| <!doctype html> |
| <html lang="en"> |
| <head> |
| <meta charset="utf-8"> |
| <meta name="viewport" content="width=device-width, initial-scale=1"> |
| <title>PatchCode (agentic-coder) β Testing Process β Qwen3.6 AEON RYS Docs</title> |
| <meta name="description" content="Historical public record for the original Qwen3.6 AEON RYS SignalLatch ckpt386 s0.10 fine-tune, testing process, runtime setup, and first-release decision."> |
| <style> |
| :root { |
| color-scheme: light; |
| --paper: #f7f5ee; |
| --paper-2: #ede9dc; |
| --ink: #111816; |
| --muted: #586560; |
| --line: #d2d7ce; |
| --panel: #ffffff; |
| --panel-2: #fbfaf5; |
| --code: #eef1ea; |
| --green: #0f7667; |
| --green-dark: #09564c; |
| --blue: #275f86; |
| --rust: #9a5a1c; |
| --red: #943b34; |
| --gold: #b48412; |
| --shadow: 0 18px 46px rgba(17, 24, 22, 0.10); |
| --soft-shadow: 0 10px 24px rgba(17, 24, 22, 0.07); |
| --radius: 8px; |
| } |
| |
| * { box-sizing: border-box; } |
| html { scroll-behavior: smooth; } |
| body { |
| margin: 0; |
| color: var(--ink); |
| background: |
| linear-gradient(90deg, rgba(17, 24, 22, 0.035) 1px, transparent 1px) 0 0 / 38px 38px, |
| linear-gradient(180deg, var(--paper), #fbfaf5 34rem, var(--paper)); |
| font: 16px/1.62 ui-sans-serif, system-ui, -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; |
| letter-spacing: 0; |
| overflow-x: hidden; |
| } |
| |
| a { |
| color: var(--green-dark); |
| text-decoration-thickness: 0.08em; |
| text-underline-offset: 0.18em; |
| } |
| |
| code, pre { |
| font-family: ui-monospace, SFMono-Regular, Menlo, Consolas, "Liberation Mono", monospace; |
| font-size: 0.93em; |
| } |
| |
| .wrap { |
| width: min(1180px, calc(100% - 36px)); |
| margin: 0 auto; |
| } |
| |
| header { |
| position: sticky; |
| top: 0; |
| z-index: 20; |
| background: rgba(247, 245, 238, 0.92); |
| border-bottom: 1px solid var(--line); |
| backdrop-filter: blur(14px); |
| } |
| |
| nav { |
| display: flex; |
| align-items: center; |
| justify-content: space-between; |
| gap: 18px; |
| min-height: 58px; |
| font-size: 0.9rem; |
| } |
| |
| nav .brand { |
| color: var(--ink); |
| font-weight: 900; |
| white-space: nowrap; |
| text-decoration: none; |
| } |
| |
| nav .links { |
| display: flex; |
| align-items: center; |
| justify-content: flex-end; |
| gap: 12px 16px; |
| flex-wrap: wrap; |
| } |
| |
| nav .links a { |
| color: var(--muted); |
| text-decoration: none; |
| font-weight: 700; |
| } |
| |
| nav .links a:hover { color: var(--green-dark); } |
| |
| .hero { |
| padding: 76px 0 46px; |
| border-bottom: 1px solid var(--line); |
| background: |
| linear-gradient(180deg, rgba(237, 233, 220, 0.9), rgba(247, 245, 238, 0.22)), |
| linear-gradient(90deg, rgba(15, 118, 103, 0.08), rgba(39, 95, 134, 0.08)); |
| } |
| |
| .eyebrow { |
| display: inline-flex; |
| align-items: center; |
| gap: 10px; |
| color: var(--rust); |
| font-size: 0.78rem; |
| font-weight: 900; |
| text-transform: uppercase; |
| letter-spacing: 0.08em; |
| margin-bottom: 18px; |
| } |
| |
| .eyebrow::before { |
| content: ""; |
| width: 34px; |
| height: 2px; |
| background: var(--rust); |
| } |
| |
| h1 { |
| margin: 0; |
| max-width: 1080px; |
| font-size: clamp(2.35rem, 5.3vw, 5.6rem); |
| line-height: 0.99; |
| letter-spacing: 0; |
| overflow-wrap: anywhere; |
| } |
| |
| .dek { |
| max-width: 920px; |
| margin: 24px 0 0; |
| color: var(--muted); |
| font-size: clamp(1.03rem, 1.7vw, 1.28rem); |
| } |
| |
| .actions { |
| display: flex; |
| gap: 12px; |
| flex-wrap: wrap; |
| margin-top: 28px; |
| } |
| |
| .button { |
| display: inline-flex; |
| align-items: center; |
| justify-content: center; |
| min-height: 42px; |
| padding: 0 15px; |
| border: 1px solid var(--line); |
| border-radius: 7px; |
| background: var(--panel); |
| color: var(--ink); |
| text-decoration: none; |
| font-weight: 800; |
| box-shadow: var(--soft-shadow); |
| } |
| |
| .button.primary { |
| background: var(--green); |
| border-color: var(--green); |
| color: #ffffff; |
| } |
| |
| .quick { |
| display: grid; |
| grid-template-columns: repeat(4, minmax(0, 1fr)); |
| gap: 12px; |
| margin-top: 36px; |
| } |
| |
| .metric { |
| min-height: 126px; |
| background: var(--panel); |
| border: 1px solid var(--line); |
| border-radius: var(--radius); |
| padding: 16px; |
| box-shadow: var(--soft-shadow); |
| } |
| |
| .metric strong { |
| display: block; |
| font-size: 1.72rem; |
| line-height: 1.08; |
| margin-bottom: 8px; |
| overflow-wrap: anywhere; |
| } |
| |
| .metric span { |
| display: block; |
| color: var(--muted); |
| font-size: 0.92rem; |
| } |
| |
| main { padding: 26px 0 74px; } |
| |
| section { |
| padding: 42px 0; |
| border-top: 1px solid var(--line); |
| } |
| |
| h2 { |
| margin: 0 0 16px; |
| font-size: clamp(1.7rem, 3vw, 2.65rem); |
| line-height: 1.12; |
| letter-spacing: 0; |
| } |
| |
| h3 { |
| margin: 0 0 10px; |
| font-size: 1.12rem; |
| letter-spacing: 0; |
| } |
| |
| p { |
| margin: 0 0 16px; |
| max-width: 900px; |
| } |
| |
| .muted { color: var(--muted); } |
| |
| .grid-2 { |
| display: grid; |
| grid-template-columns: minmax(0, 0.98fr) minmax(320px, 1.02fr); |
| gap: 32px; |
| align-items: start; |
| } |
| |
| .cards { |
| display: grid; |
| grid-template-columns: repeat(2, minmax(0, 1fr)); |
| gap: 14px; |
| } |
| |
| .card { |
| background: var(--panel); |
| border: 1px solid var(--line); |
| border-radius: var(--radius); |
| padding: 18px; |
| box-shadow: var(--shadow); |
| } |
| |
| .card.flat { |
| background: var(--panel-2); |
| box-shadow: none; |
| } |
| |
| .card.good { border-left: 5px solid var(--green); } |
