# Ciru main-model recommendation

> Update, 23 October 2026: The [completed sustained four-turn comparison](latest-readout.md) now provisionally favors Raon Q8 for long agent jobs, with important correctness defects. The six-form panel and its earlier Ornith-first recommendation below are preserved as prior evidence.

**Keep Ornith1.5 Ciru Ultra as the main model. Use MiniCPM4.7 for fast, readily checked work, and Apodex as an additional analytical perspective. Sage is installed, but its current package should not replace the main coding agent.**

This recommendation now uses **36 fresh attempts across six distinct authored forms per model**, under verified publisher settings: two algorithms, one ownership/atomicity contract, and three reasoning problems. There were 32 natural completions and four loop stops. It replaces the earlier premature briefing, which did not have enough corrected-settings evidence. These local custom tests support a practical package choice; they do not establish a universal base-model ranking.

## What changes the choice

Ornith delivered all three coding repairs with no concrete contract or complexity defect found in source review. Its incident analysis also separated observations from alternative causes well. It was not flawless: it overstated a harmed-complier bound, and its numerical JSON confused the requested removal direction with a valid repair in the opposite direction.

MiniCPM was much faster. Its new contract repair passed every mechanism group, and its new graph solution was correct and polynomial. Those are counterevidence to the earlier broad input-contract concern. However, one earlier program enumerated dense matches, and two distinct reasoning forms produced unsupported causal explanations. Its strong causal-bounds answer remains positive evidence. It is a useful fast option with external checks, rather than a model to dismiss.

Apodex gave the most complete causal-identification answer. It then treated temporal overlap as established in the incident task and confused numerical intervention direction with a sufficient cause in the arithmetic task. Its coding repairs passed the original checks, but one complexity explanation was false. These are findings about the tested experimental Q8 package, not a verdict on released IU4 or the underlying model.

Sage's two distinct algorithm attempts repeatedly reopened settled proofs without delivering a repair. Both loops were stopped with retained traces and snapshots. Its research answers contain useful reasoning, but the repeated autonomous-coding completion problem outweighs its measured MTP gain for this main-agent decision.

Raon remains a viable alternative, with strong algorithm and numerical results; its strict validation and causal-bound errors prevent a clear promotion here. Flint is a useful coding alternative, but two research attempts stopped in loops. One was interface-sensitive: it repeated an absolute-path write after explicit feedback requiring a relative path. A corrected argument was feasible and untried; this was not an impossible command or an accuracy zero.

No model is removed or declared generally incapable.

## Coding evidence

Each column is one authored task, not independent trials per assertion.

| Package | Signed circular range | Relabel/ownership contract | Bounded regex walk |
|---|---|---|---|
| Sage | Loop stopped; no authored repair | All five groups passed | Loop stopped; unchanged starter |
| Raon | Original correctness/scale and source review pass | Four of five groups; accepts forbidden dictionary/tuple inputs | 86/86; correct polynomial solution |
| Flint | Original correctness/scale and source review pass | All five groups passed | 86/86; correct polynomial solution |
| Ornith Ultra | Original correctness/scale and source review pass | All five groups passed | 86/86; correct layered-cost/tight-edge reconstruction |
| Apodex Q8 | Checks pass; stated complexity is false | All five groups passed | 86/86; correct polynomial solution |
| MiniCPM Q8 | Original checks pass; dense-match scaling defect | All five groups passed | 86/86; correct polynomial solution |

The range task had 2,427 small checks and three 200,000-element scale inputs. A separately declared saved-program diagnostic exposed MiniCPM's dense-match enumeration: it did not finish the 20,000-element case within eight seconds; all four retained comparator programs returned the correct answer in 0.11–0.38 seconds. That timeout is a scaling observation, not an accuracy zero.

Apodex has a different quadratic list-shifting family, but that exact family already passed at 200,000 elements in 1.54 seconds. Its false complexity claim is not an observed practical timeout. The new graph program does not repeat MiniCPM's earlier enumeration mechanism. All five completed graph repairs are distinct from the defective starter; visible examples alone would not establish this.

Sources: range/source findings (full archive; not included here), Apodex interpretation (full archive; not included here), contract finding (full archive; not included here), graph source review (full archive; not included here).

## Reasoning evidence

The single prose ordinals below belong only to their respective frozen rubrics. They are not added or averaged. The causal task instead has four separate ordinal axes and exact facts in its result file.

| Package | Causal identification | Marker/timeout incident | Rounding interventions |
|---|---|---|---|
| Sage | Six scalar facts right; harmed fraction upper bound wrong | 2: useful effects/probe; overlap and witness gaps | 3: valid repairs; requested removal flag wrong |
| Raon | Valid worlds; population-effect bounds wrong | 3: valid worlds/probe; unsupported response ordering | 3: correct requested facts; all-order proof asserted |
| Flint | Unavailable after reasoning loop | 3: valid worlds/probe; cheaper exclusions omitted | Unavailable after repeated invalid writes |
| Ornith Ultra | Six scalar facts right; harmed fraction upper bound wrong | 3: valid worlds, probe and cheaper exclusions | 3: valid repairs/nonunique responsibility; removal flag wrong |
| Apodex Q8 | Most complete on this form; correct sharp bounds | 2: temporal overlap treated as known | 1: intervention-direction and sufficient-cause errors |
| MiniCPM Q8 | Correct scalar and harmed-fraction bounds | Files absent; separate inline answer 1, with wrong discriminator/cause | 1: exact accumulation order wrongly made necessary |

For the causal problem, the sharp population-effect interval is [-0.36, 0.24] and the harmed-complier interval is [0, 0.25]. For the numerical problem, R=0, D=1/4, all twelve rounded-product reorder cases return 0, and both interventions return 0. **Restoring either rounding class to D repairs it; removing either class alone from R does not reproduce D.** Valid repair statements were preserved separately from wrong removal flags. MiniCPM delivered both files on this numerical form; the earlier missing-file episode is not a repeated delivery result.

