arxiv: 2403.12132 · v4 · submitted 2024-03-18 · 🌌 astro-ph.GA · astro-ph.CO· physics.comp-ph

Recognition: unknown

The haloes that reionized the Universe

Nachiket Joshi , Mahavir Sharma (IIT Bhilai)

Authors on Pith no claims yet

Pith reviewed 2026-05-06 19:17 UTC · model claude-opus-4-7

classification 🌌 astro-ph.GA astro-ph.COphysics.comp-ph PACS 98.80.-k98.62.Ai95.85.Bh

keywords cosmic reionization21 cm cosmologyescape fractionbursty star formationhigh-redshift galaxiesJWST21cmFASTSquare Kilometre Array

0 comments

The pith

Reionization was driven by bursty galaxies in ~10⁹ solar-mass haloes, with a skewed-Gaussian escape-fraction law that predicts a 21 cm power-spectrum peak at 180 MHz.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks which galaxies actually reionized the Universe and answers in two steps. First, by overlaying JWST measurements onto a cosmological hydrodynamical simulation at redshift six and above, the authors show that the galaxies bright and active enough to matter are the ones undergoing bursts of star formation, with specific star formation rates above 10⁻² per million years and UV magnitudes brighter than −17. These galaxies live mostly in haloes near 10⁹ solar masses. Second, they fit the escape fraction of ionizing photons as a function of halo mass with a skewed Gaussian that peaks at this mass scale and flattens at higher masses, then feed that prescription into a standard semi-numerical reionization code. The resulting ionized-fraction history matches observations, and the model predicts a peak in the 21 cm power spectrum at 180 MHz (z ≈ 6.8). The point a sympathetic reader should take away is that the bulk of reionizing photons came from a fairly narrow halo-mass window, and that this window leaves a specific, falsifiable signature in upcoming low-frequency radio surveys.

Core claim

Comparing JWST observations against a large hydrodynamical simulation, the authors identify the galaxies that drove cosmic reionization as bursty star formers (specific star formation rate above 10⁻² Myr⁻¹, UV magnitude brighter than −17), most of which sit in dark matter haloes around 10⁹ solar masses. They argue that the ionizing-photon escape fraction, as a function of halo mass, is best described by a skewed Gaussian peaking near this mass scale with a flat tail at higher masses. Plugging this prescription into a semi-numerical reionization code recovers the observed ionized-fraction history and yields a concrete prediction: a peak in the 21 cm power spectrum at 180 MHz (z ≈ 6.8) that up

What carries the argument

A skewed-Gaussian escape-fraction model conditioned on halo mass — peaking near 10⁹ solar masses with a flat high-mass tail — calibrated against the bursty-galaxy population identified in a JWST/simulation comparison and propagated through a semi-numerical reionization code (21cmFAST) to generate the ionization history and 21 cm observables.

If this is right

Reionization was driven by a fairly narrow halo-mass window centred near 10⁹ solar masses, not by either the rarest bright galaxies or an indefinite tail of ultra-faint dwarfs.
Upcoming low-frequency radio surveys should see a peak in the 21 cm power spectrum near 180 MHz (z ≈ 6.8), giving a clean observational test.
Bursty star formation, rather than a smooth duty cycle, is the relevant mode for producing the ionizing-photon budget of the epoch of reionization.
Escape fractions in cosmological reionization modelling should be treated as a non-monotonic function of halo mass, peaking and then flattening, rather than a power law.
The galaxies most relevant to reionization fall in the JWST-accessible UV magnitude range (≤ −17), so direct counts of bursty sources at z ≥ 6 should constrain the model further.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The flat high-mass tail in the escape-fraction curve hints that feedback in larger haloes saturates rather than fully chokes off Lyman-continuum leakage, which could be tested against resolved simulations of individual massive galaxies.
If burstiness is the controlling variable, the predicted 21 cm signal should be more spatially patchy than in models with smooth star formation, possibly distinguishable in higher-order statistics beyond the power spectrum.
Cross-correlating the predicted 180 MHz peak with JWST counts of bursty z ≈ 6–7 galaxies in the same fields would tighten the link between the identified source population and the reionization signal.
The framework implicitly downweights ultra-faint dwarfs as reionization drivers, which is testable against future deep lensed-field surveys probing M_UV > −17.