| .card.warn { border-left: 5px solid var(--gold); } |
| .card.risk { border-left: 5px solid var(--red); } |
| .card.blue { border-left: 5px solid var(--blue); } |
| |
| .card p:last-child { margin-bottom: 0; } |
| |
| .callout { |
| border-left: 4px solid var(--green); |
| background: rgba(15, 118, 103, 0.08); |
| padding: 18px 20px; |
| border-radius: 0 var(--radius) var(--radius) 0; |
| margin: 18px 0; |
| max-width: 980px; |
| } |
| |
| .callout.warn { |
| border-left-color: var(--gold); |
| background: rgba(180, 132, 18, 0.10); |
| } |
| |
| .flow { |
| display: grid; |
| grid-template-columns: repeat(5, minmax(0, 1fr)); |
| gap: 10px; |
| margin: 20px 0; |
| } |
| |
| .flow .step { |
| background: var(--panel); |
| border: 1px solid var(--line); |
| border-radius: var(--radius); |
| padding: 14px; |
| min-height: 118px; |
| } |
| |
| .step b { |
| display: block; |
| color: var(--green-dark); |
| margin-bottom: 8px; |
| } |
| |
| .figure { |
| margin: 24px 0 0; |
| padding: 14px; |
| border: 1px solid var(--line); |
| border-radius: var(--radius); |
| background: var(--panel); |
| box-shadow: var(--soft-shadow); |
| } |
| |
| .figure img { |
| display: block; |
| width: 100%; |
| height: auto; |
| border-radius: 6px; |
| background: #ffffff; |
| } |
| |
| .caption { |
| margin-top: 10px; |
| color: var(--muted); |
| font-size: 0.9rem; |
| } |
| |
| .table-wrap { |
| overflow-x: auto; |
| border: 1px solid var(--line); |
| border-radius: var(--radius); |
| background: var(--panel); |
| box-shadow: var(--soft-shadow); |
| margin: 18px 0 26px; |
| } |
| |
| table { |
| width: 100%; |
| min-width: 740px; |
| border-collapse: collapse; |
| font-size: 0.92rem; |
| } |
| |
| th, td { |
| padding: 11px 12px; |
| border-bottom: 1px solid var(--line); |
| text-align: left; |
| vertical-align: top; |
| } |
| |
| th { |
| background: var(--paper-2); |
| color: var(--ink); |
| font-weight: 900; |
| } |
| |
| tr:last-child td { border-bottom: 0; } |
| .num { text-align: right; white-space: nowrap; } |
| .pass { color: var(--green-dark); font-weight: 800; } |
| .fail { color: var(--red); font-weight: 800; } |
| .soft { color: var(--muted); } |
| |
| pre { |
| overflow-x: auto; |
| background: var(--code); |
| border: 1px solid var(--line); |
| border-radius: var(--radius); |
| padding: 16px; |
| color: var(--ink); |
| white-space: pre-wrap; |
| word-break: break-word; |
| } |
| |
| .split-list { |
| display: grid; |
| grid-template-columns: repeat(2, minmax(0, 1fr)); |
| gap: 16px; |
| margin-top: 18px; |
| } |
| |
| ul { |
| padding-left: 20px; |
| margin: 0 0 16px; |
| } |
| |
| li { margin: 6px 0; } |
| |
| footer { |
| padding: 32px 0 52px; |
| border-top: 1px solid var(--line); |
| color: var(--muted); |
| font-size: 0.92rem; |
| } |
| |
| @media (max-width: 920px) { |
| nav { align-items: flex-start; flex-direction: column; padding: 12px 0; } |
| nav .links { justify-content: flex-start; } |
| .hero { padding-top: 46px; } |
| .quick { grid-template-columns: repeat(2, minmax(0, 1fr)); } |
| .grid-2 { grid-template-columns: 1fr; } |
| .flow { grid-template-columns: 1fr; } |
| .cards, .split-list { grid-template-columns: 1fr; } |
| } |
| |
| @media (max-width: 560px) { |
| .wrap { width: min(100% - 24px, 1180px); } |
| .quick { grid-template-columns: 1fr; } |
| .metric { min-height: auto; } |
| table { min-width: 680px; } |
| } |
| </style> |
| <style>main img{max-width:100%;height:auto;display:block;margin:1.4em auto .5em;border-radius:6px}main table{display:block;overflow-x:auto}</style></head> |
| <body> |
| <header> |
| <div class="wrap"> |
| <nav aria-label="Page navigation"> |
| <a class="brand" href="../index.html">Qwen3.6 AEON RYS Docs</a> |
| <div class="links"> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/">Docs home</a> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/qwen36-aeon-rys-signallatch/index.html">SignalLatch</a> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/ckpt386-s010-testing-process/index.html">ckpt386 process</a> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/signallatch-v1-1-production-process/index.html">v1.1 production</a> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/signallatch-v1-1-all-results/index.html">v1.1 all results</a> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/rys-layer-duplication-guide/">RYS arch guide</a> |
| <a href="https://noonr48.github.io/qwen36-aeon-ik-llama/qwen36-aeon-rys-15-20/index.html">AEON 15/20</a> |
| </div> |
| </nav> |
| </div> |
| </header> |
|
|
| <div class="hero"> |
| <div class="wrap"> |
| <div class="eyebrow">PatchCode fine-tune record</div> |
| <h1>How the PatchCode agentic-coder fine-tune was made, tested, and selected.</h1> |
| <p class="dek">This page is the self-contained public record for the PatchCode behaviour distil — an agentic-coder joint LoRA (checkpoint 3661, merged at λ=0.5) on top of the SignalLatch release. It covers the IQ4_NL quant bake-off, the run-to-run noise analysis, and the exact dataset pipeline.</p> |
| <div class="actions"> |
| <a class="button primary" href="https://huggingface.co/jackasda211233/Qwen3.6-27B-AEON-RYS-Agentic-Coder-PatchCode-GGUF">PatchCode on Hugging Face</a> |
| <a class="button" href="https://noonr48.github.io/qwen36-aeon-ik-llama/ckpt386-s010-testing-process/index.html">SignalLatch record</a> |
| <a class="button" href="https://noonr48.github.io/qwen36-aeon-ik-llama/qwen36-aeon-rys-signallatch/index.html">SignalLatch overview</a> |
| <a class="button" href="https://noonr48.github.io/qwen36-aeon-ik-llama/rys-layer-duplication-guide/">RYS arch guide</a> |
