imaging of CAPERS-LRD-z9 and list the measured magnitudes in Figure 1 (top panel).

We searched the Barbara A. Mikulski Archive for Space Telescopes (MAST) for MIRI data at the position of CAPERS-LRD-z9. They were located in a gap of the PRIMER observations but were covered by the COSMOS-3D survey (PID 5893) on 2025 April 20. Raw data were downloaded from MAST and reduced with the Rainbow MIRI software (P. G. Pérez-González et al. 2024) using the JWST Calibration Pipeline version v1.18.0, reference files in jwst_1364.pmap. The Rainbow software implements a superbackground strategy to remove and homogenize the background, improving the detectability of faint sources. We refer the reader to P. G. Pérez-González et al. (2024) and G. Östlin et al. (2025) for details on the method and a discussion of improvements compared to exposure time calculator estimations. COSMOS-3D obtained data in the F1000W and F2100W filters, with exposure times of 927 s and 1848 s, respectively. The average 5σ detection limits for pointlike sources measured in a circular aperture of radius r = 0$\mathop{.}\limits^{^{\prime\prime} }$ 3 and 0$\mathop{.}\limits^{\unicode{x02033}}$7 for F1000W and F2100W are 24.8 and 22.8 mag, respectively (calculated with the method described below).

We measured photometry in the two MIRI bands assuming a pointlike nature and applying aperture corrections based on the empirical PSFs provided by the JWST Calibration Pipeline. We used several circular apertures in each band. We selected radii ranging from 0$\mathop{.}\limits^{\unicode{x02033}}$2 to 0$\mathop{.}\limits^{\unicode{x02033}}$5 for F1000W and 0$\mathop{.}\limits^{\unicode{x02033}}$4 to 1$\mathop{.}\limits^{\unicode{x02033}}$0 for F2100W, the ranges of aperture sizes chosen to be able to detect faint sources and limited to maximum/minimum aperture losses of 50%/25% for pointlike sources. The background was measured in a r = 10″ circular region around the source. For the background noise (used to get photometric uncertainties and the 5σ detection limits quoted above), following the method explained in P. G. Pérez-González et al. (2023; based on P. G. Pérez-González et al. 2008), we selected random noncontiguous pixels (separated by more than 3 pixels) to avoid the effects of noise correlation (i.e., an underestimation of the rms) introduced by the drizzling method when mosaicking the data.

For F1000W, we obtained consistent photometric measurements (within the errors) with all apertures, the measurements ranging from 24.90 ± 0.29 mag for r = 0$\mathop{.}\limits^{^{\prime\prime} }$2 to 25.63 ± 0.59 mag for r = 0$\mathop{.}\limits^{^{\prime\prime} }$5. The final magnitude used in the rest of the Letter, the one with the maximum S/N obtained with r = 0$\mathop{.}\limits^{^{\prime\prime} }$3 (aperture correction 0.60 ± 0.03 mag, the errors accounting for a 1 pixel centering error), is 25.02 ± 0.26 mag.

For F2100W, we obtained negative fluxes or S/N < 2 measurements for all apertures, so we conclude that there is a nondetection in this band and assume the 5σ upper limit quoted above of 22.8 mag (corresponding to an aperture of radius r = 0$\mathop{.}\limits^{^{\prime\prime} }$9, where the encircled energy is ∼70%).

The CAPERS team is using a combination of software tools, interactive inspection, and manual vetting to measure redshifts from the NIRSpec spectra; the full methodology will be detailed in a forthcoming publication. For CAPERS-LRD-z9, we derive the initial spectroscopic redshift using a modified version of msaexp³⁷ (G. Brammer 2022; V. Kokorev et al. 2024b). Key modifications include the ability to vary the velocity width and model individual lines with multiple components (e.g., narrow and broad).

We use msaexp to fit the continuum with cubic splines and emission lines with Gaussian profiles. For this step, we adopt nsplines = 20, leaving the position, amplitude, and width of the emission lines as free parameters. The full width at half-maximum (FWHM) is allowed to vary between 150 and 800 km s⁻¹ for narrow components and 800 and 5000 km s⁻¹ for broad ones. Uncertainties are estimated via Markov Chain Monte Carlo (MCMC) using resampling of the covariance matrix. All fits are performed on spectra corrected to match the photometry. To account for the wavelength-dependent resolution of PRISM, the model grid is initially generated on an oversampled wavelength axis and then convolved with the instrument dispersion curve.

With this automated multiline fitting approach, we derive an initial z_spec = 9.2880 ± 0.065 and a number of interesting features. Broad components in excess of ∼3000 km s⁻¹ (FWHM) are required to adequately model both the Hγ and Hβ emission, while the adjacent lines in the [O iii] λλ4959, 5007 doublet remain narrow (∼200 km s⁻¹ FWHM). The presence of broad permitted (e.g., Balmer series of hydrogen) and narrower (semi)forbidden lines like [O iii] gives us a first hint that an accreting SMBH is present in CAPERS-LRD-z9, typical of other spectroscopic investigations of LRDs with JWST (e.g., J. Matthee et al. 2024; A. J. Taylor et al. 2025; D. D. Kocevski et al. 2025).

We note that this redshift fitting routine is not always capable of accurately modeling all spectral features. While it performs well for most narrow and broad emission lines and is efficient for processing large spectral samples, its reliability decreases in the presence of complex line blends, uncommon line ratios, low signal-to-noise regions, or noisy continua. Accurate identification of key features, such as broad lines, is essential for robust classification of the object, especially when using the PRISM data. Therefore, in subsequent sections, we adopt a more refined fitting procedure for select regions of the spectrum.

