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 |
vocab.md
View on GitHub# Vocabulary Translation Tables — Supporting Reference for the Research Skill
> **When to read this file:** the research skill §1 step 2 (vocabulary translation before grepping) requires these tables. Read the section(s) relevant to your paper's domain. Papers and our codebase use different words for the same mechanism — these tables are the translation layer that prevents false-novelty claims from paper-vocabulary-only greps.
## 1. Standing latent-state vocabulary (ALWAYS include, even for non-latent papers)
Most architecture/training papers have a latent-space reframing. Always grep BOTH the paper terms AND these codebase equivalents:
| Paper vocabulary | Codebase vocabulary |
|---|---|
| "residual stream" / "hidden state" / "activation" | "belief state", "latent subspace", "sense projection" |
| "layer" / "depth" / "stage" | "decision stage", "functor application", "cgsp cycle", "consolidation tick" |
| "width" / "dimension" / "capacity" | "latent subspace", "active projection channel", "sense channel" |
| "carry-forward" / "bypass" / "skip" | "leaky integrator", "dormant subspace", "decay gate", "persistence" |
| "collapse" / "degeneration" / "valley" | "coherence decay", "re-estimation trigger", "staleness" |
| "bottleneck" / "narrowing" | "subspace projection", "channel selection", "zone gating" |
| "fixed-point" / "deterministic" / "committed" | "LatCal", "lattice calculus", "BLAKE3 commitment", "raw scalar bridge" |
| "divergence" / "flux" / "∇·F" / "density change" | "codifferential", "δ", "DEC divergence", "belief_mass_divergence" |
| "curl" / "vorticity" / "∇×F" / "circulation" | "d₁", "DEC curl", "exterior_derivative rank 1→2" |
| "boundary" / "∂M" / "frontier" / "perimeter" | "exterior_derivative", "d", "coboundary operator", "boundary_flux_mass" |
| "line integral" / "trajectory energy" / "path cost" | "line_integral", "edge field sum", "rank-1 cochain path sum" |
| "Stokes theorem" / "divergence theorem" / "Green's/Gauss" | "DEC identity d∘d=0", "curl(grad)=0", "div(curl)=0", "hodge_decompose" |
| "Hodge decomposition" / "exact/coexact/harmonic" / "Helmholtz" | "hodge_decompose", "DecFlowField", "exact_flow/coexact_flow/harmonic_flow" |
| "Fokker-Planck" / "continuity equation" / "mass conservation" | "belief_mass_divergence", "codifferential on belief cochain" |
| "cell complex" / "mesh" / "simplicial" / "cubical" | "CellComplex", "CochainField", "grid_2d" |
## 2. Standing per-NPC runtime / freeze-thaw / personality vocabulary
ALWAYS include when the paper touches per-entity state, memory, personality, swap, evaluator/judge/critic, or selective erasure/forgetting. The `riir-ai/.research/` corpus is SATURATED in this space — paper-vocabulary-only greps produce false novelty claims (see R320 failure in the main skill).
| Paper vocabulary | Codebase vocabulary |
|---|---|
| "personality swap" / "personality drift" / "character shift" | "committed personality", "freeze/thaw cadence", `ArchetypeBlendShard`, `KarcShard` |
| "selective forgetting" / "memory erasure on swap" | "non-interference branches", "branch-local", `BranchBank`, "orthogonal subspace projection" |
| "survives swap" / "replay-deterministic personality" | "sampling invariance", "quorum-verifiable", "bit-identical across nodes", "FAME Proposition 3" |
| "epoch boundary" / "checkpoint replacement" / "non-stationary utility" | "freeze/thaw cadence", "consolidation sleep-cycle", `tau_reest`, "coherence < tau" |
| "co-evolution" / "evaluator replacement" / "moving target" | "personality divergence", "direction vector drift", "emergent personality at crowd scale" |
| "evaluator" / "judge" / "critic" / "verifier" (per-entity) | "claim verifier", "CLR vote", "Salience Tri-Gate", `ConstraintPruner` |
