katopz/katgpt-rs / .agents/skills/research
research
VerifiedResearch workflow for distilling ML/AI papers into modelless inference primitives, freeze/thaw runtime patterns, latent-space operations, AND model-based training plans across the multi-repo stack. Use when reading arxiv papers, deciding which repo a paper belongs in, creating .research/ notes or .plans/ files, implementing modelless inference primitives, or routing training-vs-inference insights. Enforces the commercial strategy (public engine / private runtime / private chain / private neuron-db / private training / private SDK facade / private product-domain / public inference-substrate riir-infer / public decision-serving riir-reflex), three-track system (modelless inference + self-adaptive runtime + model-based trained weights — all three exist across the stack, not just riir-train), latent-to-latent preference, and freeze/thaw-over-fine-tuning rule. Every task ends with a mandatory Claude verdict ping-pong (§5) before any file is committed.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 80.0 KB |
| latent_raw.md | markdown | 1.6 KB |
| templates.md | markdown | 4.2 KB |
| vocab.md | markdown | 31.9 KB |
templates.md
View on GitHub# Format Templates — Supporting Reference for the Research Skill
> **When to read this file:** when creating a new `.research/NNN_*.md` note or `.plans/NNN_*.md` plan. Canonical examples to follow, not copy verbatim — adapt the sections to your paper.
## Research note format
Canonical example: `katgpt-rs/.research/238_LoRA_Muon_Spectral_Low_Rank_Manifold.md`
```markdown
# Research NNN: {Title}
> **Source:** <paper title + arxiv link + authors + date>
> **Date:** YYYY-MM-DD
> **Status:** Active | Done | Shelved
> **Related Research:** NNN (short note), ...
> **Related Plans:** NNN (short note), ...
> **Cross-ref (riir-ai / riir-chain / riir-neuron-db):** Research NNN, Plan NNN ← only if cross-repo
> **Classification:** Public | Private ← katgpt-rs notes are always Public
---
## TL;DR
<2-4 sentences: the distilled primitive, why it matters here, what it unblocks>
**Distilled for katgpt-rs (modelless, inference-time):**
<the transferable insight, stripped of training setup>
---
## 1. Paper Core Findings
...
## 2. Distillation
...
## 3. Verdict
**Tiers (high → low):**
| Tier | Criteria | Routing |
|------|----------|--------|
| **Super-GOAT** | Novel mechanism (no prior art) + new capability class + product selling point + force multiplier (≥2 pillars). Creates a moat. | Open primitive → katgpt-rs. **Architectural guide → riir-ai/.research/ OR riir-chain/.research/ OR riir-neuron-db/.research/**. Plans → appropriate repo(s). |
| **GOAT** | Provable gain (latency/quality/security) over existing approach, but not a new class of capability. Promotes to default if it wins. | Plan + implement. Feature flag + benchmark. |
| **Gain** | Incremental improvement, useful but not headline-worthy. | Plan only, behind feature flag. |
| **Pass** | Mechanism already ships OR not relevant OR training-only AND genuinely out of scope. | **No new files.** Verdict + one-line reason + closest shipped cousin, in conversation only. |
**One-line reasoning required for each verdict.** For Super-GOAT: state the selling point explicitly.
**After the tier verdict, run the MOAT gate per domain (main skill §1.6)** — a tier verdict without a domain-fit check can land a great primitive in the wrong repo and dilute the moat.
```
**File naming:** `{NNN}_{Short_Title_with_Underscores}.md` where NNN is the next free number (zero-padded to 3 digits, e.g. `239_`, `240_`). Check the folder first — numbers may be non-contiguous; pick the next free slot. **Numbers are monotonic and never reused** — even after a file is removed per the noise-reduction rule. Check `.highwater` if it exists.
## Plan format
Canonical example: `katgpt-rs/.plans/271_attention_matching_compaction.md`
```markdown
# Plan NNN: {Title}
**Date:** YYYY-MM-DD
**Research:** [katgpt-rs/.research/NNN_*.md](../.research/NNN_*.md)
**Source paper:** [arxiv ID.NNN](https://arxiv.org/abs/ID) — <short cite>
**Target:** `katgpt-rs/src/{module}/` (new module) + Cargo feature `{feature_name}`
**Status:** Active — Phase N {state}
---
## Goal
<one paragraph: what ships, what it enables, GOAT gate>
## Phase 1 — Unblocking Skeleton (CORE)
### Tasks
- [ ] **T1.1** <concrete, verifiable task>
- [ ] **T1.2** ...
```
Use `## Phase N` sections with `- [ ]` per task (mark `- [x]` when done, `- [-]` for deferred).
**GOAT gate rule** (AGENTS.md): every plan that introduces a new technique must have a feature flag and a benchmark proving the gain before promoting to default. Demote the loser if the new technique wins.
**UQ-bearing primitive GOAT gate extension (the "Report the Floor" rule, Research 322):** Any primitive claiming a probability distribution, predictive interval, quantile, coverage guarantee, or calibrated uncertainty MUST benchmark against the conformal-naive floor (`ConformalIntervalCalibrator<SeasonalNaiveForecaster>`, Plan 340 with `m=1`). If it cannot beat the floor on CRPS / empirical coverage / Winkler score, the GOAT gate FAILS. Existing UQ primitives (BoMSampler Plan 281, Sleep-Time Plan 334, Best-Belief Plan 336, KARC+overlay) are grandfathered but must include the floor at their next re-gate. The floor shipped in Plan 340 Phase 1 (2026-06-30); the rule is now enforceable. See `.benchmarks/010_report_the_floor_consolidated.md`.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |