[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77VEXQCTFBWNXC4YGMD02D4":3,"skill-same:01M77VEXQCTFBWNXC4YGMD02D4":45},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":26,"repositoryHtmlUrl":27,"repositoryHomepageUrl":28,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M77VEXQCTFBWNXC4YGMD02D4","loss-backprop-lens",null,"Use when deciding whether to edit code at all, how large the edit should be, or whether a \"working\" fix actually generalizes. Applies when iterating on a failing test\u002Frun, chasing a flake, committing a sequence of small patches, or tempted to ship on one data point. Frames code changes as gradient steps and forbids overfitting to a single sample or a single test.","CorvinLabs\u002FCorvinOS","CorvinLabs","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F289506478?v=4","01M77VCHYED8C391HPBTNYP0HS","01M77VCHZ5K92DH3W6SPN8Y4P2",25,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fbundle\u002Fskills\u002Fldd\u002Floss-backprop-lens\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: loss-backprop-lens\ndescription: Use when deciding whether to edit code at all, how large the edit should be, or whether a \"working\" fix actually generalizes. Applies when iterating on a failing test\u002Frun, chasing a flake, committing a sequence of small patches, or tempted to ship on one data point. Frames code changes as gradient steps and forbids overfitting to a single sample or a single test.\n---\n\n# Loss-Backprop-Lens\n\n## Overview\n\nWorking on a codebase is **gradient descent on code**. A test \u002F CI run \u002F E2E is a **forward pass**. The difference between expected and actual is the **loss**. Every edit is a **step** toward lower loss. This is not a metaphor — it dictates which edits are admissible and which are noise injected into your parameters.\n\n**Core principle:** A fix that reduces training loss (the current failing test) but raises generalization loss (future unseen inputs or sibling tests) is **overfitting** and is rejected, regardless of whether CI turns green.\n\nThis skill is a *mental frame*, not a rigid procedure. Apply it when one of the triggers below fires; combine with `root-cause-by-layer` for the mechanics of diagnosis.\n\n## When to Use\n\n- You have **one failing signal** and are tempted to ship a fix\n- You've made **3+ small patches to the same area** in a short window\n- A test is **flaky** and you want to retry \u002F widen \u002F skip\n- A fix **works for this test** but you aren't sure about siblings\n- You're deciding between a **local tweak** and a **structural change**\n\nDo **not** use for: greenfield feature work, pure refactors with no failing signal, single-concept bug fixes where the contract is obvious.\n\n## The Mapping\n\n| ML concept | Code equivalent |\n|---|---|\n| Forward pass | Running the test \u002F job \u002F E2E |\n| **Loss** | Difference between expected and actual output |\n| Backprop | Tracing symptom → structural origin (use `root-cause-by-layer`) |\n| Parameters (θ) | The code: logic, prompts, rules, tests, configs |\n| Gradient | The causal story \"symptom → which parameter, in which direction\" |\n| Learning rate | Edit aggressiveness: local tweak vs. redrawing a boundary |\n| **Overfitting** | Symptom patch that works **here** but doesn't generalize |\n| Regularization | Contracts, layer integrity, invariants — what keeps edits honest |\n| Training set | Tests you've already seen |\n| **Test set** | Inputs \u002F tests you **haven't** seen |\n| Batch noise | LLM, network, filesystem non-determinism — one sample is noisy |\n| Local minimum | \"Green\" state that's fragile: all tests pass, small input change breaks it |\n\n## The Four Rules\n\n### 1. No update on a single sample\n\nOne failing run is a **noisy gradient**. It may be a bug; it may be LLM format drift, infra flake, race condition, cosmic ray. Before editing code:\n\n- Reproduce the failure. If it doesn't reproduce in ≥2 of the next N runs, you have noise, not signal — do not edit.\n- If it does reproduce: now you have a gradient, walk `root-cause-by-layer`.\n\n**Exception:** The single sample itself is unambiguous signal (a clear contract violation diagnosable from the log, e.g. `ValueError: unsupported operand type(s) for +: 'Decimal' and 'NoneType'` on a known-migrated column). Then the gradient is in the log, not in repetition.\n\n### 2. Gradient before edit\n\nYou do not have an edit until you have a causal story: **symptom → mechanism → contract → structural origin → conceptual origin** (see `root-cause-by-layer`). Without the gradient, any edit is a random direction in parameter space — on average, it raises generalization loss even when it happens to lower training loss.