| "frozen snapshot" / "frozen artifact" (generic) | NAME THE CONCRETE TYPE: `KarcShard`, `ArchetypeBlendShard`, `BranchBank` snapshot, `ZoneGeometryPod`, `MerkleFrozenEnvelope`, `SleepAnticipationShard` |
| "cache invalidation on swap" / "dependent records" | `DecCache::mark_face_destroyed`, `ZoneGeometryCache::invalidate`, `topology_version` bump |
## 3. Standing compute-unit translation (MANDATORY for agent/LLM papers — R368 lesson)
Papers increasingly use "LLM forward pass" / "LLM call" as the compute unit for a decision. Our codebase uses different compute units. ALWAYS translate the compute unit, not just the semantic name:
| Paper vocabulary | Codebase vocabulary |
|---|---|
| "LLM decides what to write/record" | `SpeculativeGenerator` (draft) + `ScreeningPruner` (relevance) + `ConstraintPruner` (validity) |
| "LLM decides what to retrieve/read" | `SpeculativeGenerator` + `ScreeningPruner` + AnyRAG escalation gate |
| "LLM judges/verifies/critiques a claim" | CLR vote + SalienceTriGate + Claim Rubric L1/L2/L3 |
| "LLM reviews trajectory + rewrites code/prompts" | Raven/δ-Mem consolidation + MAPE-K self-healing (architectural analog; quality parity needs PoC) |
| "meta-LLM generates novel semantic content" | **NO modelless analog** — genuine NO-GAIN if the value IS the generation |
## 4. Standing substrate-translation vocabulary (MANDATORY for hardware/accelerator/NMP/PIM/ASIC/system papers — R418 lesson)
Papers framed in hardware vocabulary describe techniques whose *implementation substrate* is hardware, but whose *value* is substrate-independent. We simulate hardware dequant/accelerator techniques in software SIMD (Research 110 Ciot). ALWAYS translate the substrate. Before PASS-ing on a hardware paper, grep BOTH sets:
| Paper vocabulary (hardware) | Codebase vocabulary (software SIMD) |
|---|---|
| DQB / near-memory processing unit / PE | SIMD dequant fn, `dequant_via_lut`, `simd_lut_dequant` |
| HBM / pseudo-channel / memory controller | L1 cache, register file, SIMD lane |
| channel-aware layout / interleaved | cache-line-aligned, `AlignedWeightMatrix`, struct-of-arrays |
| wire-mapping FP-to-FP / 2:1 mux type conversion | bit-cast, `f16::from_bits`, SIMD shuffle/permute |
| LUT-based INT-to-FP / sign extension | pre-computed lookup table, `[f32; 256]`, SIMD gather |
| sideband tag / 3-bit tag / mode select | `QuantFormat` enum, runtime dispatch, `match` on format |
| S/Z buffer / scaling-factor cache | L1-resident scale cache, `BlockQ4K { d, dmin, scales, qs }` |
| fused DQ-GEMM kernel | fused dequantize-matmul, `simd_dot_f16_f32`, dequant-in-register |
| kernel selection / batch threshold | breakeven routing, Plan 218, memory-bound vs compute-bound |
| tensor core / CUDA core / SM | CPU SIMD lane, FPU, NEON/AVX2 lane (the *role*, not hardware) |
| io_uring / DPDK / RDMA / zero-copy bypass | zero-allocation hot path, `_buf` pattern, `pool.rs` |
| TPU systolic array / PE mesh | blocked matmul, `simd_matmul_rows`, tiled GEMM |
## 5. Standing database-substrate translation (MANDATORY for database/systems/storage/agent-state papers — R300 lesson)
Papers framed in database-engine vocabulary describe **access patterns** whose *implementation substrate* is a database engine, but whose *value* is the access pattern. We ship a database substrate (`riir-neuron-db`: Pod + lock-free `ShardIndex` + Merkle + MAPE-K + Raven/δ-Mem + `ItemEmbedIndex` + vibe KG) WITHOUT a SQL/Cypher engine. **Conflating "we don't have a fused query planner" with "we don't have a database" is the R300 false-PASS root cause.**
| Paper vocabulary (database) | Codebase vocabulary (neuron-db) |
|---|---|
| experience graph / reward-bearing search tree / UCB node | `NeuronShard` + lineage, `TrialRecord`, `KarcShard`, `EpisodeBuffer` |
| materialized view / training-data extraction | `consolidation.rs` Raven/δ-Mem output, `ShardCompactor` cold-tier product |