The spectrum of our object presented in Figure 1 coupled with our initial msaexp fit hints at the presence of broad components in both Hγ and Hβ emission lines, in contrast to the narrower profiles of the adjacent [O iii] λλ4959, 5007 Å doublet. Further, the Hβ line appears to be double-peaked, which hints at an underlying absorption component. To evaluate the potential significance of these features, we perform a more detailed line fitting to the Hγ + [O iii] λ4363 and Hβ + [O iii] λλ4959, 5007 line complexes. The redder end of PRISM, where Hβ + [O iii] λλ4959, 5007 is located, has a higher spectral resolution, and this line complex is not blended; therefore, we fit that region first. We assume that the lines in the [O iii] doublet are both narrow, while the Hβ consists of a narrow, a broad, and an absorption component. In our fitting, we assume that the velocity width of the narrow component for all three lines is the same, and we fix the ratio [O iii] λ5007/[O iii] λ4959 = 2.98 (P. J. Storey & C. J. Zeippen 2000). Velocities of the narrow, broad, and absorption components are allowed to vary between 50 and 800, 1000 and 5000, and 50 and 1000 km s⁻¹, respectively. We fix the redshift (and thus the line centers of the narrow lines using the vacuum wavelengths from P. A. M. van Hoof 2018); however, we allow for small (∼±500 km s⁻¹) offsets in the absorption component and broad component of Hβ (relative to the narrow [O iii]) to allow for kinematic motion of the gas. We model the local continuum with a first-order polynomial.

We initialize the fit by first creating a set of models on the oversampled wavelength grid. To mimic the variable resolution of PRISM, we interpolate our model onto a variable step grid while making sure that the total integrated flux is preserved. Early NIRSpec/MSA results (A. G. de Graaff et al. 2023) have shown that the spectral resolution of a pointlike source falling within a slitlet is higher compared to a uniformly illuminated slit, sometimes up to a factor of 2. We therefore conservatively increase the nominal spectral resolution by a factor of 1.7. To take into account the effects of the line-spread function, we additionally convolve our model with Gaussians of variable resolution (A. G. de Graaff et al. 2023; Y. Isobe et al. 2023). We then optimize this fit via MCMC with the emcee (D. Foreman-Mackey et al. 2013) package for python.

From our fit, we securely (combined S/N ≈ 4.2) confirm the presence of the [O iii] λλ4959, 5007 doublet at a redshift z = 9.288 ± 0.003. We adopt this redshift for all subsequent calculations. We also detect a distinct broad component in Hβ at S/N ≈ 10, with an intrinsic FWHM = 3521 ± 502 km s⁻¹ and a tenuous slight redward velocity offset from the systemic redshift of 134 ± 164 km s⁻¹. The width of the narrow lines is poorly constrained due to the low resolution of PRISM, although we report a posterior value of FWHM = 483 ± 225 km s⁻¹. The narrow component of Hβ does not appear to be statistically significant (S/N ≈ 0.8). Additionally, while including the Hβ absorption component results in a visually better fit to the double-peaked Hβ line with a line width of FWHM = 413 ± 255 km s⁻¹ and slight redward offset of 84 ± 126 km s⁻¹, this component is also statistically insignificant with S/N ≈ 1.1 and highly degenerate with the narrow Hβ component. These statistical nondetections may be expected at the resolution of PRISM, but as they are physically motivated, we include them in the model regardless to better sample their covariance with the more significant model components.

The situation with Hγ + [O iii] λ4363 is more complex, as these lines are blended at the PRISM resolution. We can, however, use the prior information obtained from the Hβ + [O iii] λλ4959, 5007 complex and attempt to adequately separate the two lines. We fix the narrow and broad line widths, broad velocity offset, and redshift to the result we obtained from the Hβ + [O iii] λλ4959, 5007 fit and only fit the integrated fluxes of narrow+broad Hγ, narrow [O iii] λ4363, and a first-order polynomial continuum. Finally, the flux ratio of the narrow Hγ to narrow Hβ is not allowed to exceed 0.47, as set by Case B recombination (D. E. Osterbrock 1989). This fit is further complicated by 6 consecutive pixels at 4.375–4.425 μm that show flux above the level of the nearby continuum. While these pixels are detected at the 1σ–3σ level above the continuum, they correspond to no known emission line. Therefore, we conservatively mask out this spurious feature. We once again forward model the effects of the PRISM resolution and fit the Hγ + [O iii] λ4363 line complex with MCMC. Similarly to Hβ, we could not derive a significant flux for a narrow Hγ (S/N ≈ 2); however, we detect and partially deblend a broad Hγ component at S/N ≈ 4.8 and [O iii] λ4363 at S/N ≈ 3.1. We note that the blue wing of the broad Hγ line appears higher than the data. This may indicate that the masked spurious feature is affecting these pixels or that the Hγ broad line width is narrower than that of Hβ, as has been seen in low-redshift quasars (e.g., M. C. Bentz et al. 2023). As our analyses are not dependent on the Hγ width, this uncertainty does not affect our interpretation of this object. We present various measured and derived properties of CAPERS-LRD-z9 in Table 1 and show our best-fit models to both line complexes in Figure 2.

Table 1. Properties of CAPERS-LRD-z9

Property	Value
R.A.	150$\mathop{.}\limits^{\unicode{x000b0}}$1362532
Decl.	2$\mathop{.}\limits^{\unicode{x000b0}}$3080298
Redshift (z)	9.288 ± 0.003
Hγ flux (narrow)	<7.1
Hγ flux (broad)	45.9 ± 9.5
[O iii] λ4363 flux	13.0 ± 4.2
Hβ flux (narrow)	<62.7
Hβ flux (broad)	140 ± 14.6
Hβ flux (absorption)	>−69.1
Hβbroad FWHM	3521 ± 502 km s−1
Hβabs FWHM	282 ± 159 km s−1
Hβbroad offset	134 ± 164 km s−1
Hβabs offset	84  ± 126 km s−1
[O iii] λ4959 flux	7.2 ± 3.5
[O iii] λ5007 flux	22.2 ± 5.3
[O iii] λ5007 FWHM	483 ± 225 km s−1
M1500 Å	−18.2 ± 0.2 mag
βUV	$-0.9{9}_{-0.13}^{+0.14}$
${\mathrm{log}}\,\left({M}_{{\rm{BH}}}/{M}_{\odot }\right)$ (canonical)	7.58 ± 0.15
${\mathrm{log}}\,\left({M}_{{\rm{BH}}}/{M}_{\odot }\right)$ (systematic bounds)	6.65–8.50
AV,BLR	$1.{9}_{-1.2}^{+1.3}$
${\mathrm{log}}\,\left({M}_{* }/{M}_{\odot }\right)$	<9.0
MBH/M* (canonical)	>4.5%
MBH/M* (systematic bounds)	>46%, >0.5%
Balmer break: $\left({f}_{\nu ,4050\,\mathring{\rm A} }/{f}_{\nu ,3670\,\mathring{\rm A} }\right)$	$4.3{5}_{-0.67}^{+0.93}$

Note. All line fluxes are given in units of 10⁻²⁰ erg s⁻¹ cm⁻². Lines fit at S/N ≤ 2 are listed as 2σ (upper/lower) limits on their fluxes. We define a “canonical” ${\mathrm{log}}\,\left({M}_{\mathrm{BH}}/{M}_{\odot }\right)$ and uncertainties as only those errors propagating from the measured broad-line properties and the empirical uncertainties given in J. E. Greene & L. C. Ho (2005) without correcting for dust. We further provide upper and lower bounds on ${\rm{log}}\,\left({M}_{{\rm{BH}}}/{M}_{\odot }\right)$ to account for systematic uncertainties in the measurement methods (see Section 4.2 for details).