| <a class="button" href="https://github.com/noonr48/qwen36-aeon-ik-llama">Runtime fork</a> |
| </div> |
| <div class="quick" aria-label="Key release facts"> |
| <div class="metric"><strong>ckpt-3661</strong><span>Final one-epoch agentic-coder joint LoRA checkpoint used for the merge.</span></div> |
| <div class="metric"><strong>λ = 0.5</strong><span>Selected merge strength (effective alpha/r = 1.0) — the trained default was over-applied.</span></div> |
| <div class="metric"><strong>IQ4_NL</strong><span>Shipped 16.6 GB GGUF — ties BF16 within noise at ~⅓ the size.</span></div> |
| <div class="metric"><strong>~58.5k</strong><span>Agentic-coding behaviour examples in the training blend (synthetic backbone + curated style slice).</span></div> |
| </div> |
| </div> |
| </div> |
| <main class="wrap"><div class="wrap"> |
| <p>This is the longer, more casual write-up for the PatchCode upload candidate (internal project name <code>merged_lam0.5</code>).</p> |
| <p>The clean model card stays short. This document is the full story: what we distilled, exactly how the dataset was built, how we tested it, why the early single-run scores fooled us, why we stopped trusting them, and why the upload candidate ended up being the plain <code>IQ4_NL</code> (reasoning-imatrix) merged GGUF rather than a heavier mixed-quant recipe.</p> |
| <p>Related public guides: |
| - runtime fork: <code>https://github.com/noonr48/qwen36-aeon-ik-llama</code> |
| - RYS layer-duplication / architecture guide: <code>https://github.com/noonr48/qwen36-aeon-ik-llama/tree/main/docs/rys-layer-duplication-guide</code> |
| - previous fine-tuned release (SignalLatch): <code>https://huggingface.co/jackasda211233/Qwen3.6-27B-AEON-RYS-SignalLatch-GGUF</code></p> |
| <p>Related release line: |
| - previous finetune: <code>Qwen3.6-27B-AEON-RYS-SignalLatch-ckpt386-s010-IQ4_NL.gguf</code> |
| - this upload candidate: <code>Qwen3.6-27B-AEON-RYS-Agentic-Coder-PatchCode.IQ4_NL.gguf</code></p> |
| <h2>Glossary</h2> |
| <ul> |
| <li><code>AEON</code>: the upstream/source model family this RYS line was built from (<code>AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored</code>).</li> |
| <li><code>SignalLatch</code> / <code>ckpt386-s010</code>: the previous finetune in this line β a behaviour LoRA (checkpoint 386) merged into the AEON RYS base at strength <code>0.10</code>. PatchCode is built on top of this.</li> |
| <li><code>PatchCode</code> / <code>merged_lam0.5</code>: the public name for this release. It is a second behaviour distil (an agentic-coder joint LoRA) merged onto SignalLatch at strength <code>0.5</code>.</li> |
| <li><code>IQ4_NL</code>: the quantized GGUF deployment format we actually upload and run.</li> |
| <li><code>imatrix</code>: importance-matrix-assisted quantization data. <code>reasoning-imatrix</code> = calibrated on reasoning/coding text (the kind that worked); <code>media-imatrix</code> = an earlier calibration kind that underperformed.</li> |
| <li><code>ik-llama</code>: the custom runtime fork. The <code>qwen3_5</code> hybrid architecture does not load on stock <code>llama.cpp</code> / <code>vLLM</code>.</li> |
| <li><code>KritaLite</code>: our hardened real-world discriminator build (a ~160k-token multi-file app, 15 binary verifier components). Single-shot coding gates saturate on this model family, so we stopped trusting them.</li> |
| <li><code>discipline</code> / <code>style_discipline</code>: a rubric measuring the distilled action-first style (no preamble, claim-requires-run, narrateβactβverify).</li> |
| </ul> |
| <h2>The short version</h2> |
| <p>We started from the SignalLatch finetune and distilled a second, agentic-coder behaviour LoRA on top of it. The goal was not a new general chat model. The goal was to make the model a better coding agent: action-first execution, claims backed by an actual run, systematic diagnoseβfix loops, stable multi-turn tool use, and fewer stalled runs.</p> |
| <p>After a full 5-phase bake-off, the model that held up was:</p> |
| <pre><code class="language-text">Qwen3.6-27B-AEON-RYS-Agentic-Coder-PatchCode.IQ4_NL.gguf |
| </code></pre> |
| <p>That means: |
| - base: <code>Qwen3.6-27B-AEON-RYS-SignalLatch-ckpt386-s010</code> |
| - adapter: agentic-coder joint LoRA, checkpoint <code>3661</code> |
| - merge strength: <code>0.5</code> (effective alpha/r = 1.0) |
| - deploy format: plain <code>IQ4_NL</code> with reasoning-imatrix |
| - runtime: custom AEON ik-llama fork</p> |
| <p>The awkward part β and the reason this write-up is long β is that the eventual ship pick was <strong>not</strong> the candidate that looked best early. A mixed-quant recipe (<code>c76</code>) hit a perfect-looking build score on the first multi-seed pass and did not reproduce. A 5-seed, same-condition confirm reversed the read. The plain <code>IQ4_NL</code> ended up tied with everything else within noise, so the decision fell to non-noise axes (size, recipe safety), where plain <code>IQ4_NL</code> wins.</p> |
| <h2>What this was meant to upgrade</h2> |
| <p>PatchCode is an upgrade over the existing SignalLatch finetune:</p> |
| <pre><code class="language-text">Qwen3.6-27B-AEON-RYS-SignalLatch-ckpt386-s010-IQ4_NL.gguf |
| </code></pre> |
| <p>The new work was not another RYS architecture pass (the architecture is unchanged and is documented in the layer-duplication guide). The new work was a behaviour distil layered on top of SignalLatch, then merged and quantized into the same practical Q4-class deployment lane.</p> |
| <p>Public framing stays narrow:</p> |
| <blockquote> |