| vector-seeded graph traversal / ANN + relational + graph | `ItemEmbedIndex::query` (latent cosine ANN) composed with `vibe.rs` KgTripleTemplate edges |
| AS-OF temporal query / bi-temporal / time travel | `MerkleFrozenEnvelope` version chain, `PersonalityLedger` ring, `AuditLog` backward replay |
| frontier selection / UCB / visit_count + reward | `BanditPruner`, `katgpt_core::manifold_bandit` (`LatentTaskTree` / `BayesianFilterArm`, Plan 370 — ungated module), `MCTS`, `GZeroPlayer` |
| governed view maintenance / retraction / invalidate derived | `MapeKLoop` self-heal cascade, `BackupMirror::integrity_check_3way` |
| change log / CDC / logical step number | `PersonalityLedger` ring, `AuditLog` entries, `TrialLog` hash chain |
| concurrent tree search atomicity / MCTS backprop race | `FunctorEntry` atomic Arc-swap, `LoRAWeightVersion` lock-free read |
| episode buffer / experience replay | `EpisodeBuffer` FIFO, Raven/δ-Mem input queue |
| unified query planner / SQL/Cypher/Vector fused | **NO direct analog** — individual access patterns map to latent ops + neuro-symbolic edges + raw commitment |
## 6. Standing coefficient / blend-shape vocabulary (MANDATORY when the paper's mechanism is a fitted/adaptive/mixture coefficient — Le Critique lesson, canonical failure #3, 2026-08-20)
Papers about adaptive blending, mixture coefficients, ensemble weighting, baseline interpolation, or "how much to trust estimator A vs B" ship in our codebase as **hand-tuned constants** with "Weight on the X axis" style doc comments — never under the words "blend"/"mixture"/"adaptive". And the shape ships where you don't expect it: the Le Critique miss was a blend-vocab grep scoped to katgpt-core while the hit sat in riir-clippy's `self_evolve.rs`.
**Grep set (run repo-wide across ALL repos, never scoped to the expected home):**
`W_[A-Z]|const [A-Z_]+: f32|_WEIGHT|weight: f32|alpha: f32|trust: f32|blend_factor`
| Paper vocabulary | Codebase vocabulary (shipped instances) |
|---|---|
| "adaptive mixture coefficient" / "fitted blend weight" / "interpolation ratio" | `W_EVO = 0.6 / W_RATE = 0.4` (riir-clippy `select_best_candidate` — the Le Critique direct hit, ρ hand-pinned at 0.4); `alpha: f32` (`DualLeoMixer`, fixed 0.3); `latent_weight` (retrieval fusion — measured INERT, T3.5); `RetrievalWeights.trust = 0.2` (EvidenceTier); `KvRoutingConfig::blend_factor` (fixed thresholds — EVPO-style) |
| "ensemble weighting" / "combine predictors by measured accuracy" | the same constants — if FIXED, the paper's contribution is usually making them data-fit (TETHER ρ*) |
| "realized outcome" / "return" / "reward stream" (what the coefficient is fit against) | `EvolveRecorder::record_outcome` (riir-clippy, fires on every applied fix), trajectory success counts, bench verdict streams — if the feed already fires, the adaptive version is an A/B away |
**Rule:** a paper whose core mechanism is FITTING a coefficient structurally matches every hand-pinned constant implementing the same convex combination. Coverage requires reading the **containing expression** (the §3.6 granularity rule in the main skill) — and checking whether the outcome stream it would be fit against already fires.
## 7. Standing inference-substrate vocabulary (MANDATORY for quant/kernel/loader/encoder papers — riir-infer)
Papers about quantization formats, kernel ports, checkpoint loaders, or encoder forwards describe mechanisms whose substrate home is now `riir-infer` (public, carved from riir-ai 2026-09-22). Grep BOTH sets before any "already ships" / routing verdict in this space:
| Paper vocabulary | Codebase vocabulary (riir-infer) |
|---|---|
| "quantization format" / "weight-only quant" / "GPTQ/AWQ/EXL3-class" | `quant/` module, the EXL3-trellis lane (CPU reference = oracle + fast arm + `riir-infer-gpu` kernels), GDN quant certification (`deltanet/`) |