Download table as:  ASCIITypeset image

Zoom In Zoom Out Reset image size

To determine if any of the emission from CAPERS-LRD-z9 might originate from an extended stellar population, we use the GALFIT software (C. Y. Peng et al. 2002) to model the galaxy as a point source in several NIRCam bands to search for signs of an underlying galaxy in the residual image. For this modeling, we provide GALFIT with empirical PSFs constructed from the PRIMER-COSMOS mosaic and noise images that account for both the intrinsic image noise (e.g., background and readout noise) and added Poisson noise due to the objects themselves. We show the results of this morphological modeling in Figure 3 for F200W and F444W. We see no signs of any extended structure, indicating that the source is unresolved or the surface brightness of the resolved component is too low for detection with the sensitivity of our image. We estimate an upper limit on the size of the host to be ${r}_{h}\lt 0\mathop{.}\limits^{^{\prime\prime} }04$ and $0\mathop{.}\limits^{^{\prime\prime} }08$ in F200W and F444W, respectively, which correspond to physical sizes of ≲175 pc and ≲350 pc.

Zoom In Zoom Out Reset image size

We model the SED of CAPERS-LRD-z9 as a composite galaxy+AGN, similar to previous work on LRDs with Balmer breaks (e.g., I. Labbe et al. 2024; B. Wang et al. 2024). However, here we model the Balmer break as part of the AGN continuum (rather than old stars), driven by absorption from a shell of dense gas around the accretion disk/broad-line region (BLR; e.g., K. Inayoshi & R. Maiolino 2025; X. Ji et al. 2025; R. P. Naidu et al. 2025). We explored modeling the spectrum using only a stellar component, with a flexible nonparametric star formation history (SFH) driving the strong Balmer break with an old stellar population (>300 Myr). Comparing the fluxes on either side of the break (in 100 Å wide windows centered at 3670 and 4050 Å), our best-fit stellar-only model has a break strength of ∼2.8. This is insufficient to match the strength of the break in CAPERS-LRD-z9 (which we measure as ${f}_{\nu ,4050\,\mathring{\rm A} }/{f}_{\nu ,3670\,\mathring{\rm A} }=4.3{5}_{-0.67}^{+0.93}$), similar to other recently discovered LRDs with extreme Balmer breaks (RUBIES-BLAGN-1, $2.7{3}_{-0.17}^{+0.20}$, B. Wang et al. 2025; A2744-QSO1, $3.3{3}_{-0.15}^{+0.15}$, Y. Ma et al. 2025, A. Weibel et al. 2025; MoM-BH*-1, $7.{7}_{-1.4}^{+2.3}$, R. P. Naidu et al. 2025; RUBIES-UDS-154183, $6.{9}_{-1.5}^{+2.8}$, A. de Graaff et al. 2025). The stellar-only model is overall a poor fit to the data, with a Bayesian information criterion (BIC) difference from our fiducial model (described below) of ΔBIC > 80. This, combined with the significant detection of broad Hβ, drives our assumption that the rest-frame optical is dominated by the AGN, enshrouded in dense gas.

We therefore pursue a more complex modeling, first using the Cloudy photoionization code (version 23.01; M. Chatzikos et al. 2023) to generate a large grid of AGN model SEDs, varying both the intrinsic AGN accretion disk SED and the gas conditions around the source. We follow the framework described in K. Inayoshi & R. Maiolino (2025) and R. P. Naidu et al. (2025) and model the intrinsic AGN accretion disk SED as a series of power laws combined with a “big bump” temperature.³⁸ This continuum is passed through a shell of gas surrounding the central source that is defined in terms of its density, metallicity, and turbulent velocity. The turbulence is required to reproduce the break shape; without it, the break is very sharp at λ = 3650 Å. We vary the ionization parameter that sets the ratio of H-ionizing photons to gas densities at the illuminated face of the gas cloud. When all other parameters are fixed, varying the ionization parameter is equivalent to varying the distance between the central source and the gas shell. We stop the calculation at a set of fixed line-of-sight column densities through the gas cloud. The full grid of Cloudy model parameters is provided in Table 2.

Table 2. Cloudy Parameter Grid and Fitted Values

Parameter	Grid Values	Fitted Value
${\mathrm{log}}\,{T}_{{\rm{BB}}}$ (K)	4, 4.7, 5, 5.7	5.0
αOX	−2.5, −2.0, −1.5	−1.5
αUV	−0.1	⋯
αX	−0.5	⋯
${\mathrm{log}}\,{n}_{{\rm{H}}}\,({\mathrm{cm}}^{-3})$	9, 9.5, 10, 10.5, 11, 11.5, 12	$9.{9}_{-0.2}^{+0.2}$
vturb (km s−1)	100, 200, 300, 400, 500	$32{0}_{-60}^{+80}$
[Fe/H]	−2	⋯
${\mathrm{log}}\,U$	−3.5, −3, −2.5, −1.5, −0.5	−1.5
${\mathrm{log}}\,{N}_{{\rm{H}}}\,({\mathrm{cm}}^{-2})$	21, 22, 23, 24, 25, 26	>25.9a
Cf	0–1	$0.1{2}_{-0.01}^{+0.01}$
AV	0–3 (SMC law)	$0.5{3}_{-0.08}^{+0.09}$

Note. The first four parameters describe the incident continuum spectrum, where T_BB is the “big bump” temperature, α_OX is the optical-to-X-ray index, α_UV is the power-law UV slope, and α_X is the power-law X-ray slope. The incident continuum is passed through dense gas defined by the gas density (n_H), metallicity ([Fe/H]), and turbulent velocity (v_turb). The level of irradiation at the face of the cloud is set by the ionization parameter (${\mathrm{log}}\,U$), and the Cloudy calculations are stopped at a range of line-of-sight column densities (N_H).