Every research form received two independent anonymous reviews, fresh blind adjudication, and parent factual/excerpt checks before unblinding. Strict-checker false negatives for decorated incident schedules and a probe label were reconciled without changing frozen outputs. Missing files, schema errors, common wording ambiguities and reasoning errors remain separate.

Results: causal (full archive; not included here), incident (full archive; not included here), numerical (full archive; not included here), incident checker reconciliation (full archive; not included here).

## Latency tradeoff

Ornith's recommendation has a substantial time cost. Observed active API minutes on the range, contract and graph tasks were:

| Package | Range | Contract | Graph |
|---|---:|---:|---:|
| Ornith | 12.7 | 6.9 | 17.8 |
| Apodex | 15.5 | 6.5 | 5.5 |
| MiniCPM | 1.4 | 2.4 | 1.8 |

These single-task times include prefill and reasoning, exclude loading/parking/tools/grading, and use each publisher's mode. They are not fixed-output throughput or a strength score. For a quick change with good tests, MiniCPM can be the practical choice; for the default autonomous workload, this panel favors Ornith's more consistent artifact and causal behavior.

Timing metadata: 002 (full archive; not included here), 003 (full archive; not included here), 006 (full archive; not included here).

## Sage installation and learned speculation

Sage is installed as **model 37**, official Q8, native MTP3, vision, thinking on, configured 262,144 context. Raon remains **model 38**, Q8 with learned DFlash7. No ngram speculation was used.

The new Sage probe uses temperature1.0, top-p0.95, top-k20, presence penalty1.5, repetition penalty1.0 and seed123, with the actual template's open thinking prefix, one slot, F16 KV, and 512 generated tokens per call.

| Prompt | AR decode tokens/s | MTP3 decode tokens/s | AR elapsed | MTP3 elapsed |
|---|---:|---:|---:|---:|
| Code, 162 prompt tokens | 45.64 | 53.31 | 11.465 s | 9.883 s |
| Prose, 65 prompt tokens | 45.43 | 51.10 | 11.411 s | 10.167 s |

The code pair shows **16.8% faster decode** and **13.8% less whole-request elapsed time**. Native MTP counters report304/620 accepted drafts on code and302/625 on prose; loaded slot flags and runtime identities verified the modes.

Exactly four planned calls ran, with no repeats or warmups. After the first successful AR call, a metadata validator rejected the redundant disabled-mode label `none,none`. That result was retained, a reviewed label-normalization adapter ran only the remaining three calls, and raw echoes were preserved. Both code calls were first after loading. AR prose became first after a reload while MTP prose followed code, so the prose comparison has a cold/warm confound.

This is a directional short-prompt measurement, one call per cell, not a repeatability estimate or useful-answer latency test. Forced512 tokens include thinking. Earlier greedy measurements remain historical; the old63,792-token greedy probe had no whole-request gain, and no new long-context speed claim is made.

Four-call results (full archive; not included here), continuation review (full archive; not included here).

## Exact packages and limits

| Model | Tested package |
|---|---|
| Sage | Official Q8 GGUF; native MTP3; vision; C1/F16 KV |
| Raon | Official Q8; learned DFlash7; text; C1/F16 KV |
| Flint | Publisher-trained INT4 stored as Q4_1 GGUF; native runtime; no learned speculation |
| Ornith | Ciru Ultra Q8 custom vLLM; benchmark C8/44GiB; restored main C12 |
| Apodex | Experimental Ciru Q8 vLLM, MTP2, C8/16GiB; not llama.cpp, Q4 or released IU4 |
| MiniCPM4.7 | Healthy Ciru Q8 1.0.1 custom runtime, adaptive DSpark, C8/24GiB, instruct mode |

All fresh quality calls used positive-temperature publisher regimes with actual request receipts. vLLM received `repetition_penalty`, not the ignored llama.cpp spelling. The historical47-minute Apodex loop used seeded random sampling, not greedy, but lacked recommended repetition1.05. Earlier60-task data include38 sampler-mismatched attempts and24 wall-censored attempts; those data supplied hypotheses and do not override this corrected panel.

There were no new quality output/reasoning/turn caps or ten-minute deadlines. Real loops were stopped; long but productive attempts continued. This matters because correct Ornith/Apodex coding attempts exceeded the old cutoff. An unavailable answer is not an accuracy zero, and no earlier censored answer is assumed correct or hopeless.

These are six targeted forms with one trajectory per package/form, not an official benchmark suite or a statistical broad-capability ranking. Configuring256K does not qualify near-limit quality, vision quality or concurrent C12 serving. The unused second contract remained unexposed. No released-IU4 rescue, quantization sweep, or kernel port was added.

Ciru is restored to selected profile31, `ornith-ultra-c12`, active with its original disabled startup state. Selector and client state are preserved; private benchmark listener, owned controllers and all72 recorded sandbox owners are clear. No further generation is scheduled.

Final host receipt (full archive; not included here).