| "dequant kernel" / "fused dequant-GEMM" | `simd/` dequant forwards, fused dequant-matmul, dequant-in-register |
| "checkpoint loader" / "GGUF/safetensors" / "tokenizer" | `gguf_loader.rs`, `safetensors_loader.rs`, `tokenizer.rs` (BPE vs SentencePiece selection) |
| "architecture port" / "model forward" | `transformer/`, `gemma_layer.rs`, `llama_layer.rs`, `ternary_layer.rs`, `rope.rs`, `wall.rs` |
| "rotation" / "Hadamard" / "incoherence processing" | the Bonsai2 rotation lane (`bonsai2_hadamard`) |
| "kernel port" / "backend port" / "Metal/CUDA/Triton port" | the laya lane discipline: flat `Vec<f32>` ops, `crates/riir-infer-laya` MSL kernels, `crates/riir-infer-gpu` (wgpu/CubeCL), `laya-riir-cuda` |
| "bit-identical port" / "parity gate" / "golden forward" | the G5 gate: top-1 ≥ 99.9% + p-drift ≤ 1e-3 vs frozen captures, per checkpoint, BEFORE any published number |
## 8. Standing decision-serving vocabulary (MANDATORY for calibration/abstention/selective-prediction/serving papers — riir-reflex)
Papers about calibrated decision serving, abstention, selective prediction, or comparison arenas have a concrete public home: `riir-reflex` (public 2026-09-23). Grep BOTH sets before any novelty or routing verdict in this space:
| Paper vocabulary | Codebase vocabulary (riir-reflex) |
|---|---|
| "selective prediction" / "abstention" / "reject option" | the `noul` answer, fused abstain (calibration score + `CorpusDistanceGate`), abstention first-class in `decision_wire` |
| "confidence calibration" / "temperature scaling" / "conformal" | `SigmoidGateCalibrator`, sigmoid normalization (never softmax), per-suite fitted birth thresholds (harness cal-slice percentile) |
| "corpus as model" / "retrieval-augmented zero-shot" / "non-parametric classifier" | `Lz4FlexDrafter` corpus-is-the-model option scoring, `pick_domain` corpus routing over hashed-feature embedding (`embed.rs`) |
| "decision API" / "typed decision endpoint" | `decision_wire` (`choice`/`score`/`noul`), stdio line-JSON + HTTP edge (`serve.rs`), `game_heads.rs` fitted serving heads |
| "comparison arena" / "benchmark harness" / "leaderboard" | `harness/` (9 dataset suites, corpus-cap + `--cal-select-cap` levers), `lanes/` (laya CPU/Metal/CUDA, CLM, GLiNER, AgentJev) |
| "distribution shift probe" / "OOD rig" | pair-head A/B probes, held-out rig policies, G5 parity gating every published laya number |
## 9. Unified decision rule — substrate-as-instantiation vs mechanism-as-value
This rule prevents false-PASS/false-redirect across four paper classes (R418 hardware + R368 LLM + R300 database + Flow Sampling training-math):
- **Value = technique / access-pattern / decision-structure / math** (not the substrate) → substrate is one *instantiation* → translate to our substrate → GOAT/Gain candidate.
- **R418 hardware:** LUT lookup / shared ALU / co-located layout / dispatch tag (StreamDQ — DQB-in-HBM is instantiation, SIMD dequant is ours; shipped 2.3×).
- **R368 LLM:** decision structure (AutoMem LOG/PLAN — LLM is instantiation, probe/draft/pruner is ours). **When you see "N LLM calls/step", FIRST ask "what decision is each call computing?"** — not "violates 20Hz budget, NO-GAIN". R169 guard applies ONLY to genuine LLM-dependent process.
- **R300 database:** access pattern (Trellis — Axiom/Velox is instantiation, `riir-neuron-db` Pod + Merkle + vibe KG is ours). `riir-neuron-db` IS a database; prefer latent-first + neuro-symbolic translations over `AnyRAG` external escalation.
- **Flow Sampling training-math:** closed-form drift / conditional score / Riemannian correction / regression target (training loop is instantiation, dllm + Latent Field Steering + freeze/thaw is ours).
- **Value = substrate-fabrication-advance itself** (new transistor geometry, new query optimizer algorithm, new optimizer like Muon, new loss function, new RL algorithm, semantic code generation) → no modelless analog → PASS or → riir-train.
