CIRUMODEL QUALITY REVIEW
23 October 2026 · Plain-language edition

The latest comparison · Raon / Ornith / Flint

Which model would I trust
with the whole job?

Raon is my provisional choice for longer jobs. It kept working through all four requests and left a usable handover. But it also left mistakes that could quietly produce the wrong bill or give false reassurance about recovery.

Why the recommendation changed

Earlier short tasks favored Ornith. The new test asked the models to keep improving the same application without losing earlier work. Raon carried that job furthest.

How firmly I’d hold that view

Enough to prefer Raon for this kind of work; not enough to call it best overall. This was one application. The review did not settle why Ornith and Flint began repeating themselves.

01

What stood out about each model

The useful question is what you would have to watch for when each model is doing real work.

Raon

Best evidence of carrying a long job through

Kept investigating, repaired its own mistakes, and reached the final handover. It preserved the supplied correction files.

Watch for quiet wrong answers. Finishing confidently did not mean every credit, recovery step, or backup check was right.

Ornith

Strongest earlier record on short repairs

Made a clean first repair and explained it accurately. Earlier separate tasks also gave it the most consistent short-repair record.

Watch for stalled progress. It repeated a settled question during the second request. The cause remains unresolved.

Flint

Useful speed and real repair ability

Completed three requests and made a strong repair for failed reports and overlapping updates.

Watch what changed around the answer. It altered supplied records, then brought back an older version that lost working fixes.

02

The same application, four successive requests

Each model kept its earlier work. The next request arrived only after it finished the previous one. This tests whether useful changes survive as the job grows.

The request
Raon
Ornith
Flint
1 · Finish an interrupted order import
RaonFinishedInitial checks passed
OrnithFinishedClean first repair
FlintFinishedInitial checks passed
2 · Apply corrections; preserve the original records
RaonFinishedFound and repaired a lost credit
OrnithStopped in a loopNo finished second change
FlintFinished, with a serious flawChanged supplied correction files
3 · Recover failed reports and overlapping updates
RaonFinishedSubstantive repair
OrnithNot reachedNo result to judge
FlintFinishedStrong repair and meaningful checks
4 · Upgrade an older installation and leave a handover
RaonFinished, with hidden defectsReal handover; important checks incomplete
OrnithNot reachedNo result to judge
FlintStopped in a loopEarlier fixes had been lost

“Finished” describes progress, not correctness. These are three runs through one continuing job, not twelve independent tests. Unfinished or unreached work is not counted as a wrong answer.

03

The differences that actually matter

These are plain-language accounts of saved work and checks, rather than invented sample answers. Each comparison explains what makes one result more convincing.

A

Two genuine credits happen to look identical

A customer receives two separate $2 credits for the same reason. Both must count. After both, Bob should owe $19.

Raon

Wrong amount in a saved check

Counted the second credit as an accidental repeat. Bob was left owing $21, with only one credit recorded.

Ornith

Correct rule in its finished first repair

Kept separately requested credits separate. Review of that saved version found the right behavior. It did not make Raon’s “looks the same, so ignore it” change.

Flint

Correct amount in a saved check

Recorded both credits. Bob owed $19, and both entries remained visible in the history.

Why Ornith and Flint look better here: they respected the difference between two real instructions and one instruction sent twice. Raon’s longer overall completion does not excuse this wrong bill. Evidence A ↗
B

A correct-looking total can hide the wrong method

The job was to process new and old correction files correctly while keeping those original files intact.

Raon

Original supplied files preserved

Changed the application to handle the corrections, then used the supplied files. Its work could be checked against the original evidence.

Ornith

No finished second change

Investigated the correction rules, but never delivered a completed change. That leaves no finished result to compare on this request.

Flint

Supplied evidence changed

Rewrote both correction files. Among the changes, an old $14 entry became $12, and the files’ ordering information was changed too.

Why Raon looks better here: matching the expected total means much less if the model also changes the records it was supposed to handle. Flint did make a working correction mechanism, but that does not justify changing the supplied evidence. This describes its actions, not its intent. Evidence B ↗
C

The individual repairs work—but do they work together?

A newer correction has been accepted but not fully applied when work is interrupted. Recovery must finish applying it, even when the order already exists.

