Persistent cross-session cognition and reproducible inference evidence around DeepSeek-V3, without changing model weights.
In 60 seconds
The strongest fit is an external persistent-cognition layer around inference, not a rewrite of DeepSeek-V3. Build durable memory and replay evidence first; use the small DREAM components for offline consolidation; defer full DREAM until persistence, erasure and replay are proven.
Adds durable cross-session experience, offline decay/prune/strengthen/merge, and a generation-depth gate before recalled memory can influence later prompts. It is external memory around inference; it does not change DeepSeek weights.
Composite 1: STIER-S-130 + BLD-ENG-260 + BLD-ENG-271 → inference/generate.py:121-144
Seals prompt/token/seed/config/checkpoint/environment metadata and joins it to checkpoint-conversion lineage and tamper evidence, without claiming bitwise equality across hardware or precision modes.
Composite 2: VPACK-DETERMINISTIC-REPLAY-VAULT + BLD-ENG-249 + STIER-S-71 → inference/generate.py; inference/convert.py; inference/fp8_cast_bf16.py
Turns the expert-selection counts DeepSeek already computes and discards into baselines, drift alerts, and resource-waste signals while leaving routing decisions unchanged.
Composite 3: STIER-S-104 + STIER-S-109 → inference/model.py:535-598, 669-693
Makes any future durable memory redact-before-write and selectively erasable, so adding persistence does not silently create an unmanaged retention surface.
Composite 4: BLD-ENG-247 + STIER-S-PHA03 + STIER-S-130 → new persistence adapter at inference/generate.py
Each claim above is carried, with its basis, in the numbered steps below. Nothing may appear here that is not derived from a row in Step 4, Step 5A or Step 5B. Licence gate: $499 USD one time under CMPSBL(R) Perpetual Edition License 1.0; no recurring fee/no expiry. This run is evaluation-only until licensed implementation source is supplied. The nonprofit twin is a separate licensor grant.
Step 0
Run before any file is copied. A STOP here ends the evaluation; it is not a defect to work around. The supplied host code is MIT and the model is separately governed by the DeepSeek Model License. The authoritative CMPSBL LICENSE/LICENSING.md sets a $499 one-time perpetual fee; older annual/free-door wording elsewhere in the kit is superseded.
| Host licence (SPDX) | MIT (code); separate DeepSeek Model License for model use/distribution |
|---|---|
| Where it was found | LICENSE-CODE; LICENSE-MODEL |
| Purchase or grant record | No CMPSBL purchase/grant record naming this host was supplied in this pre-purchase evaluation. |
| Licence | CMPSBL(R) Perpetual Edition License 1.0 (LicenseRef-CMPSBL-Perpetual-1.0), $499 USD one time; no recurring fee/no expiry. |
| Verdict | PROCEED with read-only evaluation. STOP before copying, generating, wiring or actuating CMPSBL implementation material until licensed source/purchase evidence is supplied. |
| Clause that drove it | CMPSBL LICENSE section 1 grant; section 2 non-transfer/non-sublicense/no redistribution; section 1.4 one-time fee. Host MIT and DeepSeek Model License remain independently applicable. |
| Scope limits imposed | Evaluation only in this run. Preserve host license notices. The nonprofit twin cannot be transferred by a purchaser under the standard license; it must be issued by the licensor under its own grant record or written permission. |
Sub-package exceptions — workspaces published under a different licence than the repo root, e.g. an npm-published package that must stay permissive: No separately licensed host subpackage was found in the supplied archive. The code/model license split is root-level and both remain applicable.
Verdict
Worth $499 if DeepSeek-Infer is becoming a persistent assistant or research harness; unnecessary if it remains only the reference demo.
The purchase price is not the hard part. The integration burden is the Python bridge, persistence schema, testing and privacy review. The library is compelling here only when the product goal includes durable memory, offline consolidation and replayable inference evidence.
Step 5A — the section that sells
Each entry below is a component that does not exist in the catalog and does not exist in this repository. It comes into being only when the named ingredients are combined — library parts with each other, or library parts with this host's own modules. Four host-specific composites survived the full screen. Composite 1 is the highest-leverage result; Composite 4 is the minimum safety contract if persistent memory is introduced.
A durable, self-consolidating memory layer around DeepSeek inference. Selected experiences survive process restarts, are consolidated offline, and are filtered by synthesis generation depth before later recall.
| Composition | CMPSBL STIER-S-130 "Cross-Session Persistence" + BLD-ENG-260 "Dream Synthesizer" + BLD-ENG-271 "Dream Generation Gate" + host inference/generate.py |
|---|---|
| Problem it removes | generate.py:121-144 stores conversation only in a Python messages list; process exit destroys it. model.py:439-444 separately marks attention caches persistent=False. |
| Runtime behaviour | After each turn, selected experience is stored outside the model. On an idle cadence, fragments are decayed, pruned, strengthened and merged. Before a later prompt is assembled, only bounded ranked memories are injected. |
| New because | Persistence only saves state; Dream Synthesizer only consolidates supplied fragments; Generation Gate only bounds/ranks synthesized candidates. Together they create persistent cognition none provides alone. |
| Attachment point | inference/generate.py:121-144 at the message lifecycle and chat-template boundary; a host persistence adapter owns durable storage. |
| Leaves untouched | inference/model.py architecture, checkpoint tensors, MLA caches, expert routing, tokenizer and sampling implementation. |
| Effort | ~6-10 engineer-days (basis: 228 combined entry LOC + persistence adapter + Python bridge + memory schema + scheduling + recall/privacy/regression tests). |
| Verified (weakest ingredient, verbatim) | C:ok I:ok B:— T:— S:scan. Behavior: Not claimed. Test coverage: Not claimed. Security is an author-run static scan, not an independent audit. |
user turn -> redact/write -> durable episode store
|
v
idle DreamSynthesizer
|
Generation Gate
|
v
next turn -> ranked recall -> chat template -> DeepSeek generate()
A replay-and-lineage receipt for inference and checkpoint conversion. It records enough run context to make reproduction attempts explicit and auditable without overstating hardware-level determinism.