## 10. Standing weight-space-programming vocabulary (MANDATORY for weight-edit / patch / writer-edits-weights / DSPy-optimizer papers)
**The in-workspace corpus is DENSE — grep it BEFORE any verdict (internal-first rule, main skill §1 step 4).** Canonical notes (katgpt-rs `.research/` unless noted): **648** Weight-Space Programming (the Takehi/Teoh/Zhang/Khattab "Sparks of WSP" distill — the parent note; the obstacle-vs-ability law) · **649** probe-authored circuit grammar (folded into `weight_edit_overlay::probe_circuit_grammar`) · **655** weight-space learning / model zoos · **659** wsp — closed-form MEMIT weight-write (Issue 990 `closed_form_weight_write`; PoC NEGATIVE at fixture scale, retained substrate) · **664** LP-authored weight edits (Issue 993, landed on the 980 lane) · riir-train Issue 665 (WSP weight-production lane). Landed lanes: katgpt-core **`weight_edit_overlay`** (Issue 980, Bench 976 GOAT G1–G4; gated by the opt-in `gain_cost_halt` feature — the arming condition IS the halter's obstacle telemetry) + **`closed_form_weight_write`** (Issue 990; gated by the opt-in `quant_error_lora` feature), with the riir-infer serving consumer **`closed_form_write`** (`WriteOverlays`, Issue 1189's overlay-at-dequant lane; implies `katgpt-core/quant_error_lora`). Research 659 records the canonical miss this table exists to prevent: *"Missed by this note's first-draft internal grep (keyword miss: 'weight-space' vocabulary) — caught by owner challenge"*. **cmpnd's `dspy-wsp` (2026-11-13) is the DECISION-MODEL instantiation of the same paper** — a writer LLM edits a decision student's weights via a hash-checked patch, GEPA population search, ReAnchor threshold-refit composition — and its decision backends serve `/v1/systemone`, the wire our laya `RiirAgent::system_one` + AgentJev/Jev-Mem lanes dial, benchmarking on Typed Decisions (our `typed_decisions` suite family). Distilled same-day -> Research 671 (the decision-model instantiation: the simulate_head closed-form law 5.75x, the two-axis ReAnchor composition, the writer-class swing; the untried head-surface cell filed as riir-infer Issue 1216). **676 AHA (2026-11-16) closes the lineage's last agent inputs** — matched-pair self-supervised direction (same-prefix resampled branches, detector-labeled) + LP-hit-rate layer search + the cut-off-rate headroom bound; Issue 1224 LANDED the third authoring arm same day (Bench 985 — parity bar met, claim construction-only per the falsifier non-firing; `matched_pair_authoring` + `arm_from_telemetry_with_outcomes`, bench 985). Grep anchors: `matched pair`, `cut-off rate|headroom`, `presence_penalty` (sampler purity), `AHA`. **679 WSP-Doesn't-Need-Transformers (2026-11-17, Ok & Lee — the architecture-choice successor)**: three writers free-design → ALL converge on sparse concept tables + fixed computed-address readers (PKM-shaped — Plan 408 covers the addressing); BM25-beats-written-transformers interference census (delete-one-row 0.5–1.8 vs 4.8+ flips); the BM25-over-author-text floor + LOO row-flip census instruments → riir-reflex Issue 121 (report-only). Grep anchors: `Doesn't Need Transformers`, `concept table`, `pair table`, `BM25 floor`, `row-flip census`, `Ok`. **riir-train Issue 696 (2026-11-17 — the Cohen & Shapira compress-backprop meta-model blog, the TRAINED-MAP quadrant)**: a small trained hypernet reads a task description and writes ΔW for a frozen base — measured evidence for the class (59/75 benchmarks over base+description; 16-row LoRA beaten, ≥64 rows loses), the three-rung signal ladder (adapter-replication → through-base → intrinsic-reward RL) recorded as `shine_d2l_train`'s named fix with the production run still deferred (Plan 302), modelless half audited-discarded with a served-arm re-open trigger (m3 addendum 2026-11-18: the ESC hosted think leg is a CANDIDATE trigger condition — Bench 088 is a wiring smoke, not a served-traffic read; the gold-labeled slot-residual + offline permutation probe in 696's addendum decides — read 696 before re-running this class). Grep anchors: `compress backpropagation`, `rung ladder`, `meta-update`, `Cohen`.
