# Verified controls and covariance-sensitive uncertainty support

User-directed continuation, 11 October 2026. The range-coalescing repair
(`21701526`) successfully retrieved exactly the registered native pixels.
The previous stopped journals and insufficient-evidence report remain historical
attempt records; no earlier failure or scientific threshold was erased.

## New verified gate evidence

Of 1,000 frozen clumps, 941 have a central-column possible control, and **939**
have at least one control also matching both thirteen-point profile summaries.
There are 12,648 such possible pairs. The native profile and bin measurements
were independently reconstructed from complete saved bytes, separate spherical
rotations and an independent MIME decoder.

Only **333/1,000** have a profile-matched control where both arms satisfy the
registered sum(sigma)/N <= 0.30 uncertainty bound. This fails the minimum 800.
The original occurrence comparison therefore remains unmeasured and must not
proceed to source counts under this failed gate. It is not a measured null or
absence of embedded activity. Final sensitivity/crowding/balance was not certified.

## Post-read covariance diagnosis, separately registered

The native products contain twelve marginal temperature-bin standard deviations.
They do not, by themselves, provide the correlations needed for the error of
their sum. Our registered sum-of-sigmas bound allows arbitrary covariance and
is conservative, conditional on those marginal sigmas; it excludes systematic
opacity/calibration errors. No missing bin or unknown source was counted as zero.

A frozen offline diagnostic tested five hypothetical common-correlation
scenarios. An independent verifier used explicit positive-semidefinite 12x12
covariance matrices and the quadratic form for the variance of their sum.

| Hypothetical bin correlation rho | Clumps with a possible uncertainty-qualified control |
|---:|---:|
| 0 (independent bins) | 814 |
| 0.25 | 650 |
| 0.50 | 514 |
| 0.75 | 418 |
| 1 (original conservative bound) | 333 |

These are descriptive scenarios, not fitted correlations, confidence bounds,
new control certification or estimates of hidden stars. Choosing independence
because it clears 800 would change the failed primary test after readout.
The original bound rejects 568 profile-supported targets on their own error;
38 more lack a qualified control despite qualifying target error. Median
conservative relative errors are 0.341 at target centres and 0.429 at controls.
There were zero unknown bin-uncertainty centres among the measured subset.

The diagnosis changes the next scientific action: investigate defensible
total-column uncertainty/covariance before any infrared occurrence inference.
An embedding cannot replace that missing uncertainty model. The comparison's
answer, physical associations, stars, births, lifetimes and detector performance
remain unestablished. Do not substitute the 333 subset for the frozen primary
population or loosen its 0.30 threshold.

## Model context and scope

[Marsh et al. (2015), equations 15–16](https://arxiv.org/html/1509.08699v1)
derive marginal posterior uncertainties from the diagonal of an inverse
curvature matrix. Independent prior occupation variables do not supply posterior
bin independence. The [Hi-GAL PPMAP release paper](https://arxiv.org/html/1707.03808v1)
describes marginal uncertainty cubes and separates random errors from omitted
systematic effects. These published statements motivate the covariance audit;
they do not measure correlations for our 1,000 clumps.

## Reproduction and storage

* `profile_uncertainty_gate_coalesced.json` and
  `profile_uncertainty_verification_coalesced.json`: producer and independent
  939/333 support records.
* `profile_support.csv`: all 1,000 exact requested identities and failures.
* `covariance_diagnostic.json` and `covariance_verification.json`: frozen
  post-read scenario computation and independent full-matrix replay.
* `docs/MATCHED_CONTROL_COVARIANCE_DIAGNOSTIC_20261011.md`: prospective diagnostic
  rules, explicit post-read role and stopping limits.

The registered acquisition used 3,685 cumulative requests, 29,761,115 received
body bytes and 339.075 seconds. Complete compressed native responses and detailed
tables remain local, with full hashes tracked. Scientific-stage storage is about
23 MiB; free disk remains about 51 GiB. There were no infrared source rows,
full uncertainty-cube downloads, galleries or model training.