^aWe provide the 1σ lower limit for the column density as our posterior is limited by the edge of the Cloudy grid.

Download table as:  ASCIITypeset image

In order to conduct joint inference on the AGN and galaxy parameters, we assemble and fit the full SED model using a modified version of bagpipes (A. C. Carnall et al. 2018), a Bayesian SED modeling code. bagpipes requires models with a continuous parameter space, rather than discrete points in a grid; we include a custom module to interpolate over the Cloudy grid (consisting of both the continuum and emission lines of the AGN processed through the dense gas shell).

However, in an effort to reduce the dimensionality of the problem, prior to fitting with bagpipes, we first perform a simple χ² minimization routine over the Cloudy grid to identify and fix the best-fit values of several grid parameters. Here, we fit only to the PRISM spectrum redward of the Balmer break and include several postprocessing steps, including broadening of the AGN emission lines, a variable covering fraction of the BLR clouds, and dust attenuation following a Small Magellanic Cloud (SMC) law (K. D. Gordon et al. 2003). Fitting to our grid, we find that the best-fit model has ${\mathrm{log}}\,{T}_{\mathrm{BB}}/{\rm{K}}=5$, α_OX = −1.5, ${\mathrm{log}}\,{n}_{{\rm{H}}}/{\mathrm{cm}}^{-3}=9.5$, v_turb = 100 km s⁻¹, ${\mathrm{log}}\,U=-1.5$, and ${\mathrm{log}}\,{N}_{{\rm{H}}}/{\mathrm{cm}}^{-2}=26$. We note that these parameters describe a Compton-thick environment that Cloudy is not designed to simulate.³⁹ Therefore, while we caution that our Cloudy models may not capture all physical processes in this regime, Cloudy remains the best accessible tool for this modeling.

In the subsequent modified bagpipes fit, we then fix T_BB, α_OX, and ${\mathrm{log}}\,U$ at their best-fit values and interpolate over the density ${\mathrm{log}}\,{n}_{{\rm{H}}}$, column density ${\mathrm{log}}\,{N}_{{\rm{H}}}$, and turbulent velocity v_turb. These parameters predominantly impact the strength and shape of the Balmer break (see X. Ji et al. 2025) and are therefore important to model flexibly to optimize the fit. We adopt uniform priors for each parameter and again include a broadening of the AGN emission lines, a variable covering fraction of the BLR clouds (C_f), and dust attenuation (A_V; SMC law); we note that we explored allowing the attenuation curve slope to vary but find no need for an extremely steep dust law (in contrast to A. de Graaff et al. 2025).

For the galaxy model, we use the BPASS v2.2.1 stellar templates (E. R. Stanway & J. J. Eldridge 2018) and a flexible SFH via the “bursty continuity” parameterization described in S. Tacchella et al. (2023). We use five fixed age bins with edges at 0, 10, 30, 100, 200, and 300 Myr and a log-uniform prior on the stellar metallicity from 0.1% to 100% solar. We include nebular emission, varying ${\mathrm{log}}\,U$ from −4 to −1, and dust attenuation (SMC law). Note that the dust attenuation is fit separately for the stars and AGN.

The final model is the combination of the stellar and AGN components, with the normalization of the latter fit as a free parameter. The model spectrum is convolved with the PRISM resolution curve, and the joint model is fit to both the NIRSpec spectrum and the NIRCam+MIRI photometry. We note that we mask the [O iii] λλ4959, 5007 doublet when fitting the spectrum, as the narrow-line region (NLR; see Section 4.4)—which is not considered in our Cloudy modeling of the BLR and nearby gas or our stellar model—may significantly contribute to these lines. We additionally mask the region around H_∞ (3630–3690 Å), where the Cloudy modeling yields artifacts due to the finite number of resolved energy levels (K. Inayoshi & R. Maiolino 2025; X. Ji et al. 2025). The resulting best-fit model is shown in Figure 4, and the results from the spectrophotometric modeling are discussed in Sections 4.3 and 4.4.

Zoom In Zoom Out Reset image size

The ratio of observed fluxes from the Balmer series lines provides a measure of dust attenuation. Here we use the observed Hβ/Hγ ratio. As illustrated in Figure 2, the Hγ line is blended with [O iii] λ4363. While we use priors based on the Hβ + [O iii] λλ4959, 5007 line widths to deblend these lines, the per-pixel S/N for Hγ is not sufficient for a double-component fit. Since we cannot isolate the narrow component of Hγ or Hβ, we compute a Balmer decrement for the BLR solely based on the ratio of the Hβ and Hγ broad components’ fluxes, ensuring flux integration across comparable velocity ranges. This yields Hβ/H$\gamma =3.{5}_{-0.9}^{+1.7}$.

To calculate A_V, we apply the SMC reddening law (K. D. Gordon et al. 2003), which has been found to match well the dust attenuation in high-z galaxies (P. L. Capak et al. 2015; N. A. Reddy et al. 2015, 2018) and reddened quasars (e.g., see A. M. Hopkins 2004). Assuming Case B recombination, the intrinsic line ratio would be (Hβ/Hγ)_int = 2.14 (D. E. Osterbrock 1989). Our observed ratio suggests significant attenuation; we estimate ${A}_{V,\mathrm{BLR}}=1.{9}_{-1.2}^{+1.3}$ (where our precision is limited primarily by the S/N ≈ 4 detection of broad Hγ). We note that while Case B recombination may not be a sound assumption for collisionally excited gas, this nonetheless provides an estimate of the reddening of the BLR.

Black hole mass estimation from broad emission lines relies on two fundamental assumptions: the broad-line emitting gas is virialized and dominated by the gravitational potential of the central black hole, and a “radius–luminosity” relationship exists between the broad-line orbital radius and the AGN luminosity (in Equation (1), the Hβ broad-line luminosity). These assumptions have generally been validated for low-redshift AGN (e.g., M. C. Bentz et al. 2006; E. M. Cackett et al. 2021), but their applicability to the broader AGN population is less certain.