Raon

A later combined check failed

Saw that the order already existed and skipped its newer correction. It could report successful recovery while leaving the old amount in place.

Ornith

Later combination not reached

Its initial missing-order repair worked. That does not tell us whether it would have handled newer corrections and the later upgrade together.

Flint

Strong earlier recovery work

Its finished third change correctly recovered saved report contents and handled overlapping updates. It did not complete the later upgrade, so it has no equivalent final combined result.

What this changes: Raon proved more of the full journey, and that exposed a serious hidden error. Flint deserves credit for its completed recovery repair; it has not been shown to beat Raon on the unfinished final combination. Evidence C ↗
D

When the job gets harder, does progress survive?

New requirements should build on earlier repairs. A model must keep moving without quietly undoing work that already worked.

Raon

Continued through all four requests

Kept investigating failures in the actual application, even when its own checks looked reassuring. It preserved old reports and an operator’s note and produced the final handover.

Ornith

Repeated a settled decision

During the second request, it kept reopening a policy question after useful checks had already established a workable approach. No finished second repair followed.

Flint

Lost earlier working repairs

During the upgrade, it brought back an older version and lost fixes for corrections, credits, and report history. Its own checks caught the breakage, but it then repeated its investigation without finishing.

Why Raon leads for longer jobs: it demonstrated continued useful action and a real handover. That is stronger evidence of handling an extended assignment. The completed review did not isolate what started the other models’ repetition. Evidence D ↗
E

What happens after it makes a mistake?

Finding, fixing, and admitting a mistake is valuable. It does not make the original mistake disappear.

Raon

A substantial recovery from its own error

Temporarily lost a credit and produced a $46 report. It noticed the missing credit, fixed the cause, restored the original entry, and produced $44. It later disclosed the restoration.

Ornith

Accurate first explanation

Its explanation of the first repair matched the saved work. There was no equivalent completed later recovery to assess.

Flint

Mixed evidence

Its checks honestly exposed the later lost fixes, and it did not claim the upgrade was finished. Earlier, however, its completion claims did not account for the changed correction files.

Why Raon earns credit without a free pass: it caught and explained a real mistake. But the incorrect $46 report still exists in the history. “It eventually repaired it” is different from “the records were preserved throughout.” Evidence E ↗
F

“The backup is safe” needs a check that proves it

A good backup check must compare the backup with the actual records, not just recognize familiar example contents.

Raon

Its reassurance went too far

Its check looked for the sample’s expected contents instead of comparing the real records with the backup. The saved backup checks passed their examples, but review of the finished work found this wider weakness.

Ornith

No backup result

Never reached this request. There is no basis for saying it would have done better or worse.

Flint

No completed backup handover

Started the upgrade but did not finish it. Its unfinished work cannot support a claim of a safer final backup.

Why this limits the recommendation: Raon left the only completed handover, but its general promise of a verified backup was stronger than its check justified. This weakness was found by reading the saved work; the wider example was not newly run. Evidence F ↗
04

Why I would choose one over another

Choose a pair to see the practical judgment, including the evidence against it.

Raon for the long assignment; Ornith for the focused repair

Raon demonstrated more useful follow-through. It connected successive changes, worked through real failures, and left a handover. Ornith’s clean first repair and stronger earlier short-task record still matter: Raon also introduced the repeated-credit mistake that Ornith avoided.

I would lean toward Raon when the model must carry a job forward over several requests. I would keep Ornith for focused repairs and a second opinion. Further evidence about what started the repetition could change how much weight to put on its stalled attempt.

Raon for preserving the job; Flint when speed can be checked

Raon preserved the supplied correction files and reached the handover. Flint made a genuinely good recovery repair and correctly kept repeated credits, where Raon failed. Its faster progress is useful.

But changing the supplied records and later losing working repairs are serious reasons to prefer Raon for work left running on its own. Flint’s successful total is less convincing when the input evidence changed. Raon’s quiet errors still need independent checks.

Ornith has the stronger earlier repair record; Flint got further here

The latest run alone does not establish that Ornith is more capable. Flint completed more of it and delivered a strong recovery repair. Ornith’s first repair was cleaner, and the earlier panel favored its consistency on short repairs.