Load-bearing premise

That the chosen shape of the escape-fraction curve (a skewed Gaussian peaking at 10⁹ solar masses with a flat tail) reflects real physics rather than just being a flexible fit tuned to reproduce the same reionization history it is then said to explain.

What would settle it

The Square Kilometre Array's measurement of the 21 cm power spectrum: detecting a peak at 180 MHz (z ≈ 6.8) supports the model; finding the peak at a substantially different frequency, or with very different amplitude, would rule out this halo-mass and escape-fraction picture.

read the original abstract

We study the reionization of the Universe due to haloes that host galaxies undergoing bursts of star formation. By comparing the recent results from the James Webb Space Telescope (JWST) with the cosmological hydrodynamical simulation EAGLE at $z\ge 6$, we find that bursty galaxies have specific star formation rate, sSFR $>10^{-2}$ Myr$^{-1}$, and magnitude, $M_{\rm UV}\leq -17$. Most of them reside in haloes of mass $\sim 10^9$ M$_\odot$ and some in more massive haloes. We then construct the models of escape fraction and find that a skewed Gaussian function with a flat tail towards the high mass end best describes the mean dependence of escape fraction on halo mass, considering the haloes hosting bursty galaxies as the primary drivers of reionization. We implement the models of escape fraction in the code 21cmFAST to study the progress of reionization and derive the evolution of the mean ionized fraction that agrees well with observations. We also calculate the brightness temperature, spin temperature, and kinetic temperature and further study the spatial fluctuations in these quantities to gain insights into the progress of reionization. We compute the 21 cm power spectrum and predict a peak in power at $180$ MHz corresponding to redshift, $z\approx 6.8$, that is testable by the upcoming Square Kilometre Array (SKA). Our findings suggest that the Universe was reionized by the haloes of $\gtrsim 10^{9}$ M$_\odot$.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Standard EAGLE+21cmFAST pipeline with one concrete SKA-testable number; whether the 180 MHz peak is a real prediction depends on a calibration question the abstract doesn't answer.

read the letter

Quick read on Joshi & Sharma. This is an abstract-only assessment, so take it as such.

What's here: they cross EAGLE with JWST z≥6 results to argue the bursty population (sSFR > 10⁻² Myr⁻¹, M_UV ≤ −17) sits mostly in ~10⁹ M☉ haloes, fit a skewed-Gaussian-with-flat-tail f_esc(M_halo), push it through 21cmFAST, recover x̄_HI(z), and predict a 21 cm power-spectrum peak at 180 MHz / z≈6.8. The pipeline is conventional but cleanly executed in spirit, and the headline number is the kind of concrete, falsifiable forecast SKA can actually go after. That alone makes the paper worth more than a glance.

What's genuinely new, modulo the literature I haven't re-checked: the specific bursty-galaxy thresholds tied to halo mass via EAGLE, the particular f_esc functional form, and the 180 MHz peak prediction. None of these are revolutionary, but they're useable downstream — other groups can compare against them.

The soft spot the stress-test note flagged is the right one, and I'd state it more directly than the reader did. The abstract says the skewed Gaussian "best describes the mean dependence of escape fraction on halo mass" without naming the objective. If f_esc is fit against an EAGLE-internal estimator (RT post-processing, column densities, anything halo-resolved), then x̄_HI(z) recovery is a genuine validation and the 180 MHz peak is a real forecast. If instead the f_esc parameters are tuned against τ_e and x̄_HI itself, then the reionization-history "agreement" is tautological and the 21 cm peak inherits whatever degeneracy survives the calibration. Both readings are consistent with the abstract. This isn't a fatal issue — it just sets how much weight the SKA prediction carries — but it has to be answered in the body of the paper, and the referee should pin it down.

Minor: a skewed Gaussian with a flat tail is a four-parameter family for a quantity we barely constrain physically. Fine as phenomenology, but the paper should own that it's phenomenology rather than imply mechanism.

Recommendation: send to peer review. The methodology is standard, the prediction is testable, and the calibration question is exactly what referees are for. Not reading-group material for me — there's no methodological novelty I'd want the group's hour on — and I'm unlikely to cite it in the next year unless the f_esc fitting story turns out cleaner than the abstract suggests. Worth letting a referee adjudicate.