| <p>This is a practical coding-agent / tool-use-oriented fine-tuned IQ4_NL variant of the SignalLatch release.</p> |
| </blockquote> |
| <p>It should not be framed as: |
| - a universal upgrade over base in every format |
| - a general chat benchmark win |
| - a stock <code>llama.cpp</code> / <code>vLLM</code> model |
| - a live-LoRA deployment recipe</p> |
| <h2>The dataset β exact pipeline</h2> |
| <p>This is the part most people ask about, so it is written out in full. The training blend is <code>~58.5k</code> examples and is made of two pieces: a large <strong>synthetic coding-agent behaviour backbone</strong> and a smaller <strong>curated action-first style slice</strong>, blended together.</p> |
| <h3>Piece 1 β synthetic coding-agent behaviour backbone (~43k)</h3> |
| <p>A standalone synthetic generator produces multi-turn coding-agent traces. It is <strong>fully synthetic</strong> β no real user data, no scraped repos. The pipeline:</p> |
| <ol> |
| <li><strong>Behaviour-driven generation.</strong> A pool of parallel workers calls a coding-agent teacher model. Each call is shaped around a named <em>behaviour</em> from a fixed behaviour pool (~30 behaviours), for example: |
| - <code>survey_before_edit</code> β read/search the real context before touching code |
| - <code>hypothesis_driven_debugging</code> β form a hypothesis, then verify |
| - <code>tool_intent_first</code> β express tool intent before prose |
| - <code>weigh_alternatives_then_commit</code> β weigh β₯3 options, commit to one, verify |
| - <code>external_awareness</code> β check versions/docs before asserting |
| - <code>recall_first_habit</code> β recall prior context before re-deriving</li> |
| <li><strong>Tool-agnostic vocabulary (anti-lock-in).</strong> Tool calls use a behavioural-category vocabulary (e.g. <code>memory_search</code>, <code>repo_search</code>, <code>render_or_visual_proof</code>), not real tool names. This is deliberate: the model learns <em>when/why to use a tool</em>, not a specific vendor's API surface.</li> |
| <li><strong>Scenarios.</strong> A synthetic scenario bank provides repo-shaped task context (file trees, failing tests, stack traces) so the traces are grounded in realistic edit/verify loops.</li> |
| <li><strong>Quality gates (per sample).</strong> Traces that fail the gates are dropped, not emitted: |
| - <code>no-op-edit</code> guard (a claimed edit that changes nothing) |
| - <code>claim-without-verify</code> reject (the assistant claims done with no run/check) |
| - <code>reasoning-empty</code> / <code>incomplete-trace</code> / <code>lang-runner-mismatch</code> / <code>prompt-over-cap</code></li> |
| <li><strong>Deficit-resume scheduling.</strong> Generation runs continuously, tracks per-behaviour deficits, and resumes after interruption until target counts are met (~30 samples/sec).</li> |
| </ol> |
| <p><strong>Corpus assembly + filtering (exact counts):</strong> |
| - raw unified coding corpus: <code>71,776</code> samples |
| - filter drops <code>10,666</code> bad samples β <code>61,110</code> kept |
| - top drop reasons: <code>prompt_over_cap</code> 3,946 Β· <code>lang_runner_mismatch</code> 3,645 Β· <code>reasoning_empty</code> 2,086 Β· <code>incomplete_trace</code> 861 Β· <code>claim_without_verify</code> 620 |
| - coding training subset used for the blend: <code>43,075</code></p> |
| <p>The broader synthetic corpus spans five behaviour layers (media-behaviour 42,973 Β· tool-depth 15,242 Β· reliability 19,393 Β· self-correction 31,476 Β· coding 7,721 = <code>116,805</code> total before filtering); the blend draws the coding-oriented subset.</p> |
| <h3>Piece 2 β curated action-first style slice (~7k)</h3> |
| <p>A smaller slice of curated execution-style traces that model the exact discipline we wanted to amplify: terse narrateβactβverify, no preamble, claim-requires-run. Composition (<code>6,953</code> total): |
| - own multi-project execution sessions (<code>5,455</code>) β span many different projects on purpose, so the style generalises instead of locking to one domain |
| - a different-domain contributor (<code>1,130</code>) β explicitly included for cross-project transfer |
| - reasoning-chain exemplars (<code>368</code>) β weigh-alternatives deliberation seeds</p> |
| <p><strong>De-identification / anti-lock-in pass:</strong> real tool names, hostnames, absolute paths, and identifiers are abstracted to behavioural-category tokens / placeholders. The supervision is <strong>assistant-turn-only</strong> β system/user/tool turns (where real project content lives) are masked (<code>IGNORE_INDEX</code>), so the model learns a <em>behaviour policy conditioned on varied context</em>, not project facts as outputs.</p> |
| <h3>Piece 3 β the blend</h3> |
| <p>A small blender oversamples the style slice so it is not drowned by the larger coding backbone, then shuffles:</p> |
| <ul> |
| <li>coding backbone: <code>43,075</code></li> |
| <li>style slice oversampled ~2.2Γ</li> |
| <li>blended training set: <code>58,576</code> β <strong>~74% coding backbone / ~26% action-first style</strong></li> |
| </ul> |
| <p>The oversample ratio was chosen so the style shows up without overfitting the smaller slice; a held-out task type was used to check it generalises rather than parrots.</p> |
| <h3>What the dataset is <em>not</em></h3> |
| <ul> |
| <li>It is not scraped real-user data or real private repos.</li> |
| <li>It is not a single-topic dataset β both pieces deliberately span many projects/domains.</li> |
| <li>It does not teach new domain <em>facts</em>; it teaches an execution <em>discipline</em>.</li> |