| Composition | CMPSBL VPACK-DETERMINISTIC-REPLAY-VAULT "Bug Reproduction & Replay Layer" + BLD-ENG-249 "Pipeline Lineage Registry" + STIER-S-71 "Tamper-Evident Chain" + host generate.py + convert.py + fp8_cast_bf16.py |
|---|---|
| Problem it removes | generate.py:111 fixes a Torch seed but emits no run receipt; convert.py:50-85 and fp8_cast_bf16.py:63-103 write transformed checkpoint files without cryptographic lineage. |
| Runtime behaviour | Seal prompt tokens, temperature, Torch seed, config hash, checkpoint identifiers, generated token IDs and environment metadata into a replay capsule; link conversion input/output hashes into a tamper-evident chain. |
| New because | Replay alone does not identify which conversion produced a checkpoint; lineage alone is mutable metadata; a hash chain alone has no inference semantics. The combination creates a portable evidence artifact. |
| Attachment point | inference/generate.py generate/sample; inference/convert.py main; inference/fp8_cast_bf16.py main, around I/O boundaries. |
| Leaves untouched | Numerical kernels and model outputs. No claim of bitwise equality across GPU models, kernels, drivers or precision modes. |
| Effort | ~6-9 engineer-days (basis: 217 combined entry LOC + hashing + schema + environment capture + replay harness). |
| Verified (weakest ingredient, verbatim) | C:ok I:ok B:— T:— S:scan. Behavior: Not claimed on the weakest ingredient. Test coverage: Not claimed. Security is an author-run static scan, not an independent audit. |
prompt/config/checkpoint -> seal replay metadata
conversion stages ---------> lineage hashes -> tamper chain
|
v
inference receipt
|
v
replay attempt
An observation-only monitor for expert-routing distributions, layer latency, and resource hotspots built from signals the DeepSeek MoE path already computes.
| Composition | CMPSBL STIER-S-104 "Performance Regression Detector" + STIER-S-109 "Resource Waste Profiler" + host Gate/MoE in inference/model.py |
|---|---|
| Problem it removes | MoE.forward computes counts = torch.bincount(...) at model.py:683 but the counts are local dispatch data and are not retained, baselined or analyzed. |
| Runtime behaviour | Capture expert-selection distributions, per-layer latency and utilization snapshots; compare against accepted baselines; flag route drift or sustained expert hotspots and quantify resource waste. |
| New because | The regression detector has no MoE semantics and the waste profiler has no expert-routing signal. DeepSeek supplies the signal that turns both into a model-specific monitor. |
| Attachment point | inference/model.py Gate.forward and MoE.forward; optional timing around inference/kernel.py. |
| Leaves untouched | Routing scores, selected experts and model outputs; first version is observation-only. |
| Effort | ~4-7 engineer-days (basis: 127 combined entry LOC + instrumentation + baselines + GPU metric collection). |
| Verified (weakest ingredient, verbatim) | C:ok I:ok B:load T:— S:scan. Behavior is load/export checked only. Test coverage: Not claimed. Security is an author-run static scan, not an independent audit. |
Gate scores -> selected expert ids -> counts/latency
|
+--------------+-------------+
v v
Regression Detector Waste Profiler
+--------------+-------------+
v
routing drift report
The safety contract for adding durable memory: secrets are removed before storage and targeted records can later be erased with a proof/status.
| Composition | CMPSBL BLD-ENG-247 "Secret Redactor" + STIER-S-PHA03 "Selective Amnesia Controller" + STIER-S-130 "Cross-Session Persistence" + host inference/generate.py |
|---|---|
| Problem it removes | Durable memory would create a data-retention surface the current process-local chat loop does not have. Persisting raw prompts without redaction and targeted erasure would create a privacy/security regression. |
| Runtime behaviour | Redact secrets before commit; tag stored records for selective erasure; retain an erasure proof/status; preserve normal persistence/conflict handling for remaining records. |
| New because | Redaction protects new writes; amnesia deletes addressable data; persistence makes data durable. Together they define the minimum safety lifecycle for cross-session memory. |
| Attachment point | write/read/delete boundary around the persistence adapter called from inference/generate.py. |
| Leaves untouched | Model weights and attention-cache internals. This governs only the external memory store. |
| Effort | ~4-7 engineer-days (basis: 220 combined entry LOC + tagging/indexing + erase-path tests + policy wiring). |
| Verified (weakest ingredient, verbatim) | C:ok I:ok B:— T:— S:scan. Behavior: Not claimed on the weakest ingredient. Test coverage: Not claimed. Security is an author-run static scan, not an independent audit. |
prompt/episode -> Secret Redactor -> persistent store
|
request to forget -> Selective Amnesia +----> erase proof
|
v
remaining memory
| Composite | Ingredients | New behaviour | Effort |
|---|---|---|---|
| Persistent Dream Memory Loop | STIER-S-130 "Cross-Session Persistence" + BLD-ENG-260 "Dream Synthesizer" + BLD-ENG-271 "Dream Generation Gate" + inference/generate.py | After each turn, selected experience is stored outside the model. | ~6-10 engineer-days (basis: 228 combined entry LOC + persistence adapter + Python bridge + memory schema + scheduling + recall/privacy/regression tests). |
| Inference Evidence Capsule | VPACK-DETERMINISTIC-REPLAY-VAULT "Bug Reproduction & Replay Layer" + BLD-ENG-249 "Pipeline Lineage Registry" + STIER-S-71 "Tamper-Evident Chain" + generate.py + convert.py + fp8_cast_bf16.py | Seal prompt tokens, temperature, Torch seed, config hash, checkpoint identifiers, generated token IDs and environment metadata into a replay capsule; link conversion input/output hashes into a tamper-evident chain. | ~6-9 engineer-days (basis: 217 combined entry LOC + hashing + schema + environment capture + replay harness). |
| MoE Routing Drift Monitor | STIER-S-104 "Performance Regression Detector" + STIER-S-109 "Resource Waste Profiler" + Gate/MoE in inference/model.py | Capture expert-selection distributions, per-layer latency and utilization snapshots; compare against accepted baselines; flag route drift or sustained expert hotspots and quantify resource waste. | ~4-7 engineer-days (basis: 127 combined entry LOC + instrumentation + baselines + GPU metric collection). |
| Memory Safety Envelope | BLD-ENG-247 "Secret Redactor" + STIER-S-PHA03 "Selective Amnesia Controller" + STIER-S-130 "Cross-Session Persistence" + inference/generate.py | Redact secrets before commit; tag stored records for selective erasure; retain an erasure proof/status; preserve normal persistence/conflict handling for remaining records. | ~4-7 engineer-days (basis: 220 combined entry LOC + tagging/indexing + erase-path tests + policy wiring). |
Build this composite first: Persistent Dream Memory Loop, but only after a minimal persistence/replay contract is testable. It changes process-local chat history into durable, bounded, offline-consolidated experience without modifying model weights.