In the specific case of CAPERS-LRD-z9, the broad and symmetric Hβ line profile is consistent with the basic virial assumption for kinematics dominated by gas orbits around a massive black hole, in contrast to the asymmetric broad-line profiles commonly observed in quasars with significant nonvirial kinematics (e.g., U. Vivian et al. 2022; L. B. Fries et al. 2024). The reliability of the radius–luminosity assumption is more difficult to assess. Compared to the radius–luminosity relation measured for Seyfert 1 AGN that was used to calibrate single-epoch masses, recent studies have shown that luminous and rapidly accreting quasars have smaller BLR sizes (P. Du et al. 2016; G. Fonseca Alvarez et al. 2020), which leads to smaller black hole masses. The empirical radius–luminosity relation is generally assumed to be caused by a photoionization-bounded BLR (e.g., K. T. Korista & M. R. Goad 2004), and the unusual SED shape of LRDs like CAPERS-LRD-z9 could result in a significantly different BLR structure compared to other quasars. Elevated Eddington ratios also affect the radius–luminosity relation and may also result in the overestimation of black hole masses by as much as an order of magnitude (e.g.,. P. Du et al. 2016; A. Lupi et al. 2024). We ultimately use the canonical relationship for low-redshift AGN of Equation (1) to estimate a mass for CAPERS-LRD-z9 while acknowledging that significant work must be done to better understand the applicability of mass estimates for high-redshift LRDs.

To compute the canonical black hole mass, we use Equation (10) from J. E. Greene & L. C. Ho (2005), where L_Hβ and FWHM_Hβ are the luminosity and FWHM of the broad component of Hβ, respectively:

$$\begin{eqnarray}\begin{array}{rcl}{M}_{\mathrm{BH}} & = & (2.4\pm 0.3)\\ & & \times 1{0}^{6}{\left(\displaystyle \frac{{L}_{{\rm{H}}\beta }}{1{0}^{42}\,\mathrm{erg}\,{{\rm{s}}}^{-1}}\right)}^{0.59\pm 0.06}\\ & & \times {\left(\displaystyle \frac{{\mathrm{FWHM}}_{{\rm{H}}\beta }}{1{0}^{3}\,\mathrm{km}\,{{\rm{s}}}^{-1}}\right)}^{2}{M}_{\odot }.\end{array}\end{eqnarray} \tag{ 1 }$$

Propagating the uncertainties on the calibration coefficients given in the empirical relation and the uncertainties on our line flux and FWHM, we derive a black hole mass for CAPERS-LRD-z9 of ${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })=7.58\pm 0.15$. We note that additional systematic uncertainties from applying the J. E. Greene & L. C. Ho (2005) relations at high redshift may be as high as 0.5–0.7 dex (G. Fonseca Alvarez et al. 2020; R. Abuter et al. 2024).

To best capture these true systematic uncertainties and account for the limitations of our PRISM data, we next compute strong upper and lower limits on ${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })$. First, we derive an upper bound by correcting our canonical measurement for dust attenuation. By adopting the estimated reddening of ${A}_{V,\mathrm{BLR}}=1.{9}_{-1.2}^{+1.3}$ (computed above), we derive a higher black hole mass of ${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })=8.10\pm 0.40$. Using our A_V ≈ 0.53 from our spectrophotometric fitting, we alternatively compute ${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })=7.70\pm 0.14$. We note that while our Balmer-decrement-based and spectrophotometric-fitting-based A_V values are discrepant at the 1.3σ level, we attribute this difference to the uncertain assumption of Case B recombination in the BLR, uncertainty on the Hγ line flux due to the limitations of the PRISM resolution, and possible collisional deexcitation of neutral gas in the dense gas region of the LRD. Nonetheless, we adopt the 1σ upper limit on the Balmer-decrement-derived value (${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })\lt 8.5$) as a strong upper limit on the black hole mass.

We also compute a lower limit on the black hole mass using the Hβ line luminosity and assuming that the AGN is accreting at the Eddington rate. For the purposes of computing a lower limit, we conservatively convert the Hβ luminosity to Hα luminosity assuming an unattenuated Case B ratio (Hα/Hβ = 2.86; D. E. Osterbrock 1989) and use the bolometric correction from J. Stern & A. Laor (2012) to derive ${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })\gt 6.65$. This method has the benefit of being insensitive to the measured Hβ line width. As such, it remains robust even if the Hβ line width is significantly overestimated due to uncertainties in the effects of the NIRSpec/PRISM spectral resolution or more exotic explanations for line broadening, such as electron scattering (e.g., V. Rusakov et al. 2025).

When considering these upper and lower limits together, we compute systematic bounds on ${\mathrm{log}}\,({M}_{{\rm{BH}}}/{M}_{\odot })$ of 6.65–8.50. However, for ease of comparisons to other reported LRD black hole masses in the literature, we adopt our canonical measurement and its associated nonsystemic error bounds for the remainder of this work (unless otherwise stated).

Properly estimating the stellar masses of LRDs has proven enormously difficult given the unclear origin of the “V-shaped” spectrum and degeneracies in decomposing the galaxy/AGN contribution (e.g., G. Barro et al. 2024a; V. Kokorev et al. 2023; L. H. B. Akins et al. 2024; L. J. Furtak et al. 2024; I. Labbe et al. 2024; B. Wang et al. 2025, 2024b). Here, we have assumed that the rest-UV is dominated by stars (with the rest-optical dominated by the AGN); however, we do not know this a priori. The stellar mass we derive is likely therefore an upper limit, as it is entirely possible that the UV is also AGN-dominated (a possibility we discuss further in Section 5.3). The joint bagpipes+Cloudy fitting described in Section 3.4 yields a host galaxy stellar mass of ${\mathrm{log}}\,({M}_{* }/{M}_{\odot })=8.{9}_{-0.1}^{+0.1}$.

We note that stellar masses can be underestimated when derived from the UV alone due to “outshining” from young O/B stars (see, e.g., D. Narayanan et al. 2024). However, as there are no prominent emission features in the rest-UV of CAPERS-LRD-z9, and the rest-optical emission lines are accounted for via the AGN, the SFH model is not biased toward a young stellar population. In fact, the age–dust–metallicity degeneracy is accounted for in our posterior estimate. We explored including an additional burst of SF at z = 20, and found a consistent stellar mass with no meaningful improvement to the quality of the fit, showing that the observations do not allow for a significant mass in older stars to be present in this object.