| </ul> |
| <h2>The training piece</h2> |
| <p>A single LoRA was joint-co-trained on the blended <code>58.5k</code> set (one adapter, not two-then-merge β a prior two-adapter Ξ»-merge plan was superseded because post-hoc merges can kill a fragile capability with no usable Ξ»).</p> |
| <p>Training config: |
| - PEFT type: <code>LORA</code> |
| - rank: <code>r=32</code>, alpha: <code>64</code> (alpha/r = 2.0) |
| - dropout: <code>0.05</code> |
| - target modules: <strong>all-linear</strong>, including the hybrid arch projections β <code>q/k/v/o_proj</code>, <code>gate/up/down_proj</code>, <code>out_proj</code>, and the linear-attn/SSM projections <code>in_proj_qkv / in_proj_a / in_proj_b / in_proj_z</code> |
| - supervision: completion-only (assistant turns only) |
| - optimiser: adamw, lr <code>5e-5</code> + warmup + cosine decay |
| - epochs: <code>1</code> |
| - backend: model-parallel <code>device_map</code> across a multi-GPU host (the max-quality path; the no-NVLink fleet ruled out DeepSpeed/FSDP here)</p> |
| <p>Completion: |
| - <code>global_step=3661</code> = <code>epoch 1.0</code> complete |
| - final <code>train_loss β 0.853</code> |
| - runtime ~91h (~89.5 s/it), grad-norm steady (no divergence) |
| - 37 checkpoints saved across the run β full trajectory available for eval</p> |
| <p>The adapter was behaviour-focused and small. It was not trained to teach broad new knowledge.</p> |
| <h2>The merge β why Ξ»=0.5</h2> |
| <p>The trained default adapter strength (alpha/r = 2.0) was <strong>over-applied</strong>. A checkpoint Γ strength eval showed half-strength beat full-strength on all three tested checkpoints:</p> |
| <table> |
| <thead> |
| <tr> |
| <th>checkpoint</th> |
| <th style="text-align: right;">Ξ»=0.3</th> |
| <th style="text-align: right;">Ξ»=0.5</th> |
| <th style="text-align: right;">Ξ»=0.7</th> |
| <th style="text-align: right;">Ξ»=1.0</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr> |
| <td>3661</td> |
| <td style="text-align: right;">0.522</td> |
| <td style="text-align: right;"><strong>0.617</strong></td> |
| <td style="text-align: right;">0.490</td> |
| <td style="text-align: right;">0.491</td> |
| </tr> |
| <tr> |
| <td>2600</td> |
| <td style="text-align: right;">0.567</td> |
| <td style="text-align: right;"><strong>0.573</strong></td> |
| <td style="text-align: right;">β</td> |
| <td style="text-align: right;">0.442</td> |
| </tr> |
| <tr> |
| <td>1800</td> |
| <td style="text-align: right;">0.540</td> |
| <td style="text-align: right;"><strong>0.564</strong></td> |
| <td style="text-align: right;">β</td> |
| <td style="text-align: right;">0.397</td> |
| </tr> |
| </tbody> |
| </table> |
| <p>At Ξ»=1.0 the adapter was net-neutral-to-harmful (one checkpoint fell <em>below</em> the un-adapted base). The mechanism: an over-loud LoRA delta pushes activations into regimes that hurt calibrated behaviour (preamble returns, over-claiming). Ξ»=0.5 (effective alpha/r = 1.0) keeps the style direction but respects base calibration. So the merge was done at <strong>Ξ»=0.5 onto SignalLatch (ckpt386-s010)</strong>, then exported to BF16 GGUF. (A future v2 could bake the good strength in by training at alpha=r=32, removing the inference-time knob.)</p> |
| <p><img alt="Merge strength sweep β Ξ»=0.5 wins on all three checkpoints; trained default Ξ»=1.0 is over-applied." src="https://noonr48.github.io/qwen36-aeon-ik-llama/patchcode-testing-process/assets/lambda_sweep.png" /></p> |
| <h2>Why the final testing moved to merged IQ4_NL</h2> |
| <p>The key question was not "best adapter in BF16" β it was "what we would actually deploy". The deploy target was a merged GGUF, <code>IQ4_NL</code>, imatrix-quantized, on the custom ik-llama runtime (Jinja + DeepSeek reasoning format + flash attention + graph split, temp <code>0.7</code>).</p> |
| <p>Live LoRA loading is not the production path for this release (the tested serving profile uses flash attention, which conflicts with live LoRA on this runtime). So the long-term path became: <strong>merge the adapter first, then export + quantize a full GGUF.</strong> That is why the upload is a merged GGUF, not an adapter.</p> |
| <p>The plain <code>IQ4_NL</code> uses the <strong>reasoning/coding imatrix</strong> (the kind that worked). An earlier build used a media-domain imatrix; it underperformed and was superseded.</p> |
| <h2>Complete test catalog β every run, at a glance</h2> |
| <p>Thirteen separate test runs fed this decision, plus the behaviour-rubric Ξ»-sweep. This is the full list β what each measured, on what, and what it said. The detail for each follows in <em>The testing ladder</em>.</p> |
| <table> |
| <thead> |
| <tr> |
| <th>#</th> |
| <th>test</th> |
| <th>measures</th> |
| <th>candidates</th> |
| <th>conditions</th> |
| <th>headline result</th> |
| <th>verdict</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr> |
| <td>1</td> |
| <td>Phase 1 β single-seed KritaLite</td> |
| <td>160k-token real-world build</td> |
| <td>IQ4_NL, c76, c373, BF16</td> |
| <td>1 seed</td> |
| <td>IQ4_NL 0.933 vs c76 0.867</td> |
| <td><strong>noise</strong> β did not reproduce</td> |
| </tr> |
| <tr> |
| <td>2</td> |
| <td>Phase 2 β 3-seed KritaLite</td> |
| <td>build</td> |
| <td>8 quants + BF16</td> |
| <td>3 seeds</td> |
| <td>c76/c404 0.933; IQ4_NL/BF16 0.867</td> |
| <td>reversed phase 1; mixed recipe led</td> |
| </tr> |
| <tr> |
| <td>3</td> |
| <td>Phase 3 β 40-recipe broad search</td> |
| <td>build</td> |
| <td>40 mixed recipes</td> |
| <td>3 seeds</td> |
| <td>all-zero</td> |