A composite inherits the weakest verification claim of its ingredients and never a stronger one. Where an ingredient reads “Not claimed”, the composite reads “Not claimed”.
Step 5B — new single-component designs
Single-purpose components designed for this repository, built from named catalog components plus named host files. Read them after the composites in Step 5A, which are the higher-value half of this step. Four narrow designs survived. They are deliberately smaller than the composites and each has a concrete host attachment and evidence path.
Replace the raw messages list as the orchestration surface with explicit thread IDs, forks, retention priority and windowing. Persistence remains a separate concern.
| Built from | CMPSBL STIER-S-103 "Context Threading Engine" (ThreadMessage, ConversationThread, ThreadManager, createThreadManager) |
|---|---|
| Plus host | inference/generate.py:121-144 |
| New because | The host has conversational history but no thread semantics; the design becomes DeepSeek-specific at the chat-template boundary. |
| Unlocks | Long conversations can fork and compact without model changes. |
| Evidence path | Thread metadata plus the retained-message decision can be inspected before tokenization. |
| Effort | ~1-2 engineer-days (basis: 95 entry LOC + Python bridge + message-schema mapping). |
| Constraints | C:ok I:ok B:exec T:— S:scan; mode direct; ports none. Test coverage: Not claimed; security audit: Not claimed. |
| Confidence | high - direct attachment point and narrow behavior. |
input -> ThreadManager -> retained/forked messages -> tokenizer -> generate()
Wrap each conversion stage with input/output identifiers and hashes so a resulting shard can be traced to source files and conversion parameters.
| Built from | CMPSBL BLD-ENG-249 "Pipeline Lineage Registry" (PipelineLineageRecord, PipelineLineageRegistry) |
|---|---|
| Plus host | inference/convert.py:50-85; inference/fp8_cast_bf16.py:63-103 |
| New because | The catalog component is generic; this artifact is a DeepSeek checkpoint receipt tied to the host conversion semantics. |
| Unlocks | A release can show which source tensors and settings produced each distributed shard. |
| Evidence path | A small receipt adjacent to converted weights; it proves lineage, not numerical correctness. |
| Effort | ~1-2 engineer-days (basis: 39 entry LOC + hashing + conversion-hook tests). |
| Constraints | C:ok I:ok B:— T:— S:scan; mode direct; ports none. Behavior: Not claimed. Test coverage: Not claimed; security audit: Not claimed. |
| Confidence | high - conversion boundaries are explicit and side-effect free to observe. |
source safetensors -> convert/dequant -> output shard
| |
+---------- lineage ----------+ -> receipt
Record accepted latency/throughput baselines for representative FP8 quantization, dequantization and GEMM shapes; fail a benchmark gate when a change exceeds the selected tolerance.
| Built from | CMPSBL STIER-S-104 "Performance Regression Detector" (PerformanceBaseline, RegressionAlert, setBaseline, checkRegression, listBaselines) |
|---|---|
| Plus host | inference/kernel.py plus a new host benchmark harness |
| New because | The component is generic; the design turns DeepSeek Triton kernels into a versioned performance contract. |
| Unlocks | Kernel changes stop shipping on anecdotal spot checks alone. |
| Evidence path | Benchmark result plus named baseline plus regression alert. |
| Effort | ~2-4 engineer-days (basis: 65 entry LOC + GPU benchmark matrix + CI/environment controls). |
| Constraints | C:ok I:ok B:exec T:— S:scan; mode direct; ports none. Test coverage: Not claimed; security audit: Not claimed. |
| Confidence | medium-high - benchmark stability across GPU types must be controlled. |
kernel shape/dtype -> benchmark -> checkRegression -> pass / alert
After basic consolidation works, use memory-store saturation, novelty yield and rejected-candidate rate to choose a shallow or deeper consolidation cycle.
| Built from | CMPSBL BLD-ENG-162 "Dream Lucidity Control" (DreamCycle, DreamLucidityControl) |
|---|---|
| Plus host | Persistent Dream Memory Loop around inference/generate.py |
| New because | Dream Lucidity Control has no DeepSeek memory signals by itself; the host-specific governor emerges from those measured signals. |
| Unlocks | Prevents maximum-depth consolidation from becoming an always-on default. |
| Evidence path | Per-cycle depth decision plus consolidation-result summary. |
| Effort | ~1-2 engineer-days (basis: 58 entry LOC + metrics mapping + stop-condition tests). |
| Constraints | C:ok I:ok B:— T:— S:scan; mode direct; ports none. Behavior: Not claimed. Test coverage: Not claimed; security audit: Not claimed. |
| Confidence | medium - useful only after real consolidation metrics exist. |
memory metrics -> DreamLucidityControl -> cycle depth -> DreamSynthesizer
Build this one first: Checkpoint Conversion Receipt for the lowest-risk first validation. Once persistence is approved, Persistent Dream Memory Loop remains the highest-leverage overall build.
Scoreboard
Every figure below carries the basis it was derived from. A number without a basis is a defect in this report.
| Measure | Value | Basis |
|---|---|---|
| Engineered catalog coverage | 1,162 / 1,162 | Basis: complete parse of Part 2 engineered rows; no sampled classes. |
| Discovery Vault coverage | 6,479 / 6,479 | Basis: complete parse of DISCOVERY-VAULT-INDEX.md; source unavailable pre-purchase. |
| Five-axis candidates screened | 32 | Basis: 32 full catalog rows selected after host-first inventory and then evaluated on all five axes. |
| Items recommended | 13 | Basis: 12 CARRY + 1 CARRY-LATE after twins, host mismatches and shared-entry exclusions. |
| Quick wins | 4 | Basis: four low-risk actions in Step 4, each tied to an existing host boundary. |
| New single-component designs | 4 | Basis: Step 5B; each uses a named catalog ID plus a named DeepSeek host file. |
| New composite components | 4 | Basis: Step 5A; each combines 2-3 catalog parts with DeepSeek host code. |
| (A) cost avoided | $6,000-$10,800 | Basis: ~5-9 engineer-days x 8 hours/day x $150/hour. |
| (B) capability acquisition | $12,000-$26,400 | Basis: ~10-22 engineer-days x 8 hours/day x $150/hour; counterfactual capability, not revenue. |
| (C) portfolio reuse | $0 | Basis: no concrete sibling product plan is established by this archive strongly enough to book carry-over value. |
| Integration cost | $9,600-$18,000 | Basis: ~8-15 engineer-days x 8 hours/day x $150/hour for bridge, persistence, tests, privacy/security review and learning. |
| Net after $499 licence | -$499 to +$27,101 | Basis: low = $18,000 gross - $18,000 integration - $499; high = $37,200 gross - $9,600 integration - $499. |
| Licence-only break-even | ~3.3 engineer-hours | Basis: $499 / $150 per hour; this excludes integration labor and is not overall break-even. |
Step 1
Inventory was completed before catalog matching. The supplied repository is a compact inference reference implementation plus checkpoint conversion utilities, configs, documentation and licences; it is not a production serving stack.