**Grep set (repo-wide, both layers):** `weight[-_ ]space|weight_edit|WeightEditOverlay|closed_form_weight_write|QuantErrorLora|LoraPair|WeightEpoch|check_epoch_tag|LaneSlot|MerkleFrozenEnvelope|factor_tail|forward_delta_into|system_one`
| Paper vocabulary | Codebase vocabulary |
|---|---|
| "weight-space programming" / "the writer edits the student's weights, no gradients, no training" | `weight_edit_overlay` (katgpt-core Issue 980: `WeightEditSpec` / `EditForm::{DirectionSubtraction, BypassRows}` / `ProvenanceTrailer` / `WeightEditArtifact`; `forward_delta_into` applies deltas in the DEQUANT domain — never requantizing a packed base; `compose_layer_into` / `revert_layer` the bit-identical dense path; arming = `gain_cost_halt` `Halt{Oscillation}` obstacle telemetry, refusing "no obstacle found") — the authored-overlay carrier; `closed_form_weight_write` (Issue 990, the `qwen0.8B-wsp` recipe family) the closed-form writer |
| "writer" (the frontier LLM authoring the edit) / "student model" | ⚠ NAME COLLISION: `LoraPair { reader, writer }` (katgpt-types `lora.rs`) — OUR "writer" is the causal-decode LoRA half, never an authoring LLM. The authoring role ships as the NOTES' authors themselves: LLM author (648), probe author (`probe_circuit_grammar`, 649), LP author (664, one PDHG solve). "student" ≈ the serving lane / specialist; distill teacher/student lanes live in riir-train + reflex `--distill` |
| "patch" (changed entries only, a few MB; `patch/manifest.json` + `edit.npz`) | the overlay artifact: `WeightEditArtifact` + `ProvenanceTrailer`; the rank-r patch forms: `QuantErrorLora::from_error` (dropped energy REPORTED via `thin_svd_into`) and `sst_certified::factor_tail`'s `TailFactor` payload `c = U_⊥ᵀ Δ V_⊥` (`compose_tail` same-anchor chaining; basis mismatch refuses loud) |
| "full checkpoint" (`write_checkpoint`) | freeze/thaw snapshot: `MerkleFrozenEnvelope::freeze` / `thaw_envelope` (riir-neuron-db), `WeightEpoch` versioning (katgpt-types `lora.rs`) |
| "apply checks sha256 of base + edited tensor — anyone with the public base rebuilds bit-for-bit" | digest-checked intake + mount: `fetch_artifacts.sh` (BLAKE3 + exact-size BEFORE use), `check_winner_file` (refuses a mismatched artifact), `graft_artifact.rs` (digest-pinned sidecar envelope + monotonic mount), envelope BLAKE3 commitments + `verify_descent` lineage |
| "swap edits in/out at serve time" / "reload in place, clear the prefix cache" | `LaneSlot` + `check_epoch_tag` monotonic apply — advance / idempotent / fork refused / downgrade refused (instinct `arsenal_ops.rs`, arsenal law A2) + the arsenal laws' A8 no-runtime-minting (instinct AGENTS.md — apply-side only); riir-engine `LoRAHotSwap` / `dispatch_lora_merge` |
| "guardrails refuse embeddings / output layer / final norm; refuse whole-tensor rescales" | the overlay law: the base artifact never moves (base digest pinned on the envelope; an empty overlay demotes byte-identically; structural undo = drop the row, rows-identical); sst `compose_tail` same-anchor / basis-mismatch loud refusals |
| "simulate_head — the writer recomputes the head from reads and derives the edit closed-form" | closed-form derivation doctrine (main skill §3.5 Path 2 — deterministically constructed overlays, never GD); shipped instances: the fact-write's protected ridge solve (`dW = R·Kpᵀ·M⁻¹`, f64 Cholesky, Issue-940 deterministic fold order) + Research 664's LP solve (`min ‖v‖₁` under activation constraints) |
| "forward-pass tool" / "teacher_forced reads (hidden states, MLP activations, logits)" | the `FactWriteModel` seam (teacher-forced capture of the MLP down-projection INPUT at answer positions); `probe_circuit_grammar` (activation probes as the read channel); laya `RiirAgent` reads |
| "GEPA population search over candidate edits (Pareto-frontier parent selection)" | `ruliology_search` L3 Pareto-front candidate search (riir-refine); Pareto rank-0 + Wilson bounds (reflex `stats.rs`); bandit evidence selection (`BanditPruner`, katgpt-pruners; `katgpt_core::manifold_bandit`'s `LatentTaskTree` / `BayesianFilterArm`, Plan 370); GEPA-D `gepa_reflective` (katgpt-pruners, Plan 164, default-on) for the config-evolution population |