| <td><strong>harness bug</strong> (missing <code>config.json</code>) β void</td> |
| </tr> |
| <tr> |
| <td>4</td> |
| <td>Phase 4 β search re-gate</td> |
| <td>build</td> |
| <td>53 candidates</td> |
| <td>bug fixed</td> |
| <td>none beat the curated originals</td> |
| <td>broad search doesn't help this merge</td> |
| </tr> |
| <tr> |
| <td>5</td> |
| <td>Phase 5 β discipline rubric</td> |
| <td>action-first style</td> |
| <td>5 quants</td> |
| <td>3 seeds</td> |
| <td>BF16/IQ4_NL 0.931; c76 0.903</td> |
| <td>IQ4_NL & BF16 lead discipline</td> |
| </tr> |
| <tr> |
| <td>6</td> |
| <td>Phase 5 β <code>agent_eval_http</code></td> |
| <td>7-task agentic pass-rate + turns</td> |
| <td>5 quants</td> |
| <td>1 pass</td> |
| <td>c76 27/7 turns; baseline 27/11; c373 31/17 (thrash)</td> |
| <td>c76 leads process-efficiency</td> |
| </tr> |
| <tr> |
| <td>7</td> |
| <td>Q5 confirm</td> |
| <td>build + long-context + discipline</td> |
| <td>Q5_K_M (uniform, 20 G)</td> |
| <td>3 seeds</td> |
| <td>0.867 / 0.988 / 0.806</td> |
| <td>doesn't clear "both" (build+disc β₯ 0.90)</td> |
| </tr> |
| <tr> |
| <td>8</td> |
| <td>Overnight 2 β precision Γ promotion matrix</td> |
| <td>build + discipline</td> |
| <td>8 (q5/q6/q8 Γ uniform/promoted)</td> |
| <td>3 seeds</td> |
| <td>none clear both; promotion kills discipline</td> |
| <td>precision is <strong>not</strong> the build lever</td> |
| </tr> |
| <tr> |
| <td>9</td> |
| <td><strong>Confirm β 5-seed head-to-head</strong></td> |
| <td>build + long-context + discipline</td> |
| <td>IQ4_NL vs c76</td> |
| <td><strong>5 seeds, same-condition</strong></td> |
| <td>IQ4_NL 0.920 vs c76 0.907 (Ξ 0.013 βͺ 0.067)</td> |
| <td><strong>TIED within noise β the decisive test</strong></td> |
| </tr> |
| <tr> |
| <td>10</td> |
| <td>Q8 confirm</td> |
| <td>build + long-context + discipline</td> |
| <td>Q8_0 vs IQ4_NL</td> |
| <td>5 seeds</td> |
| <td>Q8 0.867 vs IQ4_NL 0.920</td> |
| <td>no edge; near-lossless buys nothing</td> |
| </tr> |
| <tr> |
| <td>11</td> |
| <td>Agentic-loop</td> |
| <td>40 held-out mini-projects; pytest-verified convergence + turns + recovery + stall</td> |
| <td>IQ4_NL, c76, Q8</td> |
| <td>40 tasks</td> |
| <td><strong>all 100 % convergence, ~6.6β7.2 turns, 0 % stall</strong></td> |
| <td><strong>did not discriminate</strong> β every quant (incl. base) converges; a family property, not a PatchCode distinction</td> |
| </tr> |
| <tr> |
| <td>12</td> |
| <td>SignalLatch 4-gate suite</td> |
| <td>coding/habits + hard-reasoning + long-context (exact + rubric)</td> |
| <td>IQ4_NL vs BF16</td> |
| <td>n=12 hard, n=4 longctx</td> |
| <td>IQ4_NL 0.887 vs BF16 0.846 (hard); 0.979 vs 0.941 (long); 0 errors</td> |
| <td>IQ4_NL tracks/edges BF16 within noise</td> |
| </tr> |
| <tr> |
| <td>13</td> |
| <td>Behaviour rubric β Ξ»-sweep</td> |
| <td>action-first style + coding discipline + held-out generalization</td> |
| <td>base (SignalLatch) vs PatchCode @ Ξ»{0.3,0.5,0.7,1.0,1.3}, ckpt{3661,2600,1800}</td> |
| <td>15 cases Γ strengths</td> |
| <td>base 0.486 β PatchCode Ξ»0.5 0.617 (~β
the tokens)</td> |
| <td>PatchCode beats base; Ξ»0.5 is the sweet spot</td> |
| </tr> |
| </tbody> |
| </table> |
| <p><strong>The only test that discriminated was #9</strong> (the 5-seed confirm) β and it discriminated by showing everything is <em>tied within noise</em>, which pushed the decision onto non-noise axes (size + plain-quant recipe), where IQ4_NL wins. Tests #1 and #3 were void (noise / harness bug). Tests #5β#8, #10 and #11 all failed to separate the finalists. #12 confirms IQ4_NL is not a quality cliff below BF16. #13 is the one place PatchCode clearly beats its SignalLatch base.</p> |
| <h2>The testing ladder (5 phases + confirms)</h2> |
| <p>Single-shot and hard-suite gates <strong>saturate</strong> on this model family (every quant scores ~the same, including BF16). The discrimination that actually changed the decision came from a 160k-token real-world build (KritaLite) run multi-seed, plus a discipline rubric, plus an agentic-process efficiency probe. The phases:</p> |
| <p><strong>Phase 1 β single-seed real-world build.</strong> Made the plain <code>IQ4_NL</code> look like the winner (0.933 vs c76's 0.867). This was <strong>noise</strong> β it did not reproduce.</p> |
| <p><strong>Phase 2 β multi-seed KritaLite (3 seeds).</strong> Reversed phase 1: <code>c76</code>/<code>c404</code>/<code>c373</code> hit 0.933; plain <code>IQ4_NL</code> dropped to 0.867. Now a mixed-quant recipe looked like the winner.</p> |
| <p><strong>Phase 3 β 40-recipe broad search.</strong> Returned all-zero. Root cause was a <strong>harness bug</strong> (the eval script imports a <code>config.json</code> that was not copied into the eval root), not real scores.</p> |
| <p><strong>Phase 4 β search re-gate (bug fixed).</strong> Re-scored all 53 candidates correctly. No new recipe beat the curated originals; the broad search does not help this merge.</p> |
| <p><strong>Phase 5 β discipline + agentic process.</strong> Plain <code>IQ4_NL</code> and BF16 led the action-first <em>discipline</em> rubric (0.931); <code>c76</code> led <em>process efficiency</em> (fewest turns/tools/errors).</p> |
| <p><strong>Overnight 2 β base-precision Γ attention-promotion matrix (3 seeds).</strong> Decomposed the build/discipline tradeoff. No candidate clears "both" (build β₯ 0.90 <strong>and</strong> discipline β₯ 0.90): |
| - promotion destroys discipline regardless of base precision |
| - uniform higher precision does <strong>not</strong> fix build (build is not precision-limited)</p> |