| Host unit | Path | Exported surface | LOC | Job it does |
|---|---|---|---|---|
| Unit 1 | inference/model.py | ModelArgs; ParallelEmbedding; linear; Linear; ColumnParallelLinear; RowParallelLinear; RMSNorm; precompute_freqs_cis; apply_rotary_emb; MLA; MLP; Gate; Expert; MoE; Block; Transformer | 808 | Core MLA/MoE transformer, model-parallel layers, transient KV/PE caches, expert routing. |
| Unit 2 | inference/generate.py | sample; generate; main | 185 | Interactive/batch generation, process-local messages list, tokenizer chat template, fixed Torch seed. |
| Unit 3 | inference/kernel.py | act_quant_kernel; act_quant; weight_dequant_kernel; weight_dequant; fp8_gemm_kernel; fp8_gemm | 196 | Triton FP8 quantization/dequantization/GEMM kernels. |
| Unit 4 | inference/convert.py | main | 96 | Converts HF safetensor checkpoints into model-parallel demo shards; no digest/provenance receipt. |
| Unit 5 | inference/fp8_cast_bf16.py | main | 112 | FP8-to-BF16 conversion with a small loaded-file cache; warns and continues on missing scale_inv. |
| Unit 6 | inference/configs/config_16B.json | ModelArgs values | 19 | 16B demo model configuration. |
| Unit 7 | inference/configs/config_236B.json | ModelArgs values | 20 | 236B demo model configuration. |
| Unit 8 | inference/configs/config_671B.json | ModelArgs values | 22 | 671B DeepSeek-V3 demo configuration: 256 routed experts, 8 activated experts. |
| Unit 9 | inference/configs/config_v3.1.json | ModelArgs values | 23 | V3.1 configuration with FP8 scale_fmt ue8m0. |
| Unit 10 | inference/requirements.txt | torch; triton; transformers; safetensors | 4 | Pinned runtime dependencies. |
| Unit 11 | README.md | Documentation | 361 | Architecture, benchmarks, demo instructions; explicitly labels DeepSeek-Infer as example only and recommends serving frameworks. |
| Unit 12 | README_WEIGHTS.md | Documentation | 94 | Weight layout, MTP module, FP8 scaling/dequantization documentation. |
| Unit 13 | LICENSE-CODE | MIT | 21 | Code licence. |
| Unit 14 | LICENSE-MODEL | DeepSeek Model License | 91 | Separate model-use, redistribution, attribution and use-restriction terms. |
Step 2
Every candidate carried forward from the catalog, scored on the five axes before ranking. Rows that fail an axis stay in the table with the reason — a silently dropped candidate is a defect. All 1,162 engineered rows were parsed first. Thirty-two full rows then formed the host-specific candidate set and were evaluated on all five axes. Discovery Vault screening was run separately across all 6,479 index rows.
| Candidate (ID) | Direct fit | Combinatorial | Value | Cost / risk | Novelty | Carried forward? |
|---|---|---|---|---|---|---|
| STIER-S-130 Cross-Session Persistence [twin of] BLD-ENG-080 | Direct: closes the durable cross-session state gap around generate.py. | Combines with Dream Synthesizer + Generation Gate into persistent cognition. | High if DeepSeek-Infer becomes a stateful assistant/research harness; low if it stays a demo. | Low component size; real cost is persistence backend + Python bridge + privacy review. | Novel in this repo: yes. | CARRY |
| BLD-ENG-260 Dream Synthesizer | Direct: provides offline decay/prune/strengthen/merge over stored experience. | Core of the Persistent Dream Memory Loop. | High for long-running assistants because memory quality can improve without weight updates. | 1–5 day component band; requires stored-memory schema and scheduling. | Novel: yes; this repo has no offline consolidation loop. | CARRY |
| BLD-ENG-271 Dream Generation Gate | Direct: bounds recursive synthesis depth and semantic drift. | Safety gate for Dream Synthesizer or full DREAM. | High trust value once synthesized memories can feed later prompts. | Hours-scale component; policy thresholds still require evaluation. | Novel: yes. | CARRY |
| BLD-ENG-162 Dream Lucidity Control | Direct: controls consolidation depth; no host equivalent. | Adds adaptive depth control after the simple Dream loop is stable. | Medium: operational tuning rather than foundational capability. | Hours-scale component; host metrics needed to drive it intelligently. | Novel: yes. | CARRY |
| PRIM-ENG-01 DREAM | Direct fit only after persistence exists; not a first-step drop-in. | Can mine sub-threshold experience signals and cluster novel patterns over idle cycles. | Potentially very high research value; not required for ordinary inference. | 748 entry LOC, 47,391 closure LOC, PORT-PERSISTENCE, C:proj(12); Python bridge and validation dominate cost. | Novel: strongly yes. | CARRY-LATE |
| BLD-ENG-247 Secret Redactor | Direct: prevents secrets from entering a new persistent memory store. | Forms the write-side boundary of the Memory Safety Envelope. | High downside prevention; little user-facing product value by itself. | Hours-scale; must be applied before persistence writes, not after. | Novel: no; standard requirement, correctly placed. | CARRY |
| STIER-S-PHA03 Selective Amnesia Controller [twin of] BLD-ENG-192 | Direct: provides controlled deletion once durable memory is introduced. | Combines with redaction + persistence for an erasable memory system. | High trust/privacy value for user-specific memory. | 1–5 day component band plus host data-index work. | Novel in this repo: yes. | CARRY |
| VPACK-DETERMINISTIC-REPLAY-VAULT Bug Reproduction & Replay Layer | Direct: seals input/seed/result into replay capsules. | With lineage + tamper chain becomes an Inference Evidence Capsule. | High engineering/debug value for expensive distributed inference runs. | 1–5 day band; must define what constitutes replay-equivalent across GPU/precision. | Novel: yes in this repo. | CARRY |
| BLD-ENG-249 Pipeline Lineage Registry | Direct: records stage input/output hashes and provenance walks. | Attaches to checkpoint conversion and generation evidence. | High for reproducibility and release confidence. | Hours-scale component; host hashing and artifact IDs must be wired. | Novel: moderately. | CARRY |
| STIER-S-71 Tamper-Evident Chain [twin of] BLD-ENG-009 | Direct: makes receipts tamper-evident rather than mutable log records. | Third ingredient in Inference Evidence Capsule. | Medium-high trust value; especially useful for benchmark/release evidence. | Hours-scale component. | Novel: moderately. | CARRY |
| STIER-S-104 Performance Regression Detector [twin of] BLD-ENG-076 | Direct: compares performance against baselines. | With MoE routing counts can detect routing/performance drift. | High for kernel and routing regressions. | Hours-scale component plus benchmark harness creation. | Novel: no; the combination with expert routing is the novel part. | CARRY |