| "paired gain, same seeds, ±SE" / paired non-inferiority | reflex harness paired LB95 + paired non-inferiority δ = max(1.0pp, 2.5·SE); instinct T2 strict-superiority paired gate (paired (pick − A0) LB95 > 0, ONE frozen read); bench paired A/B under common random numbers |
| "holdout the writer never sees / valset hidden until the final check" | the contamination discipline (`grammar-not-answers` skill: never ingest test rows in any proportion; echo-gate + leak-scan; the cal slice must not sit inside its own reference corpora) |
| "ReAnchor refits decision thresholds (Noul/Score/Choice), changes no weights, stacks with the edit" | the TWO ORTHOGONAL AXES law — serving-config fit vs weight edit compose: reflex per-suite cal-slice threshold fitting (`--cal-select-cap`, `--gate-fit-calibrated`, ESC gate ρ fits) rides the threshold axis; overlays ride the weights axis |
| "decision head / pointer head / joint schema head"; "System One decision model" (state + typed questions → per-option probabilities) | decision_wire (`choice`/`score`/`noul`) + `label_heads.rs` / `readout.rs` (reflex); `RiirAgent::system_one` (riir-infer-laya) — the same serving shape; `/v1/systemone` is the wire our AgentJev + Jev-Mem lanes dial (identity RESOLVED 2026-11-14, Research 671: our AgentJev lane dials malevrigns/agent-jev — third-party, not TypeSafe hosted Jev, not NanoJev/C-Tianyu; dspy-wsp supports none of the laya families — family yes, identity no) |
| "editing removes obstacles; it doesn't add abilities" | the modelless-correction doctrine (§3 last row): freeze-state corrections repair systematic biases; "NO-GAIN if the value IS the generation" — Research 648's obstacle-vs-ability law; an edited model fixes a failure mechanism, never grows a capability |
## 11. Standing persistent-latent retention vocabulary (MANDATORY for recurrent-state / gated-recurrence / memory / SSM papers — the Latent-Lattice lesson, canonical failure #6, 2026-11-19)
Papers with per-layer/per-entity persistent state updated across time (lattice latents, RNN companions, SSM state, belief updates, memory streams) ship here as **leaky/EMA integrators whose retention coefficient is a fixed config constant** — never under the paper's words. THE ASYMMETRY this table exists to catch (census 2026-11-19, riir-train research 500 §4): **drive c is input-derived EVERYWHERE; retention g was FIXED EVERYWHERE** (until Plan 639). A mechanism-name grep finds the sites; only the COEFFICIENT question — *is the decay fixed config or derived from the input?* — finds the delta. **Split every recurrence into (drive c, retention g, readout) and signal-diff EACH class separately** (main skill §3.6 at formula level; a component-level "recurrence: covered" verdict is void).
**Grep set (repo-wide, ALL repos):** `retention|_decay|decay_|leaky|carry|EMA|exponential_moving|momentum|gamma|latch|sticky|surprise|novelty|belief_retention|reach_decay|decay_confidence`
| Paper vocabulary | Codebase vocabulary (shipped instances) |
|---|---|
| "gated recurrence" `r_t = g⊙r_{t−1} + (1−g)⊙c` / "input-dependent decay" / "selective retention" (the Mamba-Δ class) | `evolve_belief_additive` (katgpt-sense `reconstruction.rs:986` — the EXACT shape, retention `config.belief_retention[m]` = `[0.9; 6]` FIXED); `leaky_step` (`katgpt-types/src/leaky_core.rs`); GDN's gated delta rule (riir-infer `deltanet/` — the paper-class recurrent state itself) |
| "EMA" / "leaky integration" / "exponential moving average" | refine `ema_step` (`evolve.rs:83`, outcome-signed α=0.1/β=0.05 FIXED); goal_salience `reach_decay = 0.995`/tick (FIXED); desperation_dual β=0.03 (FIXED); TemporalDerivativeKernel dual fast/slow EMA (0.3/0.03 FIXED + `sigmoid_surprise_gate`) |
| "decay over time" / "confidence fade" / "forgetting" | `decay_confidence` family (`σ(−λ·Δt + 4.6)`, riir-games-shared `game_traits/spatial.rs` — FIXED λ); **`block λ` per-axis affect decay** (`entity_cognition/block_lambda.rs` — measured persistence ρ → λ per axis, COMPILE-TIME PINNED: the static precursor of input-derived retention; glossary row "block λ"); CommandDrive decay-per-tick |
| "FiLM" / "conditioning" / "gain and shift from a latent" | AffectProjection `fast_sigmoid(gain·x + bias)` → 5 emotion scalars (riir-engine `mop_runtime/hla_bridge.rs`); ProjectionWeights per-module sigmoid gates (katgpt-sense); `functor_gate(coherence, β, τ)` (riir-engine `latent_functor/zone_gating.rs:68`); factorized_action FiLM (katgpt-core) |