\n\n### 3. Learning rate must match the loss structure\n\nThe size of your edit must match the scope of the loss:\n\n| Loss pattern | Admissible edit size |\n|---|---|\n| One-off, surface-level symptom | Local tweak at the named layer |\n| Same defect recurring across tests \u002F iterations (≥3 in a short window) | **Architectural** edit — redraw the boundary, introduce a missing abstraction, even if it costs more lines today |\n| One \"structural\" bug affecting one call site | Fix at the boundary; do not escalate to architecture |\n\n**Red flag:** 3+ consecutive commits in the same function, each fixing a different small symptom. You are inside a local minimum doing gradient descent on the wrong objective. Stop, lift your head, consider an architectural edit or deleting the feature.\n\n### 4. Regularization beats local fix\n\nA fix that makes the current test green but violates a contract, a layer boundary, an invariant, or a published rule **raises generalization loss**. Rejected, even when the gate is satisfied.\n\nExamples of regularizers (things you do not trade for a green gate):\n- Explicit contracts between components\n- Layer boundaries (domain ≠ transport ≠ persistence)\n- Single source of truth\n- Invariants (documented, enforced, tested)\n- \"Explicit over implicit\"\n\nIf a proposed fix violates any of these, it is an overfit even if it passes locally.\n\n## Red Flags — STOP, you are overfitting\n\n- \"One failing run, let me patch it now\" → single sample\n- \"I'll add a retry to handle the flakiness\" → training on noise\n- \"It's a 3-line change, safer than a refactor\" → ignoring the loss pattern\n- \"Each fix is small but the bugs keep coming\" → local-minimum trap, step size too small\n- \"This fix passes THE test\" → check the *unseen* test\n- \"The test is wrong, let me adjust the assertion\" → fitting test to code instead of code to spec\n\n## How to Apply — checklist\n\n1. **Is the signal real?** Reproduce or find the unambiguous fingerprint in the log. If neither, do not edit.\n2. **What's the pattern?** One-off or recurring? Look at the last 3 commits in the touched area.\n3. **Pick the right step size:** local for one-off at a clear boundary; architectural for recurring or for a repeated local-minimum dance.\n4. **Check generalization:** \"If a sibling input \u002F sibling test were in the suite, would my fix still be correct?\"\n5. **Regularize:** Does the fix respect contracts, boundaries, invariants? If not, it's overfit — reject even if green.\n\n## Common Rationalizations\n\n| Excuse | Reality |\n|---|---|\n| \"One sample is enough, I can see the bug\" | Sometimes — only when the log itself is unambiguous signal. \"I can see it\" is not the same as \"I checked reproducibility.\" |\n| \"Small diff is safer\" | A small diff in a local-minimum loop is 5 small diffs you haven't written yet. The architectural edit has a smaller total diff. |\n| \"This is just a quick fix, not architecture\" | Five consecutive quick fixes IS architecture — the kind nobody designed. |\n| \"The retry handles the flake\" | You just trained on noise. The real gradient is still there. |\n| \"I'll step back after I ship this one\" | You won't. You're in gradient descent with a miscalibrated learning rate; each step lowers the bar for the next. |\n\n## Related\n\n- **`root-cause-by-layer`** — mechanics of computing the gradient (the 5-layer ladder).\n- **`docs-as-definition-of-done`** — regularization via documentation: a fix is not done until the doc-level mental model is current.\n- **`dialectical-reasoning`** — used at step 4 to force a counter-case on the proposed fix before shipping.\n\n## Real-World Cue\n\nIf your recent commit log in one function reads:\n```\nfix: strip whitespace\nfix: handle empty string\nfix: lowercase before compare\nfix: null-check result\nfix: strip quotes     ← you are here\n```\n…you are not fixing bugs. You are climbing an ungoverned gradient. Stop, step back, write the missing abstraction (a normalizer pipeline, a contract, a typed input), then delete the five patches.\n","# Loss-Backprop-Lens\n\n## Overview\n\nWorking on a codebase is **gradient descent on code**. A test \u002F CI run \u002F E2E is a **forward pass**. The difference between expected and actual is the **loss**. Every edit is a **step** toward lower loss. This is not a metaphor — it dictates which edits are admissible and which are noise injected into your parameters.\n\n**Core principle:** A fix that reduces training loss (the current failing test) but raises generalization loss (future unseen inputs or sibling tests) is **overfitting** and is rejected, regardless of whether CI turns green.\n\nThis skill is a *mental frame*, not a rigid procedure. Apply it when one of the triggers below fires; combine with `root-cause-by-layer` for the mechanics of diagnosis.