| STIER-S-109 Resource Waste Profiler | Direct fit to GPU/compute utilization only if the host supplies ResourceUsage samples. | With Gate/MoE counts turns hot expert distributions into actionable waste signals. | Medium-high for large-scale inference cost control. | Hours-scale component; telemetry collection is additional host work. | Novel: yes as applied to MoE routing. | CARRY |
| STIER-S-103 Context Threading Engine | Direct: replaces an unstructured Python list with thread/window/priority semantics. | Pairs with persistent recall without changing model internals. | Medium: better long-session behavior, but not durable memory by itself. | Hours-scale component plus Python bridge. | Novel: no; known assistant architecture. | CARRY |
| BLD-ENG-289 Working Memory | Overlaps transient context/attention rather than durable memory. | Could sit above prompts, but adds another short-term buffer before persistence is solved. | Low now. | 178 LOC plus bridge for a problem the model/context path already handles. | Not novel. | REJECT |
| STIER-S-134 Embedding Store [twin of] BLD-ENG-073 | In-memory embeddings do not close the durable-store gap and the host does not expose an embedding pipeline here. | Could become useful later with an explicit embedding producer and persistence adapter. | Medium later, low now. | Small, but incomplete as a memory system. | Not novel. | REJECT |
| BLD-ENG-029 Nocturne Consolidation | Same consolidation problem as BLD-ENG-260 with a narrower API. | Redundant beside the chosen Dream Synthesizer. | No incremental value after BLD-ENG-260. | Avoid duplicate implementation. | No. | REJECT |
| VPACK-NOCTURNE-CONSOLIDATION Memory Consolidation & Pattern Reinforcement Layer | Valid capability but overlaps the chosen small direct Dream path. | Could replace BLD-ENG-260 if its pack interface better fits a later delivery. | Medium alternative, not additive. | 107 entry LOC; avoid two consolidation engines. | No. | REJECT |
| SUBSYS-08 Memory Stream | Contains persistent-agent/DREAM concepts, but the host is a tiny Python inference demo. | Would import a 29,458-LOC closure plus PORT-PERSISTENCE/PORT-PRESENTATION and React peer surface. | Low net fit here. | High integration mismatch; use small direct parts instead. | Novel but over-scoped. | REJECT |
| SENGINE-DREAM-ENGINE Dream Engine | Capability matches, but row is shared-entry and cannot be sized from pre-purchase metadata. | Whole-engine route duplicates the smaller direct Dream pieces. | Potential value, insufficient pre-purchase evidence for sizing. | Shared-entry 1,890 LOC belongs to class, PORT-PERSISTENCE. | Novel, but evidence ceiling blocks recommendation. | NOT-EVALUABLE |
| SENGINE-MEMORY-ENGINE Memory Engine | Capability matches durable/context memory. | Would overlap the smaller persistence + context components. | Potentially high, but not defensibly sizeable here. | Shared-entry 1,890 LOC, PORT-PERSISTENCE. | Not novel conceptually. | NOT-EVALUABLE |
| LAYER-COGNITIVE-MEMORY Long-Term Knowledge Graph Memory Layer | Layer description matches the target outcome. | Could be a broader alternative after purchase. | Potentially high. | Shared-entry 494 LOC; class-wide surface cannot be used for estimation. | Not novel. | NOT-EVALUABLE |
| LAYER-NOCTURNE-CONSOLIDATION Memory Consolidation & Pattern Reinforcement Layer | Matches consolidation, but duplicates direct Dream components. | Broad layer is unnecessary for the first integration. | Medium. | Shared-entry 494 LOC; exclude from sizing. | No. | NOT-EVALUABLE |
| STIER-S-CJ122 Dream Cross-Pollination Engine | No independent dream pools exist in the host to cross-pollinate. | Only useful after multiple memory domains/agents exist. | Low now. | Premature architecture. | Novel but unjustified. | REJECT |
| STIER-S-90 Forensic Replay Engine | Replay is valuable, but chosen replay-vault + lineage composition is more specific to inference receipts. | Would duplicate replay responsibilities. | Medium, redundant. | 126 LOC + bridge; avoid competing replay systems. | No. | REJECT |
| STIER-S-OBS02 Distributed Tracing Synthesizer | The demo uses distributed model parallelism but no request/service graph needing end-to-end spans. | Reconsider in SGLang/vLLM serving integration, not this reference repo. | Low here, high in a service host. | 90 LOC + instrumentation. | No. | REJECT |
| BLD-PRM-001 Seeded RNG | DeepSeek already calls torch.manual_seed(965); a separate Mulberry32 RNG would not make CUDA kernels bitwise deterministic. | Replay should record the actual Torch seed and environment instead. | Low. | Wrong abstraction for GPU inference determinism. | No. | REJECT |
| BLD-ENG-176 Episodic Replay | Episodic replay is aimed at reinforcement/learning; this repo has no training loop. | Could be revisited in a training/evaluation repository. | Low here. | Adds a learning abstraction without a learner. | No. | REJECT |
| BLD-ENG-080 Cross-Session Persistence [twin of] STIER-S-130 | Twin of STIER-S-130. | No new capability. | None beyond selected twin. | Duplicate. | No. | REJECT |
| BLD-ENG-009 Tamper-Evident Chain [twin of] STIER-S-71 | Twin of STIER-S-71. | No new capability. | None beyond selected twin. | Duplicate. | No. | REJECT |
| BLD-ENG-192 Selective Amnesia Controller [twin of] STIER-S-PHA03 | Twin of STIER-S-PHA03. | No new capability. | None beyond selected twin. | Duplicate. | No. | REJECT |
| BLD-ENG-073 Embedding Store [twin of] STIER-S-134 | Twin of STIER-S-134. | No new capability. | None beyond rejected twin. | Duplicate. | No. | REJECT |
| BLD-ENG-094 Embedding Similarity | Basic vector similarity could support semantic memory, but this DeepSeek demo exposes no embedding-production or durable vector-store path. | Useful only after an explicit embedding extractor/store exists; it does not itself create persistent recall. | Low now; medium later if semantic retrieval becomes a requirement. | 102 entry LOC plus Python bridge, but would trigger additional embedding and storage infrastructure. | No; similarity math alone is not the missing durable-memory capability. | REJECT |
Candidates screened: 32 full candidates after 1,162-row catalog parse · carried forward:
13 (12 CARRY + 1 CARRY-LATE) · rejected: 15; 4 additional rows NOT-EVALUABLE for sizing. Rows flagged
[shared-entry] in the catalog were not scored from their
catalog columns; how they were handled: four [shared-entry] rows were excluded from size/effort scoring because their Entry LOC and API surface belong to class-wide entry points; pre-purchase source was not available to isolate them.