From our computed stellar mass upper limit and our canonical measurement of M_BH, we compute M_BH/M_* >4.5%. Invoking our systematic lower and upper bounds for M_BH, we alternatively compute M_BH/M_* > 0.5% and M_BH/M_* > 46%, respectively. We plot this black hole to stellar mass ratio in Figure 5, along with a large body of literature samples (Y. Harikane et al. 2023; D. D. Kocevski et al. 2023, 2025; V. Kokorev et al. 2023, 2024a; R. L. Larson et al. 2023; L. J. Furtak et al. 2024; I. Labbe et al. 2024; R. Maiolino et al. 2024; R. Tripodi et al. 2024; H. Übler et al. 2024; B. Wang et al. 2025; M. Yue et al. 2024b; H. B. Akins et al. 2025a; I. Juodžbalis et al. 2025). Figure 5 demonstrates the “overmassiveness” of CAPERS-LRD-z9 in the context of both literature samples of JWST-detected AGN (colored points) as well as the established mass ratio for local quasars of 0.1% (dashed black line). For comparison, we overlay the predicted evolution of M_BH/M_* ratios from ∼10³ early black hole populations formed via two distinct seeding channels: heavy seeds (red contours) and light seeds (blue contours), based on the SP1 model presented in H. Hu et al. (2025).

Zoom In Zoom Out Reset image size

The observed value of CAPERS-LRD-z9 is broadly consistent with both scenarios, but the lower limit nature of this value may hint toward the heavy-seed regime. We discuss possible explanations for this effect and its possible interplay with black hole seeding and growth in Section 5.1.

Our best-fit AGN model has ${\mathrm{log}}\,{n}_{{\rm{H}}}/{\mathrm{cm}}^{-3}\sim {9.9}_{-0.2}^{+0.2}$ and ${\mathrm{log}}\,{N}_{{\rm{H}}}/{\mathrm{cm}}^{-2}\gtrsim 26$.⁴⁰ These are similar to the best-fitting parameters for MoM-BH*-1 and RUBIES-UDS-154183 (A. de Graaff et al. 2025; R. P. Naidu et al. 2025) and suggest very extreme gas conditions in the BLR. To examine whether these conditions extend to the NLR, we look to the forbidden oxygen lines. We measure an extremely high ratio of [O iii] λ4363/[O iii] λ5007 (RO3) = 0.58 ± 0.23 in CAPERS-LRD-z9. Though [O iii] λ4363 is blended with Hγ at the PRISM resolution, our MCMC line fitting (described in Section 3.2) yields a significant [O iii] detection (S/N ∼ 3).

As the [O iii] λ4363 and λ5007 lines arise from different upper energy levels, RO3 is commonly used as a temperature diagnostic. High RO3 traces high-temperature regions of the ISM and is found to be increasingly common in z ≳ 8 galaxies (J. Brinchmann 2023; H. Katz et al. 2023; F. Cullen et al. 2025). RO3 has also been used as an AGN diagnostic, particularly at high redshift, where the canonical Baldwin, Phillips, and Terlevich diagram breaks down due to lower metallicity (B. E. Backhaus et al. 2023; G. Mazzolari et al. 2024). Here, the [O iii] λ4363 line likely traces the NLR or ISM gas heated by the AGN in addition to star formation.

The ratio we observe for CAPERS-LRD-z9 (RO3 ≈ 0.58 ± 0.23) is particularly extreme, even compared to other z > 7 BLAGN (V. Kokorev et al. 2023; R. Tripodi et al. 2024; H. Übler et al. 2024). Such a high ratio is only possible at densities approaching (or exceeding) the critical density of [O iii] λ5007 (n_e ∼ 7 × 10⁵ cm⁻³), where collisional deexcitation becomes important. The critical density for [O iii] λ4363 is significantly higher (∼10⁸ cm⁻³). Using pyneb (V. Luridiana et al. 2015), we find that the RO3 we measure corresponds to densities ≳10⁶ cm⁻³, even at very high temperatures of T_e ∼ 10⁵ K. At lower temperatures (T_e ∼ 10⁴ K), the implied density is even higher: n_e ≳ 10⁸ cm⁻³. This suggests that the NLR itself could be quite dense, corroborating the extreme densities inferred for the BLR.

We plot our canonical black hole mass for CAPERS-LRD-z9 in Figure 6, along with other notable z > 6 AGN from the literature. We also plot simple models of Eddington-limited black hole growth (with 10% radiative efficiency) for both stellar (light) seeds (∼10² M_⊙) and more massive heavy seeds (≳10⁴ M_⊙). The red shaded region corresponds to heavy seeds with formation redshifts of 25–15 that begin accreting at the Eddington limit immediately after formation. For the stellar remnant light seeds, we plot formation redshifts of 30–15 and require an additional 100 Myr to pass after the formation redshift before accretion begins due to progenitor gas heating by the precollapsed stars (e.g., J. L. Johnson & V. Bromm 2007). It is clear that CAPERS-LRD-z9’s SMBH is too massive to form from a simple stellar remnant that grows at the Eddington rate, similar to what has been inferred for other recently discovered high-z AGN (e.g., R. L. Larson et al. 2023). However, a heavy seed growing at the Eddington rate can easily reproduce the observed black hole mass, making it a plausible formation pathway. This scenario naturally establishes the overmassive M_BH/M_* configurations present at high z, as shown in Figure 5 (see also the red shading representing the heavy-seed model prediction). Furthermore, the emerging SMBH demographics suggest that the black hole to stellar mass ratio may bifurcate into high- and low-M_BH/M_* branches as we approach the highest redshifts. Such a bifurcation has been predicted for a hybrid seeding scenario, where the two branches are linked to heavy and light seeds, respectively (also see Figure 3 in J. Jeon et al. 2025b).

Zoom In Zoom Out Reset image size

A key uncertainty in all direct-collapse black hole (DCBH) models is the choice of formation criteria, such as the required level of soft-UV (Lyman–Werner) background flux or the critical metallicity above which gas cooling would become too efficient, resulting in cloud fragmentation that suppresses DCBH formation (e.g., V. Bromm & A. Loeb 2003; M. C. Begelman et al. 2006; K. Omukai et al. 2008; J. H. Wise et al. 2019). The value of these parameters is reflected in the resulting number densities of DCBH seeds and their descendants (e.g., A. Trinca et al. 2022; H. Hu et al. 2025; J. Jeon et al. 2025b). In turn, any firm limits on the DCBH abundance would significantly constrain the formation scenarios, as could be accomplished by linking the DCBH seeding pathway to a subset of the observed LRD population.

