Burrito Core’s maintainer claims its released GPT-OSS training and inference stack restores reliable tool calling and refusal behavior while sustaining fast 128K-context inference on a single RTX 3090, potentially making GPT-OSS more practical for local agents.

state: expiredheat: lowuncertainty: highconvergesscott: mediumlocal-inference inference-harnesses tool-calling open-modelsiamskeole skeoleBurrito Core

What is this?

Burrito Core is presented as a GPT-OSS training, evaluation, and inference stack maintained by iamskeole/skeole. Its maintainer claims that an inference harness restores reliable tool calling and refusal behavior while supporting 128K-context inference on one RTX 3090, backed by 320,192 evaluation runs across eight seeds, 3.49B tokens, and 1,062 GPU-hours. The supplied search results are unrelated to the software, so the release, benchmark methodology, performance, and behavioral improvements remain uncorroborated beyond the case’s own titles and summary.

Why it matters to Scott

The claim directly converges with Scott’s position that agent capability belongs to the model-plus-harness unit, and with his Ask compatibility layer for repairing inconsistent tool-call formats. If independently reproduced, the single-3090 long-context stack could be tested against his active Ollama/gamepc substrate and create a dated-receipts opportunity; for now, the claimed evaluation and performance gains remain uncorroborated.

ip:concept.model-plus-harness-benchmark-unitip:concept.evaluation-driven-developmentdev:project.askdev:concept.multi-format-tool-call-parsingdev:concept.hardware-aware-local-inferencedev:project.gamepcradar:concept.agent-harnessesradar:concept.tool-callingradar:concept.local-inferenceradar:stencil-harness-coding-improvementradar:vllm-silent-tool-parser-failuresradar:ctx-cliff-local-inference-benchmark

queries asked of Scott's wikis

local long-context inference economics on consumer GPUs
tool-calling reliability in agent harnesses
harness fixes versus model capability
open-weight models for local coding agents
behavioral evals for refusals and tool use
GPT-OSS deployment and agent integration

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (3) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	Don't trust me bro: 3.49B tokens, 320,192 evals, 8 seeds, at batch size 1 over 1,062 GPU hours on a single RTX 3090. And an inference harness that fixes gpt-oss. LocalLLaMA	skeole	0	28
🟧 hn	Show HN: Don't trust me bro: fixing GPT-OSS (3.49B tokens, 1k GPU hours, 1x3090)	iamskeole	2	0
🟧 echo.github ⭐	The author’s public write-up says: “320,192 runs. 8 seeds ... 1,062 GPU hours on a single RTX 3090.” It identifies burrito-evals as the comp	iamskeole	—	—

Interpretation history

2026-09-08T00:25:52Z
The refreshed discussion adds no technical results beyond the already-known template incorporation, and no concrete validation milestone is pending. Retire active monitoring as faded, not disproved; independent reliability tests or measured single-3090 long-context performance would justify reopening.
2026-09-06T00:22:57Z
This staleness check adds no substantive evidence: the downstream template incorporation remains a narrow adoption signal, not confirmation of reliability or long-context performance. Keep the stack as an optional local-agent experiment rather than a validated improvement, and reduce polling frequency.
2026-09-03T23:28:03Z
The refreshed discussion is repetitive amplification and skepticism, with no new implementation results, independent reproduction, or benchmark audit. The case remains a testable maintainer-reported harness whose core reliability and single-3090 performance claims are unvalidated.
2026-09-02T22:45:16Z
A downstream user says they incorporated Burrito Core’s fixes into an existing GPT-OSS Jinja template, adding a small implementation/adoption signal. This still does not independently reproduce the claimed reliability gains, refusal behavior, or 128K single-3090 performance, so the case remains unvalidated.
2026-09-01T22:23:35Z
The refreshed discussion adds no independent reproduction, benchmark audit, or technical counterevidence; it is further repetitive reaction to the maintainer’s existing claims. The case remains a released, testable harness with potentially useful local-agent implications, but not a validated advance.
2026-09-01T20:56:00Z
The refreshed comments remain repetitive skepticism about presentation and model choice, with no independent reproduction, benchmark audit, or technical counterevidence. The case still means a testable maintainer-reported harness release rather than a validated advance.
2026-09-01T19:02:55Z
The refreshed discussion adds an adjacent agent-side parser idea but no independent reproduction, benchmark audit, or substantive counterevidence. The case remains a testable maintainer-reported implementation awaiting external receipts.
2026-09-01T16:55:13Z
No independent reproduction or new technical evidence arrived; the added discussion is mostly skepticism about clarity and model choice rather than validation or substantive counterevidence. The released stack remains testable and relevant, but the episode has cooled pending receipts from another implementer.
2026-09-01T16:48:25Z
grounded: converges/medium — The claim directly converges with Scott’s position that agent capability belongs to the model-plus-harness unit, and with his Ask compatibility layer for repair
2026-09-01T16:46:22Z
origin walked (codex/luna, conf 0.97): anchor hn.story.49523381 -> echo.github.8fb39667c6 by iamskeole
2026-09-01T16:44:59Z
case created — The Reddit and Hacker News observations trace to the same substantial evaluation and released implementation, so they form one moving episode.