Step 3
What the host's runtime, build, licence and release shape forbid, independent of whether a component is attractive. DeepSeek is Python/PyTorch/Triton and model-parallel; the catalog implementation is canonical TypeScript. The integration must therefore respect the licensed bridge and preserve DeepSeek model/code boundaries instead of pretending the kit contains a native Python implementation.
| Constraint | Source of truth | What it rules out |
|---|---|---|
| Language/toolchain boundary | inference/requirements.txt; kit non-TypeScript rules |
No hand-written pretend port. Licensed implementation must use the supported Python bridge boundary around canonical TypeScript semantics. |
| GPU/distributed runtime | inference/model.py; inference/generate.py |
No component may assume a single-process global state when DeepSeek runs model-parallel across ranks. |
| Durable-memory boundary | inference/generate.py:121-144; inference/model.py:439-444 |
Do not count process-local chat history or persistent=False attention caches as cross-session memory. |
| Reference-demo release shape | README.md DeepSeek-Infer section |
Do not force production request tracing, rate limiting or serving orchestration into this example repo when the README assigns production serving to external frameworks. |
| Host licence scope | LICENSE-CODE; LICENSE-MODEL |
Preserve MIT code terms and independently satisfy the DeepSeek Model License. CMPSBL evaluation is read-only until a valid purchase/grant supplies licensed implementation source. |
| CMPSBL runtime mode | recommended catalog rows |
No carried component is governed-required. Twelve carried rows are direct; full PRIM-ENG-01 DREAM is self-contained and additionally requires PORT-PERSISTENCE. |
| Persistence adapter | PRIM-ENG-01 constraints |
PORT-PERSISTENCE must be implemented before full DREAM can run; portable defaults are not production persistence. |
| Memory safety | new external-memory surface |
Any durable user memory requires write-time redaction, retention/erasure policy and host security/privacy review; S:scan is not an independent audit. |
Every recommended component's Verified column, quoted
verbatim with Not claimed left in those words:
STIER-S-130 Cross-Session Persistence [twin of] BLD-ENG-080: C:ok I:ok B:exec T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point loads and zero-arity exports were executed twice and compared structurally. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
BLD-ENG-260 Dream Synthesizer: C:ok I:ok B:— T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Not claimed. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
BLD-ENG-271 Dream Generation Gate: C:ok I:ok B:— T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Not claimed. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
BLD-ENG-162 Dream Lucidity Control: C:ok I:ok B:— T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Not claimed. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
PRIM-ENG-01 DREAM: C:proj(12) I:ok B:exec T:— S:scan - Compile: Compiles under project settings; 12 strict-mode errors (source was authored with strict off). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point loads and zero-arity exports were executed twice and compared structurally. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: self-contained. Ports/peers: ports: PORT-PERSISTENCE.
BLD-ENG-247 Secret Redactor: C:ok I:ok B:— T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Not claimed. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
STIER-S-PHA03 Selective Amnesia Controller [twin of] BLD-ENG-192: C:ok I:ok B:exec T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point loads and zero-arity exports were executed twice and compared structurally. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
VPACK-DETERMINISTIC-REPLAY-VAULT Bug Reproduction & Replay Layer: C:ok I:ok B:load T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point imports cleanly and exposes exports; no zero-arity surface was executed without fabricated input. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
BLD-ENG-249 Pipeline Lineage Registry: C:ok I:ok B:— T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Not claimed. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
STIER-S-71 Tamper-Evident Chain [twin of] BLD-ENG-009: C:ok I:ok B:exec T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point loads and zero-arity exports were executed twice and compared structurally. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
STIER-S-104 Performance Regression Detector [twin of] BLD-ENG-076: C:ok I:ok B:exec T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point loads and zero-arity exports were executed twice and compared structurally. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
STIER-S-109 Resource Waste Profiler: C:ok I:ok B:load T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point imports cleanly and exposes exports; no zero-arity surface was executed without fabricated input. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
STIER-S-103 Context Threading Engine: C:ok I:ok B:exec T:— S:scan - Compile: Compiles (batch tsgo --strict --noEmit). Integration: Installed with install.mjs into an empty host project and type-checked there with tsgo --noEmit --skipLibCheck. Behavior: Entry point loads and zero-arity exports were executed twice and compared structurally. Test coverage: Not claimed. Security: Static unsafe-construct sweep: clean (author-run 13-rule deterministic scan; not an independent security audit). Mode: direct. Ports/peers: none.
Governed runtime: No carried recommendation is governed-required. Twelve are direct; PRIM-ENG-01 DREAM is self-contained. No governed slices are required by the carried set. — any
component whose runtimeRequirements.requiresGovernedRuntime is
true must ship with its listed slices or not at all.
Adapter ports that must be implemented before first call: PORT-PERSISTENCE for full PRIM-ENG-01 DREAM. The smaller first-stage memory components list no catalog adapter ports, but the host still must supply the actual durable backend and Python bridge.