Some studies have suggested that the conditions required to form DCBH seeds are too rare for this channel to account for the majority of the SMBHs being detected by JWST (e.g., A. K. Bhowmick et al. 2024), whereas others argue for less restrictive conditions (e.g., S. Chon & K. Omukai 2025). However, observations suggest that ultradense star clusters may be common or even ubiquitous in high-redshift galaxies (A. Adamo et al. 2024; S. Fujimoto et al. 2024; L. Mowla et al. 2024), providing a promising additional heavy-seeding mechanism.

Furthermore, the heavy-seed explanation is not the only way to produce CAPERS-LRD-z9’s black hole mass by z = 9.288. Allowing (even mild) super-Eddington accretion with a stellar remnant seed can also reproduce the black hole mass of CAPERS-LRD-z9. For example, in Figure 6, we show that a 10² M_⊙ stellar remnant formed at z = 30 growing at an average Eddington ratio of 1.85 after a 100 Myr delay in starting accretion can easily reproduce the black hole mass of CAPERS-LRD-z9. In reality, we expect that super-Eddington accretion occurs in short bursts (e.g., E. Takeo et al. 2020; H. Suh et al. 2025). However, we plot the simplified average Eddington ratio growth curve here for visual clarity.

As referenced in Section 1, the “dense gas” model of LRDs has garnered significant attention in the literature in recent months. This model, introduced by K. Inayoshi & R. Maiolino (2025), has recently been applied to two newly discovered LRDs—MoM-BH*-1 at z = 7.76 (R. P. Naidu et al. 2025) and RUBIES-UDS-154183 at z = 3.55 (A. de Graaff et al. 2025)—as well as the triply imaged z = 7.04 LRD A2744-QSO1 (L. J. Furtak et al. 2024; X. Ji et al. 2025). We plot these objects, and the z = 4.47 LRD (UNCOVER-45924) from I. Labbe et al. (2024), in Figure 7. All spectra are normalized to the spectrum of CAPERS-LRD-z9 at rest-frame 0.51 μm.

Zoom In Zoom Out Reset image size

All of these objects are notable for their strong Balmer breaks, which, in the cases of MoM-BH*-1 and RUBIES-UDS-154183, were insufficiently fit by evolved stellar populations (A. de Graaff et al. 2025; R. P. Naidu et al. 2025). The Balmer breaks and SEDs of both objects were better fit by invoking the dense gas model of K. Inayoshi & R. Maiolino (2025). Similarly, while another object—A2744-45924—can be fit with a stellar model, the implied stellar mass is large (∼10¹¹ M_⊙) and in tension with the Atacama Large Millimeter/submillmeter Array dynamical mass measurement of ∼10⁹ M_⊙ (H. B. Akins et al. 2025b).

While CAPERS-LRD-z9 shows a factor of ∼1.5 softer break $\left({f}_{\nu ,4050\,\mathring{\rm A} }/{f}_{\nu ,3670\,\mathring{\rm A} }=4.3{5}_{-0.67}^{+0.93}\right)$ than MoM-BH*-1 ($7.{7}_{-1.4}^{+2.3}$) and RUBIES-UDS-154183 ($6.{9}_{-1.5}^{+2.8}$), all of these objects show a remarkably similar rest-frame optical SED and Balmer break shape, suggesting common physical and emission structures. However, unlike MoM-BH*-1, CAPERS-LRD-z9 demonstrates only moderate Hβ absorption and shows a strong detection of Hγ in emission. Higher-resolution spectroscopy of CAPERS-LRD-z9 beyond that of NIRSpec/PRISM would allow for a higher-fidelity examination of these features.

Given the similarity in spectral shapes between CAPERS-LRD-z9 and the sources in A. de Graaff et al. (2025) and R. P. Naidu et al. (2025), we propose that the rest-frame optical light from CAPERS-LRD-z9 is indeed dominated by AGN emission through a shell of dense (≳10¹⁰ cm⁻³) neutral gas, and that the rest-UV emission blueward of the Balmer break may be stellar emission from the host galaxy (e.g., D. D. Kocevski et al. 2023) or scattered light originating from the AGN (e.g., J. E. Greene et al. 2024). From our Cloudy modeling, we infer a gas density of ${\mathrm{log}}\,{n}_{{\rm{H}}}/{\mathrm{cm}}^{-3}=9.{9}_{-0.2}^{+0.2}$ and column density ${\mathrm{log}}\,{N}_{{\rm{H}}}/{\mathrm{cm}}^{-2}\gtrsim 26$. These results are remarkably comparable to MoM-BH*-1 and RUBIES-UDS-154183, further suggesting that these objects may have similarities in their physical structure. However, we note that this dense regime is beyond Cloudy’s recommended use cases, and alternate analyses may be necessary in future work to fully understand the nature of these exotic sources.

We also note that our best-fit spectrophotometric model overpredicts the MIRI F1000W flux by χ = −1.8. These data were used in the fitting procedure, but the model was not flexible enough to simultaneously fit the NIRSpec+MIRI data. We attempt to better match the MIRI photometry by including a flexible dust attenuation law for the AGN component of the model (see the Appendix for details). In this revised fit, we find a best-fit S. Salim et al. (2018) δ parameter—which governs the steepness of the attenuation curve, with negative values producing increasingly steeper curves—of $\delta =-4.6{6}_{-0.24}^{+0.31}$. While the resulting fit matches the MIRI F1000W flux with χ = −0.27 with no detrimental effects on the rest of the fit, δ = −4.66 represents an attenuation curve that is unphysically steep, as it is by definition a D. Calzetti (2001) curve multiplied by an additional factor of (λ/5500 Å)^4.66. For reference, an SMC law can be approximated with a comparatively small δ = −0.45. While steep attenuation curves were also explored in A. de Graaff et al. (2025) to reconcile their MIRI observations with the dense gas LRD model, we agree with their assessment that such curves are likely unrealistic. Alternate models and modeling parameters are thus necessary to reproduce the rest-NIR properties of the LRDs, including—perhaps—modifying the intrinsic SED of the accretion disk. Such modeling may soon be robustly enabled by the increasing known population of dense gas modeled LRDs.