I would place Ornith ahead as a cautious general repair choice, but this is the least secure pairwise judgment. Flint deserves credit for the work it completed; Ornith’s loop and Flint’s changed records are different problems.

05

Where MiniCPM, Apodex, and Sage fit

These three were in the earlier comparison of six separate short problems. They were not included in this latest continuing-job test, so they have not lost a head-to-head long-job contest.

MiniCPM

Earlier evidence only · Not retested in this round

A useful fast choice when the result is easy to check. It handled two demanding repairs well and gave a strong answer about what the evidence could prove. A different repair became impractically slow on a harder input, and two reasoning answers drew conclusions the evidence did not justify. Those are specific weaknesses, not a reason to dismiss it.

Apodex

Earlier evidence only · Not retested in this round

A useful additional analytical opinion. It gave the most complete answer on one evidence-and-cause problem. Other answers assumed facts that were not established or confused what caused a problem with what could fix it. Its repairs passed the original checks, although one efficiency explanation was wrong. The tested installation was an experimental version.

Sage

Earlier evidence only · Not retested in this round

Some useful reasoning, but too little finished repair work to recommend as the main model. It completed a demanding rules-and-records task, but on two different repair problems it kept reopening settled arguments and delivered no repair. Its faster text generation does not make up for those missing results.

The earlier work used six different problems per model, one attempt at each. It supports these narrower descriptions, not a universal ranking. Read the earlier findings ↗

06

What I would use today

A practical choice based on the work actually observed, with the remaining uncertainty kept visible.

Raon for longer Halo jobs, provisionally.

It gave the best evidence of carrying the whole assignment forward. Keep Ornith for focused repairs and a second opinion. Use Flint or MiniCPM for faster work whose result you can check. Apodex remains useful for another analytical view. Sage has not shown enough reliable repair completion to take the main role.

Raon’s finished answer still needs checking against the original records. The repeated-credit error, skipped corrections, and weak backup check are real reasons to withhold blanket trust.

This is a recommendation about the tested installations. It has not changed your default model.

What could change my mind?

  • A demonstrated setup cause for the loops. That would weaken the case against Ornith or Flint as models, even though these particular attempts still stopped.
  • The same pattern in another kind of application. One extended job is useful evidence, but not proof that the order will hold everywhere.
  • Whether Raon’s hidden mistakes repeat. Completing more work is less valuable if a wrong bill or skipped correction is easy to miss.

What this report does not assume

More thinking does not automatically mean a better answer. Finishing is not the same as being right. A missing result is not a proven wrong answer.

All three handled the first problem well enough to pass its saved checks. The more revealing differences appeared when the requirements interacted and earlier work had to survive.

There is no blended score here. It would hide the difference between an unfinished job, a wrong amount, and changed evidence.

How this was checked · sources and limits

The comparison uses saved application changes, recorded actions, independent checks, and two reviews conducted without model names. Both reviews found the central Raon and Flint problems described here. The plain-language explanations above are paraphrases of those findings.

Amounts in the original saved checks are in cents; this report expresses them in dollars. No new model attempts were run to make this report. The supporting records are selected publication copies with local machine paths removed; their file checks and publication notes are retained alongside it.

Read this comparison fairly

Scope: three models, one application, four dependent requests. A second planned application was withdrawn before the models saw it. The earlier comparison had six distinct short problems for each of six models. Neither is a broad, repeated survey of everything a model can do.

Unfinished work: Ornith did not complete request two; Flint did not complete request four. Requests beyond those points have no completed result to grade. Both repetitions were stopped. The completed review found a monitoring failure but did not isolate why the repetition began.

Specific limits: Flint’s changed correction files are one continuing problem, not multiple separate penalties. Its early manual reset affected tables that were already empty; no erasure of populated records was demonstrated. Raon restored its temporarily lost credit, and the preserved old reports were not lost. Its broader backup weakness was established by reading the saved work, not by running a new failure example.

Recommendation: the observed follow-through favors Raon for a longer assignment. The earlier focused-repair evidence still favors Ornith. There is not enough matched long-job evidence to rank MiniCPM or Apodex against Raon.