Step 4
Ranked by expected value to this product — not by catalog prestige. The list favors durable memory, replay/provenance and observable MoE behavior. It deliberately demotes generic production-service infrastructure because the host README assigns production serving to external frameworks.
| # | Item | Catalog IDs | Host paths | Effort | Value | Confidence |
|---|---|---|---|---|---|---|
| 1 | Cross-Session Persistence | STIER-S-130 | inference/generate.py:121-144 |
~1-2 engineer-days (basis: 75 entry LOC + persistence adapter + Python bridge) | Very high if a stateful assistant/research harness is desired | high |
| 2 | Dream Synthesizer | BLD-ENG-260 | new memory job; inference/generate.py boundary |
~2-4 engineer-days (basis: 106 entry LOC + memory schema + scheduling + bridge) | Very high after persistence | high |
| 3 | Bug Reproduction & Replay Layer | VPACK-DETERMINISTIC-REPLAY-VAULT | inference/generate.py; inference/convert.py; inference/fp8_cast_bf16.py |
~2-4 engineer-days (basis: 110 entry LOC + receipt schema + environment capture + bridge) | High reproducibility/debug value | high |
| 4 | Pipeline Lineage Registry | BLD-ENG-249 | inference/convert.py; inference/fp8_cast_bf16.py |
~1-2 engineer-days (basis: 39 entry LOC + hashing + hook tests) | High release/provenance value | high |
| 5 | Dream Generation Gate | BLD-ENG-271 | memory promotion path |
~0.5-1.5 engineer-days (basis: 47 entry LOC + host policy thresholds + bridge) | High safety value for synthesized memory | high |
| 6 | Secret Redactor | BLD-ENG-247 | memory write boundary |
~0.5-1 engineer-day (basis: 41 entry LOC + persistence write hook + fixtures) | High downside prevention | high |
| 7 | Selective Amnesia Controller | STIER-S-PHA03 | memory delete boundary |
~2-4 engineer-days (basis: 104 entry LOC + memory indexing + erasure tests) | High privacy/trust value | medium-high |
| 8 | Performance Regression Detector | STIER-S-104 | inference/kernel.py; MoE metrics |
~1-3 engineer-days (basis: 65 entry LOC + benchmark harness + baseline policy) | High engineering-quality value | high |
| 9 | Resource Waste Profiler | STIER-S-109 | inference/model.py:683 + GPU metrics |
~1-3 engineer-days (basis: 62 entry LOC + GPU telemetry mapping) | Medium-high compute-cost value | medium |
| 10 | Tamper-Evident Chain | STIER-S-71 | evidence receipt store |
~0.5-1.5 engineer-days (basis: 68 entry LOC + receipt-chain persistence) | Medium-high integrity value | high |
| 11 | Context Threading Engine | STIER-S-103 | inference/generate.py:121-144 |
~1-2 engineer-days (basis: 95 entry LOC + Python bridge + schema mapping) | Medium long-session value | medium-high |
| 12 | Dream Lucidity Control | BLD-ENG-162 | Dream consolidation scheduler |
~1-2 engineer-days (basis: 58 entry LOC + metrics mapping + stop tests) | Medium tuning value | medium |
| 13 | DREAM | PRIM-ENG-01 | post-persistence idle pipeline |
~2-4 engineer-weeks (basis: 748 entry LOC + 12 strict-mode errors + PORT-PERSISTENCE + Python bridge + controlled evaluation) | Potentially very high research value; not required for ordinary inference | medium |
[HARDEN] CMPSBL STIER-S-130 "Cross-Session Persistence" -> host: inference/generate.py (185 LOC) Why: The host loses its messages list on process exit; this adds durable session state and conflict resolution at the exact chat boundary. Uses: PersistedState, MergeResult, save, load, resolveConflict, listPersistedKeys, clearAll Saves: ~4-12 engineer-hours, basis = 75 Entry LOC + 2 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:exec T:— S:scan Mode: direct Ports to implement: none
[COMBINE] CMPSBL BLD-ENG-260 "Dream Synthesizer" -> host: inference/generate.py (185 LOC) + new memory job Why: The host has no offline memory lifecycle; this supplies decay, prune, strengthen and merge once durable episodes exist. Uses: MemoryFragment, ConsolidationResult, DreamConfig, DreamSynthesizer Saves: ~1-5 engineer-days, basis = 106 Entry LOC + 4 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:— T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL VPACK-DETERMINISTIC-REPLAY-VAULT "Bug Reproduction & Replay Layer" -> host: generate.py + convert.py + fp8_cast_bf16.py (393 LOC combined) Why: The host seeds Torch but stores no run receipt; replay capsules make inputs, seeds and results explicit for reproduction attempts. Uses: cmpsbl_replay_seal, cmpsbl_replay_get, cmpsbl_replay_count, DETERMINISTIC_REPLAY_VAULT_LAYER Saves: ~1-5 engineer-days, basis = 110 Entry LOC + 4 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:load T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL BLD-ENG-249 "Pipeline Lineage Registry" -> host: convert.py + fp8_cast_bf16.py (208 LOC combined) Why: Checkpoint conversion writes output directly with no stage hashes; lineage records close that provenance gap. Uses: PipelineLineageRecord, PipelineLineageRegistry Saves: ~4-12 engineer-hours, basis = 39 Entry LOC + 2 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:— T:— S:scan Mode: direct Ports to implement: none
[COMBINE] CMPSBL BLD-ENG-271 "Dream Generation Gate" -> host: new memory promotion path at generate.py Why: Once synthesized memories can feed future prompts, generation-depth limits bound recursive semantic drift. Uses: DreamCandidate, MAX_GENERATION, memoryWeight, sortBySynthesisPriority, filterDreamCandidates, rankDreamCandidates Saves: ~4-12 engineer-hours, basis = 47 Entry LOC + 2 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:— T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL BLD-ENG-247 "Secret Redactor" -> host: new persistence write boundary at generate.py Why: Persistent memory would otherwise store raw secrets; redaction belongs at the write boundary before data becomes durable. Uses: redactSecrets, redactContext Saves: ~4-12 engineer-hours, basis = 41 Entry LOC + 2 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:— T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL STIER-S-PHA03 "Selective Amnesia Controller" -> host: new persistence delete boundary at generate.py Why: Durability without targeted erasure creates a retention trap; this adds addressable deletion and an erasure proof. Uses: ErasureRequest, ErasureProof, AmnesiaStats, createSelectiveAmnesiaController Saves: ~1-5 engineer-days, basis = 104 Entry LOC + 3 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:exec T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL STIER-S-104 "Performance Regression Detector" -> host: kernel.py (196 LOC) + model.py telemetry Why: There is no benchmark gate; accepted baselines and alerts make kernel/routing regressions observable. Uses: PerformanceBaseline, RegressionAlert, setBaseline, checkRegression, listBaselines Saves: ~4-12 engineer-hours, basis = 65 Entry LOC + 3 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:exec T:— S:scan Mode: direct Ports to implement: none