In terms of the overall assembly history of the first SMBHs, we may witness a three-stage evolutionary sequence: the initial seeding process, a period of rapid growth during the LRD stage driven by massive inflows of gas, and a final “clearing of the fog” (possibly driven by radiation-hydrodynamics processes; e.g., A. Smith et al. 2017), when the emerging SMBHs become unobscured and, with sufficient cosmic time to assemble additional stellar mass, join the “standard” quasar/AGN sequences at lower redshifts. More specifically, K. Inayoshi (2025) proposed that LRDs may be simply the first incidence of AGN activity in a galaxy’s evolution, based on the log-normal occurrence rate of LRDs with cosmic time. The nascent evolutionary stage of the LRDs in this toy model may also explain their unique and extreme characteristics. The SMBH could form early in a galaxy’s lifespan, surrounded by a large reservoir of dense neutral gas that serves as both a readily available supply of accretion material and the shell of dense gas surrounding the AGN (as in K. Inayoshi & R. Maiolino 2025). This scenario would result in an overmassive SMBH producing a dense gas reddened SED with a strong Balmer break, in combination with nascent star formation in the host galaxy that produces a blue rest-UV slope. In summation, these two components would supply the two halves of the defining LRD “V-shaped” SED for the duration of the initial “LRD phase” of AGN activity. Given the low metallicity/dust content but high column density of the nuclear gas, this model also favorably reproduces the lack of mid-IR dust emission (H. B. Akins et al. 2024; G. C. K. Leung et al. 2024; P. G. Pérez-González et al. 2024; C. C. Williams et al. 2024) and lack of X-rays (E. Lambrides et al. 2024; M. Yue et al. 2024a; R. Maiolino et al. 2025) observed in LRDs. During future AGN active periods—after a period of dormancy—the stellar mass of the galaxy will have had sufficient time to accumulate and produce a more typical black hole to stellar mass ratio of M_BH/M_* = 0.1%. Simultaneously, the dense gas near the AGN will be consumed by the AGN during the LRD phase, such that in future instances of AGN activity, the resulting evolved galaxy and AGN resemble a classical quasar/AGN.

In Section 4.3, we have estimated the stellar mass of the host galaxy of CAPERS-LRD-z9 under the assumption that its UV luminosity is primarily powered by star formation. Even a visual inspection, however, shows that the shape of the UV continuum is unlike that of star-forming galaxies at comparable redshifts (e.g., M. Castellano et al. 2022; V. Kokorev et al. 2025). As Figure 7 illustrates, despite the relatively low S/N, the slope of the continuum of CAPERS-LRD-z9 appears to change at around rest-frame ≈2500 Å, redward of which is relatively blue and rising toward shorter wavelengths, whereas at bluer wavelengths, it flattens and then perhaps decreases. Moreover, at rest-frame ≈2500 Å (observed-frame 2.6 μm), there is a hint of an apparent broad emission feature (see Figures 1 and 4). Figure 7 also shows an overall similarity between the UV continuum of CAPERS-LRD-z9 and that of UNCOVER-45924, which has a substantially higher S/N, and strong UV lines such as C iv λ1550 and C iii] λ1908 and the Fe ii lines commonly seen in quasars (e.g., M. Vestergaard & B. J. Wilkes 2001). These features are not detected with statistical significance in the spectrum of CAPERS-LRD-z9, yet the broad emission feature at ≈2500 Å may be associated with the Fe ii pseudocontinuum, as in A2744-45924. This, as well as the overall shape of the continuum, is intriguing and adds credence to the idea that the UV continuum in CAPERS-LRD-z9 could be nonstellar.

To further explore this idea, we compare the LRDs to a control sample of quasars (z < 0.95) from the Dark Energy Spectroscopic Instrument (DESI) DR1 data set (DESI Collaboration et al. 2025). We assemble this sample by selecting DESI quasars with similar Hβ_broad FWHM, L([O iii] λ5007), and [O iii] λ5007/Hβ_total flux ratios to CAPERS-LRD-z9 based on FastSpecFit (J. Moustakas et al. 2023) emission line measurements reported in the DESI AGN/QSO value-added catalog⁴¹ (S. Juneau et al. 2025, in preparation). We plot both the 10th–90th percentile of these objects and the inverse-variance-weighted mean stack of the 20 reddest objects in Figure 7 as a standard of comparison. It is immediately clear that even the stack of the reddest DESI quasars strongly differs from the LRD rest-optical continua, including a lack of the prominent Fe ii bump at 4434–4684 Å in the LRDs (which is not expected to appear in faint [L_Hβ < 10⁴⁴ erg s⁻¹] high-redshift AGN; e.g., B. Trefoloni et al. 2024). However, as mentioned above, in the rest-UV, CAPERS-LRD-z9 and UNCOVER-45924 both exhibit possible AGN signature emission features such as the ∼0.25 μm iron bump, prominently featured in the DESI stack, suggesting that the rest-UV emission may indeed have an AGN origin.

The question naturally arises how UV emission from the accretion disk can escape such a dense gas cloud. As suggested by J. E. Greene et al. (2024), the rest-UV light in LRDs could originate from accretion disk photons being either scattered or directly transmitted (via a patchy medium). If we simply rescale the rest-UV continuum from the DESI quasars shown in Figure 7 to match our data, we find that the UV slopes are roughly consistent with an implied scattering fraction of ∼10%, fully consistent with the results presented in J. E. Greene et al. (2024). The scattering medium could be neutral (Rayleigh scattering) or partially ionized gas (Rayleigh and Thomson scattering) and could be associated with either the external regions of the same dense gaseous envelope that enshrouds the AGN or different regions altogether. The type of scattering medium and its geometry will, in general, contribute to the overall spectral shape and intensity, given the specific wavelength dependence of the scattered radiation, and this might also help explain the diversity of spectral morphologies observed in this category of sources.

If the UV emission is predominantly of AGN origin, the upper limit to the stellar mass discussed in Section 4.3 will have to be revised downward by a substantial amount, which is unconstrained by the current data. This makes the ratio M_BH/M_* ≫ 4.5% substantially more extreme, with important implications for the mechanisms of formation and coevolution of SMBHs and their host galaxies.