Referee Report

4 major / 6 minor

Summary. The manuscript identifies, by comparison of EAGLE simulations with JWST observations at z ≥ 6, a population of bursty galaxies (sSFR > 10^-2 Myr^-1, M_UV ≤ -17) residing predominantly in ~10^9 M_sun haloes. It proposes that these haloes are the primary drivers of cosmic reionization, parameterizes the mean escape fraction as a skewed Gaussian in halo mass with a flat high-mass tail, runs the resulting model through 21cmFAST, and reports an ionized-fraction history consistent with current observational constraints together with a predicted 21 cm power-spectrum peak at 180 MHz (z ≈ 6.8) as a forecast for SKA.

Significance. If the result holds, the paper provides (i) a JWST/EAGLE-grounded characterization of which galaxies and haloes dominate reionization — with halo mass ~10^9 M_sun as the pivotal scale — and (ii) a concrete, falsifiable forecast for the 21 cm power-spectrum peak amplitude and frequency that is directly testable by SKA. The combination of a JWST-anchored bursty-galaxy criterion with a 21cmFAST-level forecast is useful, and the explicit prediction at 180 MHz is the kind of falsifiable observational target that the field needs. The framing of bursty galaxies in ~10^9 M_sun haloes as the dominant reionizers is a useful contribution to the ongoing debate over the relative role of faint vs. bright galaxies and low- vs. high-mass haloes.

major comments (4)

[Abstract / f_esc model construction] The abstract states that a skewed Gaussian with a flat high-mass tail 'best describes the mean dependence of escape fraction on halo mass' but does not specify the target against which 'best' is defined. This is load-bearing: if the four parameters (peak mass, scale, skewness, tail amplitude/normalization) are tuned — directly or via grid search — against τ_e and/or x̄_HI(z), then the agreement with the observed reionization history is calibration, not prediction, and only the shape (not amplitude or frequency) of the 21 cm peak retains forecast value. The manuscript must state explicitly what quantity the f_esc functional form is fit to, show the residuals of that fit, and quantify the residual freedom propagated into the 21 cm power spectrum. A halo-resolved f_esc estimator extracted from EAGLE itself (e.g., column-density-based or post-processed radiative transfer) would settle this;
[21 cm prediction at 180 MHz] The 180 MHz / z ≈ 6.8 peak is presented as the testable forecast but, as written, its predictive content depends on item 1 above. Please provide the family of 21 cm power-spectra obtained when the f_esc parameters are varied within the range allowed by their calibration target (whatever that target is), so that the reader can see how tightly the peak frequency and amplitude are pinned down. Without an uncertainty band on the predicted peak, the 'testable by SKA' claim cannot be evaluated.
[Functional-form choice for f_esc(M_halo)] The choice of a skewed Gaussian with a flat high-mass tail is presented as descriptive but is itself a modelling axiom with four free parameters acting on a single variable (M_halo). Please justify (i) why halo mass alone is adequate — environment, sSFR, gas content, and merger state are known to matter — and (ii) why this particular four-parameter family was selected rather than, e.g., a double power law or a piecewise form. A comparison of goodness-of-fit and of resulting 21 cm predictions across at least two alternative functional families would establish that the 180 MHz peak is robust to the parameterization choice rather than a feature of this specific ansatz.
[Bursty-galaxy thresholds (sSFR > 10^-2 Myr^-1, M_UV ≤ -17)] The two cuts that define 'bursty' are given as sharp thresholds. Please show how the identification of ~10^9 M_sun haloes as the dominant hosts depends on these thresholds — in particular, whether moving the sSFR cut by 0.3–0.5 dex or the M_UV cut by ±1 mag shifts the inferred dominant halo mass scale, and consequently the f_esc(M_halo) calibration and the 180 MHz peak. The current presentation reads as if these thresholds are observationally inevitable, but the abstract suggests they are chosen from the EAGLE/JWST comparison.

minor comments (6)