| <p><strong>Confirm β 5-seed, same-condition, baseline vs c76 head-to-head.</strong> The decisive run:</p> |
| <table> |
| <thead> |
| <tr> |
| <th>candidate</th> |
| <th style="text-align: right;">build (5-seed)</th> |
| <th style="text-align: right;">long-context</th> |
| <th style="text-align: right;">discipline (5-seed)</th> |
| <th style="text-align: right;">size</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr> |
| <td><strong>plain IQ4_NL (reasoning imx)</strong></td> |
| <td style="text-align: right;">0.920 (Β±0.067)</td> |
| <td style="text-align: right;">0.975</td> |
| <td style="text-align: right;">0.842 (Β±0.333)</td> |
| <td style="text-align: right;">16.6 G</td> |
| </tr> |
| <tr> |
| <td>c76 (promoted attn)</td> |
| <td style="text-align: right;">0.907 (Β±0.067)</td> |
| <td style="text-align: right;">0.935</td> |
| <td style="text-align: right;">0.867 (Β±0.292)</td> |
| <td style="text-align: right;">20 G</td> |
| </tr> |
| </tbody> |
| </table> |
| <p>build gap <code>0.013</code> βͺ <code>0.067</code> noise floor β <strong>not discriminating</strong>. c76's earlier "0.933 build win" did not reproduce (it scored 0.933 β 0.867 β 0.907 across passes β pure run-to-run variance).</p> |
| <p><strong>Q8 confirm β 5-seed, near-lossless Q8 vs plain IQ4_NL.</strong> Q8 shows no edge on any axis and is ~2Γ the size β ruled out. Near-lossless precision buys nothing measurable here.</p> |
| <p><strong>Behaviour rubric β PatchCode vs the base it was distilled from.</strong> A 15-case rubric (action-first style + coding discipline + held-out generalization) was run across merge strengths, with the adapter disabled as the "strength 0" anchor β i.e. the SignalLatch base PatchCode was built on. PatchCode at the chosen Ξ»=0.5 beat the base on score while emitting far fewer tokens:</p> |
| <table> |
| <thead> |
| <tr> |
| <th>variant (15-case rubric)</th> |
| <th style="text-align: right;">score</th> |
| <th style="text-align: right;">avg output tokens</th> |
| <th style="text-align: right;">avg time/case</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr> |
| <td>base (adapter off = SignalLatch)</td> |
| <td style="text-align: right;"><code>0.486</code></td> |
| <td style="text-align: right;"><code>311</code></td> |
| <td style="text-align: right;"><code>34s</code></td> |
| </tr> |
| <tr> |
| <td>PatchCode (ckpt-3661 @ Ξ»=0.5)</td> |
| <td style="text-align: right;"><code>0.617</code></td> |
| <td style="text-align: right;"><code>91</code></td> |
| <td style="text-align: right;"><code>13s</code></td> |
| </tr> |
| </tbody> |
| </table> |
| <p>The base tended to ramble (~311 tokens of hedging preamble β e.g. it scored 0.20 on the coding-discipline case with "I might overwrite the user's changesβ¦"); PatchCode was terse and on-target (~91 tokens) and scored higher. That is the distil's intended effect: more disciplined execution, less wasted output. Caveats: this is a behaviour rubric, not a multi-turn agent turn-count; Ξ»=0.5 is the sweet spot β higher strengths (0.7 / 1.0 / 1.3) also got terse (~60 tokens) but fell <em>below</em> the base (0.39β0.49), so terseness alone is not the win; single-temperature, small per-category N.</p> |
| <p><strong>Q5_K_M confirm β uniform Q5 (3 seeds).</strong> Does a uniform higher precision (no selective promotion) clear "both"? <code>Q5_K_M</code> (20 G, imatrix-calibrated): build <code>0.867</code> (Β±0.133), long-context <code>0.988</code>, discipline <code>0.806</code> (Β±0.292) β build and discipline both below 0.90. Uniform-precision does not fix build and erodes discipline. Ruled out.</p> |
| <p><strong>Agentic-loop β autonomous convergence (40 held-out mini-projects).</strong> Each quant ran 40 held-out mini-projects (a README plus a <em>failing</em> pytest suite) fully autonomously: reason β read β implement β run tests β fix β converge. Convergence is <strong>objective pytest pass, not self-claimed.</strong></p> |
| <table> |
| <thead> |
| <tr> |
| <th>quant</th> |
| <th>n</th> |
| <th>convergence</th> |
| <th>mean turns (converged)</th> |
| <th>recovery (mean)</th> |
| <th>stall</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr> |
| <td>c76</td> |
| <td>40</td> |
| <td><code>100%</code></td> |
| <td>6.6</td> |
| <td>0.4</td> |
| <td><code>0%</code></td> |
| </tr> |
| <tr> |
| <td>Q8_0</td> |
| <td>40</td> |
| <td><code>100%</code></td> |
| <td>7.0</td> |
| <td>0.5</td> |
| <td><code>0%</code></td> |
| </tr> |
| <tr> |
| <td>IQ4_NL</td> |
| <td>40</td> |
| <td><code>100%</code></td> |
| <td>7.2</td> |
| <td>0.4</td> |
| <td><code>0%</code></td> |
| </tr> |
| </tbody> |
| </table> |
| <p>This axis <strong>did not discriminate</strong> β every quant (including the un-adapted base behaviour) converged on all 40 tasks, so autonomous convergence is a property of the model <em>family</em> on these tasks, not a PatchCode distinction. It does not favour any ship candidate, and the decision falls to size + recipe methodology. (Per-task: 8 tasks Γ 5 reps each, all 5/5 for every quant β <code>calc</code>, <code>debug_stack</code>, <code>graph</code>, <code>lru</code>, <code>mdlist</code>, <code>minijson</code>, <code>taskq</code>, <code>tracker</code>.)</p> |
| <p><strong>SignalLatch gate suite β IQ4_NL vs BF16.</strong> The established four-type gate set (coding/habits, hard-reasoning, hard-project, long-context) run on the PatchCode merge in both formats. Both clear every gate with <strong>zero errors</strong>; IQ4_NL tracks or nominally edges BF16. The ~0.04 gaps sit inside the build noise floor, so this reads as <em>tied</em>, not an IQ4_NL win.</p> |
| <table> |
| <thead> |