| "persistent side state" / "companion memory" / "horizontal latent" | per-NPC `belief[8]` (katgpt-sense); HLA per-head accumulators (katgpt-hla); LatentSkillEvolution directions (riir-refine, durable traj_store learn rows); `CommittedFieldBlend` (archetype blend, feature-gated); `DeltaNetState` (riir-infer — the MIXING-side class); `DeltaMemoryState` + surprise-gated writes |
| "stateless vs recurrent conditioning" (the ablation axis) | static per-channel scales: DORY `attn_gate_scale`/`ffn_gate_scale` (riir-infer `dory/forward.rs:629`) vs the persistent-latent conditioners above — the paper's stateless-vs-recurrent ablation maps onto exactly this split |
| "evidence-gated update rate" / "adaptive step size" | `elo_k_eff = K·min(1, n/16)` (evidence gates WHEN-to-move, never retention); calibrator `min_obs` refit gating; sticky latches (pick_latch, EscRateGuard — the latch family: conditioning state with NO decay) |
⚠ **GLOSSARY FIRST (the M3 defense):** the fleet glossary (`riir-ai/.docs/01_orientation/glossary.md`) already coins this family — `drive`, `block λ (per-block decay)` ("instead of one global constant" — literally the fixed-vs-derived retention axis), `no_gain`'s leaky fill, `CommandDrive`'s decay-per-tick, `elo_k_eff`, `futility`'s latch. `read_file` it BEFORE inventing variants (main skill §1 step 2, canonical failure #6).
## 12. Worked examples
**DiPOD paper → riir-ai code:** paper-vocabulary grep misses `latent_functor/reestimation.rs` which ships DiPOD's "interleave self-distillation when ELBO drifts" as "coherence-driven re-estimation scheduler when coherence < tau_reest". Vocabulary translation is the only defense — notes framing can use codebase vocabulary that paper-vocabulary grep misses on BOTH layers.
**Stokes paper → katgpt-rs DEC code:** paper-vocabulary grep for `stokes|divergence theorem|fokker-planck` returns ZERO hits. Codebase-vocabulary grep for `codifferential|exterior_derivative|hodge_decompose|DecFlowField` hits `katgpt-dec/src/operators.rs`, `katgpt-dec/src/hodge.rs`, `katgpt-dec/src/flow.rs`. The Generalized Stokes' theorem machinery ships as DEC operators but no note framed it in Stokes vocabulary.
**Latent-Lattice paper → the modelless lattice (2026-11-19, canonical failure #6):** paper vocabulary ("lattice latents", "horizontal stream", "companion state") plus a mechanism-name grep found HLA / evolve_belief / zone_gating at the WRONG GRANULARITY ("pattern exists partially — validation only"). The owner challenge ("we actually do/done that") forced the coefficient question and flipped the verdict: `evolve_belief_additive` is the paper's EXACT gated-recurrence shape, and the one delta was the retention coefficient (fixed `[0.9; 6]` config vs the paper's input-derived g) — a plan-able modelless fusion (katgpt-rs Plan 639, novelty-gated retention). The fleet glossary's `block λ` row had documented the fixed-vs-per-axis retention axis one hop away the whole time. Products: riir-train research 500 §4 (the census), Plan 639, this table, and the glossary-first rule in the main skill.
**Weight-Space Programming → katgpt-rs code (§10):** paper-vocabulary grep for `weight patch|writer edits weights|WeightSpace|dspy-wsp` finds little outside `.research/648`. Codebase-vocabulary grep for `WeightEditOverlay|closed_form_weight_write|QuantErrorLora|factor_tail|check_epoch_tag|LaneSlot|MerkleFrozenEnvelope|system_one` finds the whole pipeline: production (`weight_edit_overlay.rs` Issue 980 + `closed_form_weight_write.rs` Issue 990, Research 659/664), the low-rank patch form (`sst_certified.rs` `TailFactor`), artifact integrity (`MerkleFrozenEnvelope`, `graft_artifact.rs`), serve-time swap (`check_epoch_tag`/`LaneSlot`), and the paired/holdout eval discipline (reflex harness). The decision-model serving half maps to `RiirAgent::system_one` / decision_wire / the `typed_decisions` suite — dspy-wsp's backends speak the same `/v1/systemone` shape. Without §10's translation, a WSP-family paper distill re-runs Research 659's keyword miss ("weight-space" vocabulary grep returned zero; caught only by owner challenge).
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 |