[Abstract] Specify the cosmology and the EAGLE volume/resolution used for the z ≥ 6 comparison, and the JWST sample(s) (CEERS, JADES, etc.) that anchor the bursty-galaxy thresholds.
[Abstract] State the τ_e value that the model produces alongside the mean ionized-fraction history; this is the standard cross-check expected by the reionization community.
[21cmFAST setup] Please report the 21cmFAST version, box size, cell size, and the choice of T_vir / M_min, since these are known to influence the 21 cm power-spectrum amplitude at the level of the predicted peak.
[Predicted peak] Quote the predicted Δ²_21 amplitude at 180 MHz with units, not just the frequency, so readers can compare to SKA sensitivity curves directly.
[Terminology] 'Flat tail towards the high mass end' should be defined quantitatively (over what mass range, at what fractional level relative to the peak).
[References] Cite the existing literature on f_esc(M_halo) parameterizations from radiative-transfer simulations (e.g., SPHINX, Thesan, CoDa) so that the choice of the skewed-Gaussian form can be situated relative to prior fits.

Simulated Author's Rebuttal

4 responses · 2 unresolved

We thank the referee for a careful and constructive report. The four major comments converge on a single, legitimate concern: the manuscript does not draw a sharp enough line between what is calibrated against existing reionization constraints and what is genuinely predicted for SKA, and it does not propagate the residual freedom in the f_esc parameterization, the choice of functional family, or the bursty-galaxy thresholds into the predicted 21 cm power spectrum. We accept this critique. In revision we will (i) state explicitly that f_esc(M_halo) is calibrated to x̄_HI(z) and τ_e — not derived from radiative transfer in EAGLE — and report best-fit parameters with residuals; (ii) provide an uncertainty band on the 21 cm power spectrum by running 21cmFAST over the calibration-allowed f_esc parameter range, so the predicted peak at ~180 MHz is presented with frequency and amplitude error bars and compared to SKA1-Low sensitivity; (iii) refit two alternative functional families (double power law and a piecewise form) and overlay their 21 cm predictions to test robustness of the peak; and (iv) add a sensitivity analysis of the sSFR and M_UV thresholds, reporting how the dominant host halo mass and the predicted 21 cm peak shift. We are honest that an EAGLE-internal, halo-resolved f_esc estimator from post-processed radiative transfer is beyond the scope of this paper and we will say so rather than claim otherwise. We believe these revisions directly address the referee's concerns

read point-by-point responses

Referee: Abstract / f_esc construction: 'best describes' is load-bearing — if the skewed Gaussian's four parameters are tuned against τ_e or x̄_HI(z), the reionization-history agreement is calibration, not prediction. The target, residuals, and propagated freedom must be stated; ideally an EAGLE-internal halo-resolved f_esc estimator should be used.

Authors: The referee is correct that the manuscript, as written, does not separate calibration from prediction sharply enough. To clarify: the skewed-Gaussian-plus-flat-tail form is fit to reproduce the observed x̄_HI(z) constraints (and consistency with τ_e from Planck) given the EAGLE halo mass function and the bursty-galaxy ionizing-photon budget; it is not derived from EAGLE radiative transfer. Consequently the amplitude of the reionization-history match is calibration, and only the *shape* of the 21 cm signal — its qualitative peak structure — is genuinely predictive, as the referee states. We will (i) state this explicitly in the abstract and §on f_esc construction, (ii) add a table of best-fit parameters with residuals against the calibration data and their covariance, and (iii) note that a halo-resolved f_esc from EAGLE post-processing (e.g. column-density-based or coupled radiative transfer) is beyond the present scope but is the natural next step. We do not currently have such an estimator and will not claim one. revision: yes
Referee: 180 MHz / z ≈ 6.8 peak: predictive content depends on the calibration question above. Provide the family of 21 cm power spectra obtained when f_esc parameters are varied within their calibration-allowed range, so the reader can see how tightly peak frequency and amplitude are pinned. Without an uncertainty band, 'testable by SKA' is not evaluable.