\n\n## When to Use\n\n- You have **one failing signal** and are tempted to ship a fix\n- You've made **3+ small patches to the same area** in a short window\n- A test is **flaky** and you want to retry \u002F widen \u002F skip\n- A fix **works for this test** but you aren't sure about siblings\n- You're deciding between a **local tweak** and a **structural change**\n\nDo **not** use for: greenfield feature work, pure refactors with no failing signal, single-concept bug fixes where the contract is obvious.\n\n## The Mapping\n\n| ML concept | Code equivalent |\n|---|---|\n| Forward pass | Running the test \u002F job \u002F E2E |\n| **Loss** | Difference between expected and actual output |\n| Backprop | Tracing symptom → structural origin (use `root-cause-by-layer`) |\n| Parameters (θ) | The code: logic, prompts, rules, tests, configs |\n| Gradient | The causal story \"symptom → which parameter, in which direction\" |\n| Learning rate | Edit aggressiveness: local tweak vs. redrawing a boundary |\n| **Overfitting** | Symptom patch that works **here** but doesn't generalize |\n| Regularization | Contracts, layer integrity, invariants — what keeps edits honest |\n| Training set | Tests you've already seen |\n| **Test set** | Inputs \u002F tests you **haven't** seen |\n| Batch noise | LLM, network, filesystem non-determinism — one sample is noisy |\n| Local minimum | \"Green\" state that's fragile: all tests pass, small input change breaks it |\n\n## The Four Rules\n\n### 1. No update on a single sample\n\nOne failing run is a **noisy gradient**. It may be a bug; it may be LLM format drift, infra flake, race condition, cosmic ray. Before editing code:\n\n- Reproduce the failure. If it doesn't reproduce in ≥2 of the next N runs, you have noise, not signal — do not edit.\n- If it does reproduce: now you have a gradient, walk `root-cause-by-layer`.\n\n**Exception:** The single sample itself is unambiguous signal (a clear contract violation diagnosable from the log, e.g. `ValueError: unsupported operand type(s) for +: 'Decimal' and 'NoneType'` on a known-migrated column). Then the gradient is in the log, not in repetition.\n\n### 2. Gradient before edit\n\nYou do not have an edit until you have a causal story: **symptom → mechanism → contract → structural origin → conceptual origin** (see `root-cause-by-layer`). Without the gradient, any edit is a random direction in parameter space — on average, it raises generalization loss even when it happens to lower training loss.\n\n### 3. Learning rate must match the loss structure\n\nThe size of your edit must match the scope of the loss:\n\n| Loss pattern | Admissible edit size |\n|---|---|\n| One-off, surface-level symptom | Local tweak at the named layer |\n| Same defect recurring across tests \u002F iterations (≥3 in a short window) | **Architectural** edit — redraw the boundary, introduce a missing abstraction, even if it costs more lines today |\n| One \"structural\" bug affecting one call site | Fix at the boundary; do not escalate to architecture |\n\n**Red flag:** 3+ consecutive commits in the same function, each fixing a different small symptom. You are inside a local minimum doing gradient descent on the wrong objective. Stop, lift your head, consider an architectural edit or deleting the feature.\n\n### 4. Regularization beats local fix\n\nA fix that makes the current test green but violates a contract, a layer boundary, an invariant, or a published rule **raises generalization loss**. Rejected, even when the gate is satisfied.\n\nExamples of regularizers (things you do not trade for a green gate):\n- Explicit contracts between components\n- Layer boundaries (domain ≠ transport ≠ persistence)\n- Single source of truth\n- Invariants (documented, enforced, tested)\n- \"Explicit over implicit\"\n\nIf a proposed fix violates any of these, it is an overfit even if it passes locally.\n\n## Red Flags — STOP, you are overfitting\n\n- \"One failing run, let me patch it now\" → single sample\n- \"I'll add a retry to handle the flakiness\" → training on noise\n- \"It's a 3-line change, safer than a refactor\" → ignoring the loss pattern\n- \"Each fix is small but the bugs keep coming\" → local-minimum trap, step size too small\n- \"This fix passes THE test\" → check the *unseen* test\n- \"The test is wrong, let me adjust the assertion\" → fitting test to code instead of code to spec\n\n## How to Apply — checklist\n\n1. **Is the signal real?** Reproduce or find the unambiguous fingerprint in the log. If neither, do not edit.\n2. **What's the pattern?** One-off or recurring? Look at the last 3 commits in the touched area.\n3. **Pick the right step size:** local for one-off at a clear boundary; architectural for recurring or for a repeated local-minimum dance.\n4. **Check generalization:** \"If a sibling input \u002F sibling test were in the suite, would my fix still be correct?