| <tr> |
| <th>gate (cases)</th> |
| <th>PatchCode IQ4_NL</th> |
| <th>BF16 (control)</th> |
| </tr> |
| </thead> |
| <tbody> |
| <tr> |
| <td>coding / habits</td> |
| <td><code>0.958</code></td> |
| <td><code>0.917</code></td> |
| </tr> |
| <tr> |
| <td>hard-reasoning</td> |
| <td><code>0.789</code></td> |
| <td><code>0.751</code></td> |
| </tr> |
| <tr> |
| <td>long-context (4)</td> |
| <td><code>0.979</code></td> |
| <td><code>0.941</code></td> |
| </tr> |
| <tr> |
| <td><strong>weighted overall</strong></td> |
| <td><strong><code>0.887</code></strong></td> |
| <td><code>0.846</code></td> |
| </tr> |
| </tbody> |
| </table> |
| <h2>The noise lesson (critical β reuse for every future bake-off)</h2> |
| <p>The SignalLatch-style suite is <strong>noisier than it looked</strong>: |
| - KritaLite build: Β±0.067β0.13 <strong>run-to-run</strong> variance (beyond seed). c76 scored 0.933 β 0.867 β 0.907 on the same gguf. |
| - discipline: Β±0.3 spread. |
| - build is <strong>ceiling-limited</strong> (max 0.933 = 14/15) β zero headroom to discriminate two good quants.</p> |
| <p><strong>Rule:</strong> 3-seed differences <0.13 on this suite are meaningless. Use <strong>5+ seeds, same-condition head-to-head</strong> before any ship call. Only non-noise axes (size, recipe methodology/safety, long-context at ceiling) reliably tiebreak. HumanEval was rejected β it saturates on Qwen and is the wrong mode for an agent.</p> |
| <p>This is exactly how a 3-seed pass almost shipped the <em>weaker</em> model.</p> |
| <h2>The ship decision</h2> |
| <p>With build, discipline, and long-context all <strong>tied within noise</strong>, the decision fell to non-noise axes, where plain <code>IQ4_NL</code> wins all three:</p> |
| <p><img alt="No candidate clears BOTH build and discipline (β₯0.90) β promotion destroys discipline; precision does not fix build." src="https://noonr48.github.io/qwen36-aeon-ik-llama/patchcode-testing-process/assets/bothquest.png" /></p> |
| <p><img alt="Ship scoreboard (5-seed): IQ4_NL ties the field within noise on build/long-context/discipline, and wins on size." src="https://noonr48.github.io/qwen36-aeon-ik-llama/patchcode-testing-process/assets/ship_scoreboard.png" /></p> |
| <ul> |
| <li><strong>smaller</strong> (16.6 G vs 20β29 G)</li> |
| <li><strong>marginal long-context</strong> edge (0.975 vs 0.935β0.969)</li> |
| <li><strong>plain-quant recipe</strong> β the fleet's proven pattern; promotion/mixed recipes carry evidence-harmful risk (discipline collapse) for zero measured benefit</li> |
| </ul> |
| <p>Ship: <strong>plain <code>IQ4_NL</code> (reasoning-imatrix)</strong>. The mixed-recipe <code>c76</code> is retained on disk as the build-heavy fallback if a future, harder build-gate ever discriminates beyond the noise floor (use 5+ seeds).</p> |
| <h2>What the testing says and does not say</h2> |
| <p><strong>Does say:</strong> |
| - PatchCode's distilled action-first discipline is preserved through <code>IQ4_NL</code> (tied with BF16 across build / long-context / discipline). |
| - Near-lossless precision (Q8) and attention promotion buy no measurable edge on this suite. |
| - Plain <code>IQ4_NL</code> is the defensible default on size + recipe safety.</p> |
| <p><strong>Does not say:</strong> |
| - It does not prove PatchCode is better for all tasks. |
| - It does not prove plain <code>IQ4_NL</code> is globally optimal. |
| - It does not make this a stock <code>llama.cpp</code> / <code>vLLM</code> release. |
| - It does not make live LoRA loading the recommended serving setup.</p> |
| <p>The most accurate public sentence:</p> |
| <blockquote> |
| <p>On a 5-seed, same-condition practical coding-agent bake-off, PatchCode plain <code>IQ4_NL</code> tied BF16 within noise on build, long-context, and discipline, and was the selected default on size and recipe safety.</p> |
| </blockquote> |
| <h2>Selected artifact</h2> |
| <pre><code class="language-text">Qwen3.6-27B-AEON-RYS-Agentic-Coder-PatchCode.IQ4_NL.gguf (16.6 GB β recommended) |
| Qwen3.6-27B-AEON-RYS-Agentic-Coder-PatchCode.BF16.gguf (57.6 GB β source-quality reference) |
| </code></pre> |
| <p>Recommended runtime: <code>https://github.com/noonr48/qwen36-aeon-ik-llama</code></p> |
| <pre><code class="language-bash">./build/bin/llama-server \ |
| -m /path/to/Qwen3.6-27B-AEON-RYS-Agentic-Coder-PatchCode.IQ4_NL.gguf \ |
| -c 65536 -ngl 999 -np 1 -fa on -sm none \ |
| --temp 0.7 --jinja --reasoning-format deepseek --reasoning-budget 0 |
| </code></pre> |
| <p>(<code><think></code> is emitted as a separate <code>reasoning_content</code> field β use <code>--reasoning-format deepseek</code> or fold it back so tool-action parsing sees the action.)</p> |
| <h2>Final read</h2> |
| <p>This was not a clean leaderboard. It was a real engineering pass: distil the style, build a hardened discriminator because the easy gates saturated, get fooled by a one-run perfect build score, repeat the finalists same-condition, discover the build is ceiling-limited and noisy, and ship the smallest plain-quant that ties everything within noise.</p> |
| <pre><code class="language-text">PatchCode IQ4_NL is a practical agentic-coder upgrade over the SignalLatch release. |
| It is the selected default among the tested quants, tied with BF16 within noise β |
| not a universal final answer. |
| </code></pre></div></main> |
|
|
| <footer> |
| <div class="wrap"> |
| Last updated: 2026-06-29. This page documents the PatchCode (agentic-coder) fine-tune and quant bake-off for the Qwen3.6 AEON RYS / SignalLatch line.</div> |
| </footer> |
| </body> |
| </html> |
|
|