Decision trace

09-08 10:25expireThe refreshed discussion adds no technical results beyond the already-known template incorporation, and no concrete validation milestone is pending. Retire active monitoring as faded, not disproved; i
09-08 10:25alert_silentThere is no new consequential release, implementation result, or technical finding to surface. The maintainer's existing claims remain potentially useful but do not create a new reason to interru
09-08 10:25alert_routeThere is no new consequential release, implementation result, or technical finding to surface. The maintainer's existing claims remain potentially useful but do not create a new reason to interru
09-06 10:22repriceThis staleness check adds no substantive evidence: the downstream template incorporation remains a narrow adoption signal, not confirmation of reliability or long-context performance. Keep the stack a
09-06 10:22alert_silentThere is no new release, access change, implementation result, or credible technical finding to bring forward today; the comment-count change does not warrant interrupting Scott.
09-06 10:22alert_routeThere is no new release, access change, implementation result, or credible technical finding to bring forward today; the comment-count change does not warrant interrupting Scott.
09-04 09:28repriceThe refreshed discussion is repetitive amplification and skepticism, with no new implementation results, independent reproduction, or benchmark audit. The case remains a testable maintainer-reported h
09-04 09:28alert_silentThree additional comments do not materially change the technical evidence or Scott-facing implications; wait for an external reproduction, benchmark audit, or downstream test results.
09-04 09:28alert_routeThree additional comments do not materially change the technical evidence or Scott-facing implications; wait for an external reproduction, benchmark audit, or downstream test results.
09-03 08:45repriceA downstream user says they incorporated Burrito Core’s fixes into an existing GPT-OSS Jinja template, adding a small implementation/adoption signal. This still does not independently reproduce the cl
09-03 08:45alert_silentThe new downstream incorporation is noteworthy but provides no reported test results or benchmark audit, so it does not materially validate the core claims and can wait for the next briefing.
09-03 08:45alert_routeThe new downstream incorporation is noteworthy but provides no reported test results or benchmark audit, so it does not materially validate the core claims and can wait for the next briefing.
09-02 08:23repriceThe refreshed discussion adds no independent reproduction, benchmark audit, or technical counterevidence; it is further repetitive reaction to the maintainer’s existing claims. The case remains a rele
09-02 08:23alert_silentThe new delta is only comment churn and does not change confidence in the tool-calling, refusal-behavior, or single-3090 performance claims; an external reproduction or audit can wait for the normal b
09-02 08:23alert_routeThe new delta is only comment churn and does not change confidence in the tool-calling, refusal-behavior, or single-3090 performance claims; an external reproduction or audit can wait for the normal b
09-02 08:21sensor_dirtycomment_update
09-02 06:56repriceThe refreshed comments remain repetitive skepticism about presentation and model choice, with no independent reproduction, benchmark audit, or technical counterevidence. The case still means a testabl
09-02 06:56alert_silentThe new delta is only discussion churn and does not alter the established release or substantiate its tool-calling, refusal, or single-3090 performance claims; it can wait for an external reproduction
09-02 06:56alert_routeThe new delta is only discussion churn and does not alter the established release or substantiate its tool-calling, refusal, or single-3090 performance claims; it can wait for an external reproduction
09-02 06:22sensor_dirtycomment_update
09-02 05:02repriceThe refreshed discussion adds an adjacent agent-side parser idea but no independent reproduction, benchmark audit, or substantive counterevidence. The case remains a testable maintainer-reported imple
09-02 05:02alert_silentThe new comments neither validate nor materially challenge the released stack’s tool-calling, refusal, or single-3090 performance claims, so there is no consequential delta that needs to beat the next
09-02 05:02alert_routeThe new comments neither validate nor materially challenge the released stack’s tool-calling, refusal, or single-3090 performance claims, so there is no consequential delta that needs to beat the next
09-02 03:22sensor_dirtycomment_update
09-02 02:55repriceNo independent reproduction or new technical evidence arrived; the added discussion is mostly skepticism about clarity and model choice rather than validation or substantive counterevidence. The relea
09-02 02:55alert_silentThe new delta is only minor, skeptical discussion and does not change the established fact of release or validate the maintainer’s capability and performance claims; it can wait for independent reprod
09-02 02:55alert_routeThe new delta is only minor, skeptical discussion and does not change the established fact of release or validate the maintainer’s capability and performance claims; it can wait for independent reprod
09-02 02:50alert_shadowThe public code and reported 320,192-run evaluation record make this testable now and directly relevant to Scott’s model-plus-harness thesis, Ask compatibility work, and local 3090-class substrate. Th
09-02 02:50alert_routeThe public code and reported 320,192-run evaluation record make this testable now and directly relevant to Scott’s model-plus-harness thesis, Ask compatibility work, and local 3090-class substrate. Th
09-02 02:48groundThe claim directly converges with Scott’s position that agent capability belongs to the model-plus-harness unit, and with his Ask compatibility layer for repairing inconsistent tool-call formats. If i
09-02 02:46promote_anchororigin walk conf 0.97
09-02 02:45createThe Reddit and Hacker News observations trace to the same substantial evaluation and released implementation, so they form one moving episode.