Authors: We agree. In the revision we will run 21cmFAST over the f_esc parameter posterior allowed by the x̄_HI(z) and τ_e constraints (sampling peak mass, width, skewness and tail amplitude on a grid bracketing the best fit) and present the resulting envelope of Δ²_{21}(k,ν). This will allow the reader to read off both the predicted peak frequency range and amplitude range, together with their covariance, and to compare them against forecast SKA1-Low sensitivity curves. We expect the peak frequency to be more robust than its amplitude, but we will let the figure speak for itself rather than assert this. revision: yes
Referee: Functional form: skewed Gaussian + flat tail is itself a 4-parameter modelling axiom. Justify (i) why M_halo alone suffices, given that environment, sSFR, gas content and merger state matter, and (ii) why this family rather than e.g. a double power law or piecewise form. Compare goodness-of-fit and resulting 21 cm spectra across at least two alternative families to establish robustness of the 180 MHz peak.

Authors: This is a fair criticism. (i) We use M_halo as the single variable because, in our EAGLE selection, sSFR and bursty-galaxy incidence are themselves strong functions of M_halo at z≥6, so much of the secondary dependence is absorbed; we do not, however, claim it is fundamental, and we will state this caveat explicitly. A multivariate f_esc(M_halo, sSFR, gas) treatment is deferred. (ii) The skewed-Gaussian-plus-tail family was chosen because it captures the EAGLE bursty-host distribution (a peak near ~10^9 M⊙ with a non-negligible high-mass contribution) with the smallest number of parameters we found adequate; the choice was not benchmarked against alternatives. In the revised version we will refit with (a) a double power law and (b) a piecewise/broken form, report χ² against the same calibration target, and overlay the resulting 21 cm power spectra. If the 180 MHz peak survives across families we will say so; if it does not, we will revise the central claim accordingly. revision: yes
Referee: Bursty-galaxy thresholds (sSFR > 10^-2 Myr^-1, M_UV ≤ -17) are presented as sharp and observationally inevitable, but appear chosen from the EAGLE/JWST comparison. Show how the ~10^9 M⊙ host identification, the f_esc calibration, and the 180 MHz peak shift if the sSFR cut is moved by 0.3–0.5 dex or the M_UV cut by ±1 mag.

Authors: The referee is correct that these cuts emerge from the EAGLE/JWST comparison rather than being externally imposed; the manuscript should not give the opposite impression and we will rephrase. We will add a sensitivity analysis varying the sSFR threshold by ±0.5 dex and the M_UV threshold by ±1 mag, and report for each combination: (a) the modal host halo mass and its dispersion, (b) the refit f_esc(M_halo) parameters, and (c) the resulting 21 cm peak frequency and amplitude. This will let the reader see directly whether the ~10^9 M⊙ scale and the 180 MHz peak are threshold-dependent or robust. We do not yet know the outcome of (c) at the extreme corners and will report it as found. revision: yes

standing simulated objections not resolved

We cannot, within the scope of this paper, replace the calibrated f_esc(M_halo) with a halo-resolved estimator extracted from EAGLE via post-processed radiative transfer; this is acknowledged as a limitation rather than answered.
We cannot a priori guarantee that the 180 MHz peak will survive the alternative functional-form and threshold-sensitivity tests we are committing to perform; the revised manuscript will report whatever those tests show, including a weakening of the central claim if warranted.

Circularity Check

2 steps flagged

Abstract is ambiguous about whether the skewed-Gaussian f_esc(M_halo) is fit to an EAGLE-internal escape-fraction estimator or tuned against the reionization history; only the latter would be circular, and the abstract alone does not establish which.

specific steps

other [Abstract, sentence on escape-fraction model]
"we construct the models of escape fraction and find that a skewed Gaussian function with a flat tail towards the high mass end best describes the mean dependence of escape fraction on halo mass, considering the haloes hosting bursty galaxies as the primary drivers of reionization"

The abstract does not specify the objective against which 'best describes' is evaluated, nor whether the target is a halo-resolved f_esc proxy from EAGLE (independent) or the reionization history itself (circular). The qualifier 'considering the haloes hosting bursty galaxies as the primary drivers of reionization' frames the model selection in terms of the reionization outcome the paper then claims to predict. This is a localized risk of FITTED INPUT CALLED PREDICTION, but not demonstrated from the abstract alone — full-text inspection of the fitting objective is required to confirm or rule out.
other [Abstract, ionized-fraction claim]
"We implement the models of escape fraction in the code 21cmFAST to study the progress of reionization and derive the evolution of the mean ionized fraction that agrees well with observations."