\"\n5. **Regularize:** Does the fix respect contracts, boundaries, invariants? If not, it's overfit — reject even if green.\n\n## Common Rationalizations\n\n| Excuse | Reality |\n|---|---|\n| \"One sample is enough, I can see the bug\" | Sometimes — only when the log itself is unambiguous signal. \"I can see it\" is not the same as \"I checked reproducibility.\" |\n| \"Small diff is safer\" | A small diff in a local-minimum loop is 5 small diffs you haven't written yet. The architectural edit has a smaller total diff. |\n| \"This is just a quick fix, not architecture\" | Five consecutive quick fixes IS architecture — the kind nobody designed. |\n| \"The retry handles the flake\" | You just trained on noise. The real gradient is still there. |\n| \"I'll step back after I ship this one\" | You won't. You're in gradient descent with a miscalibrated learning rate; each step lowers the bar for the next. |\n\n## Related\n\n- **`root-cause-by-layer`** — mechanics of computing the gradient (the 5-layer ladder).\n- **`docs-as-definition-of-done`** — regularization via documentation: a fix is not done until the doc-level mental model is current.\n- **`dialectical-reasoning`** — used at step 4 to force a counter-case on the proposed fix before shipping.\n\n## Real-World Cue\n\nIf your recent commit log in one function reads:\n```\nfix: strip whitespace\nfix: handle empty string\nfix: lowercase before compare\nfix: null-check result\nfix: strip quotes     ← you are here\n```\n…you are not fixing bugs. You are climbing an ungoverned gradient. Stop, step back, write the missing abstraction (a normalizer pipeline, a contract, a typed input), then delete the five patches.\n",{"name":5,"description":7},"main","3f28cfd9b82326ff325e645ff041589c3220d813","corvin_operator\u002Fbundle\u002Fskills\u002Fldd\u002Floss-backprop-lens",4,"https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS","https:\u002F\u002Fcorvin-labs.com, https:\u002F\u002Fwww.youtube.com\u002F@Corvin-Labs","https:\u002F\u002Fgithub.com\u002FCorvinLabs","2026-11-14 23:30:13",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add CorvinLabs\u002FCorvinOS --skill loss-backprop-lens",{"tool":37,"command":38},"bunx","bunx skills add CorvinLabs\u002FCorvinOS --skill loss-backprop-lens",{"tool":40,"command":41},"pnpm","pnpm dlx skills add CorvinLabs\u002FCorvinOS --skill loss-backprop-lens","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fbundle\u002Fskills\u002Fldd\u002Floss-backprop-lens\u002FSKILL.md",true,[],[46,53,59,65,72,78],{"skillNo":47,"name":48,"nameZh":6,"description":49,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":50,"categories":51,"keywords":52,"publishedAt":18},"01M77VGAEV6EQA3C62F78EDHQ3","voice","Voice mode for Claude Code — automatically reads assistant replies aloud via TTS, with German\u002FEnglish auto-detection and optional summarization for long responses. Trigger when the user mentions voice, vorlesen, sprachausgabe, TTS, \"speak\", \"say it out loud\", or asks about \u002Fvoice-* commands.","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fvoice\u002Fskills\u002Fvoice\u002FSKILL.md",[],[],{"skillNo":54,"name":55,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":18},"01M77VG6N69YBC0CR0TSMGXHTV","paper-generator-orchestrator","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fskills\u002Fpaper-generator-orchestrator\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":62,"categories":63,"keywords":64,"publishedAt":18},"01M77VG3DTRSZ1C4N57KNFARD5","academic-paper-generation","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fskills\u002Facademic-paper-generation\u002FSKILL.md",[],[],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":18},"01M77VFXA58C38HG5K5J4Q67MF","forge","Forge schema-bound, sandboxed tools at runtime. Paired with the cowork `forge` persona for chat-driven use in the voice-skill bridge. Use when you would otherwise write the same Bash snippet ≥3× with different parameters, when you need precise numerical results over a dataset, or when the user wants reproducible\u002Fdeterministic output written to a file.","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fforge\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":18},"01M77VFTJ8KS80R6PVPTBH4CWV","cowork","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fcowork\u002Fskills\u002Fcowork\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":81,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":18},"01M77VFQTC6WR708YMVHAKQWWN","quality-layer-control","Toggle quality disciplines (ADR Gate, docs-as-definition-of-done, etc.) on or off globally","https:\u002F\u002Fgithub.com\u002FCorvinLabs\u002FCorvinOS\u002Fblob\u002Fmain\u002Fcorvin_operator\u002Fbundle\u002Fskills\u002Fquality-layer-control\u002FSKILL.md",[],[]]