If the f_esc(M_halo) parameters were chosen with reference to x̄_HI(z) or τ_e data, then 'agrees well with observations' is consistency by construction rather than a derivation. The 180 MHz 21 cm peak would then be the only genuinely predictive output, and its robustness depends on how much residual freedom the skewed-Gaussian form retains after calibration. Cannot be confirmed from abstract; flagged as a structural risk only.

full rationale

With only the abstract available, the derivation chain visible is: (i) identify bursty galaxies via JWST↔EAGLE comparison (sSFR, M_UV, M_halo); (ii) "construct models of escape fraction" and select a skewed-Gaussian-with-flat-tail form as best describing f_esc(M_halo); (iii) feed into 21cmFAST and report agreement with x̄_HI(z) plus a predicted 180 MHz 21 cm peak. Layer (i) is referenced to external simulation+observation comparisons and is not circular in the abstract. Layer (iii) is mechanically a forward run of 21cmFAST given (ii), so its non-circularity hinges entirely on how the skewed-Gaussian parameters in (ii) are fixed. The abstract uses the phrase "best describes the mean dependence of escape fraction on halo mass, considering the haloes hosting bursty galaxies as the primary drivers of reionization." This language is consistent with either (a) a fit to a halo-resolved f_esc proxy from EAGLE, in which case the agreement with x̄_HI(z) is a genuine prediction, or (b) a fit/selection conditioned on reproducing reionization observables, in which case "agrees well with observations" is consistency by construction and the 180 MHz peak inherits residual freedom in the functional form. The clause "considering the haloes hosting bursty galaxies as the primary drivers of reionization" is mildly suggestive of (b) — the modelling choice is framed in terms of the reionization outcome — but it is not dispositive. Hard rule 1 forbids claiming circularity without an exhibited reduction. The abstract does not present equations or a fitting objective, so I cannot quote a step where a fitted parameter is renamed as a prediction. The honest finding is: a real, identifiable circularity risk localized to one step (parameter selection in the f_esc model), not yet demonstrated. Score 3 reflects "concern present, not established," not "circularity proven." A full-text pass should look for: (1) the explicit objective minimized when picking the skewed-Gaussian parameters, (2) whether τ_e or x̄_HI(z) data points enter that objective, and (3) whether any halo-resolved EAGLE f_esc estimator is used as the target. If (1)–(2) confirm tuning against reionization observables, the score on the abstract's "agrees with observations" claim should rise toward 6; if (3) holds independently, the score should drop to 1–2.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

No new physical entities are postulated. The load-bearing additions are (a) a chosen functional form for escape fraction vs. halo mass with associated fitted parameters, and (b) thresholds defining 'bursty' reionizers. The abstract gives no indication of how many free parameters the f_esc family has nor the calibration target, so the most honest reading is that the central forecasts inherit moderate calibration freedom. No invented particles, mediators, or forces; the methodology stays inside standard ΛCDM + galaxy formation.

free parameters (2)

Skewed-Gaussian escape-fraction parameters (location, scale, skewness, tail amplitude) = not specified in abstract
These set f_esc(M_halo) and therefore the entire reionization history and 21 cm prediction; their calibration target is unclear from the abstract.
Bursty-galaxy thresholds (sSFR > 10^-2 Myr^-1, M_UV ≤ -17) = sSFR > 10^-2 Myr^-1; M_UV ≤ -17
Define the reionizer population; chosen via comparison of EAGLE and JWST distributions.

axioms (3)

domain assumption EAGLE galaxy population at z ≥ 6 is representative of the JWST-observed reionization-era population after the bursty selection.
Required for the identification of reionizers; EAGLE was not specifically calibrated to this regime.
domain assumption 21cmFAST predictions for spin/kinetic/brightness temperatures and the 21 cm power spectrum are accurate at the level needed to identify a 180 MHz peak.
Semi-numerical reionization codes carry known systematics relative to full radiative transfer.
ad hoc to paper Halo mass is the relevant single variable for parameterizing escape fraction; skewed-Gaussian-with-flat-tail is the appropriate functional family.
Functional form chosen empirically as best fit; physical motivation not stated in the abstract.

pith-pipeline@v0.9.0 · 9922 in / 6092 out tokens · 87847 ms · 2026-05-06T19:17:23.927625+00:00 · methodology