[COMBINE] CMPSBL STIER-S-109 "Resource Waste Profiler" -> host: model.py (808 LOC), especially line 683 Why: DeepSeek already computes expert counts; mapping those counts to resource usage can expose sustained hotspots and waste. Uses: ResourceUsage, WasteReport, WasteItem, profileWaste Saves: ~4-12 engineer-hours, basis = 62 Entry LOC + 3 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:load T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL STIER-S-71 "Tamper-Evident Chain" -> host: new evidence receipt store Why: Replay and lineage receipts are mutable without integrity chaining; this makes later tampering detectable. Uses: createAuditChain Saves: ~4-12 engineer-hours, basis = 68 Entry LOC + 2 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:exec T:— S:scan Mode: direct Ports to implement: none
[HARDEN] CMPSBL STIER-S-103 "Context Threading Engine" -> host: generate.py (185 LOC) Why: The process-local list has no thread IDs, forks, windowing or priority retention; this hardens long-session orchestration. Uses: ThreadMessage, ConversationThread, ThreadManager, createThreadManager Saves: ~4-12 engineer-hours, basis = 95 Entry LOC + 3 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:exec T:— S:scan Mode: direct Ports to implement: none
[COMBINE] CMPSBL BLD-ENG-162 "Dream Lucidity Control" -> host: new Dream consolidation scheduler Why: After consolidation exists, fixed-depth dream cycles are a poor default; this introduces measured depth control. Uses: DreamCycle, DreamLucidityControl Saves: ~4-12 engineer-hours, basis = 58 Entry LOC + 2 correctness concerns; integration effort is budgeted separately. Verified: C:ok I:ok B:— T:— S:scan Mode: direct Ports to implement: none
[SKIP] CMPSBL PRIM-ENG-01 "DREAM" -> host: post-persistence idle pipeline around generate.py Why: It has the highest research ceiling, but it requires trustworthy persisted episodes, replay evidence and deletion controls first; therefore it is explicitly deferred. Uses: PrimitiveClassification, PrimitiveSource, PrimitiveCatalogEntry, PRIMITIVE_CATALOG, registerAllPrimitives, getCatalogEntry, getCatalogBySource, getCatalogNames, invokePrimitive +51 more Saves: ~1-3 engineer-weeks, basis = 748 Entry LOC + 6 correctness concerns; integration effort is budgeted separately. Verified: C:proj(12) I:ok B:exec T:— S:scan Mode: self-contained Ports to implement: ports: PORT-PERSISTENCE
| Gap | Evidence in host | What addresses it |
|---|---|---|
| No durable conversation memory | generate.py:121-144 process-local messages; no durable store | STIER-S-130 + Persistent Dream Memory Loop |
| No offline memory consolidation | No background jobs or memory lifecycle code | BLD-ENG-260; later PRIM-ENG-01 DREAM |
| No synthesized-memory depth control | No memory synthesis exists | BLD-ENG-271 + BLD-ENG-162 |
| No erase/redact contract for future memory | No durable memory today; adding it creates the obligation | BLD-ENG-247 + STIER-S-PHA03 |
| No run/checkpoint provenance | No hashlib/sha256 code; conversions write outputs directly | VPACK-DETERMINISTIC-REPLAY-VAULT + BLD-ENG-249 + STIER-S-71 |
| MoE route counts discarded | model.py:683 computes counts but does not export/store them | STIER-S-104 + STIER-S-109 composite |
| No performance regression gate | No tests/bench harness in repo | STIER-S-104 + host benchmark harness |
Low effort, high confidence, nothing on a critical path.
Items that would duplicate, undermine or dilute systems already trusted here. A report with no entries here evaluated nothing.
Step 6
Three buckets, never merged. (A) is hard, (B) is counterfactual, (C) is conditional. Assumption: one senior ML/platform engineer at a blended US rate of $150/hour, 8 hours/day. Entry LOC is used only as the catalog estimation-band key; closure LOC is not monetized. Ranges include host-specific correctness/integration concerns rather than multiplying LOC by rate.
| Bucket | Engineer-weeks | US$ (range) | Basis and assumptions |
|---|---|---|---|
| (A) Cost avoided — work already planned | ~1.0-1.8 | $6,000-$10,800 | ~5-9 engineer-days (basis: replay/provenance, conversion receipts, regression gate and basic persistence mechanics likely required if this demo becomes a durable research harness). |
| (B) Capability acquisition — would never have been built | ~2.0-4.4 | $12,000-$26,400 | ~10-22 engineer-days (basis: Persistent Dream Memory Loop, bounded synthesis, erasable memory and MoE-routing drift analysis; counterfactual capability acquisition, not revenue). |
| (C) Portfolio and future reuse — conditional | 0 | $0 | No concrete sibling repositories or next-24-month product plan is established by this archive strongly enough to book carry-over value. No assumption is made about future distinct-product grant records. |
| − Integration cost (ports, wiring, security review, learning) | ~1.6-3.0 | $9,600-$18,000 | ~8-15 engineer-days (basis: Python bridge, persistence adapter, memory schema, retrieval policy, GPU benchmark harness, host tests, privacy/security review and familiarization). |
| = Net against the applicable tier | ~0.0-4.6 equivalent weeks | -$499 to +$27,101 | Low: $18,000 gross - $18,000 integration - $499. High: $37,200 gross - $9,600 integration - $499. The equivalent-week range is not a staffing schedule. |
US senior ML/platform planning rate $150/hour; one primary engineer; 8-hour days and 5-day weeks. Low case assumes some existing persistence/benchmark infrastructure can be reused outside this archive; high case assumes more capability must be created/reviewed from scratch. The $499 fee is one time under the authoritative license.
The items least likely to arise from this repo alone are the Persistent Dream Memory Loop, Memory Safety Envelope and the MoE Routing Drift Monitor composition. Together they change the product shape from a stateless reference demo into a stateful experimental system that can accumulate curated experience while retaining evidence about what influenced a run.
Yes if the intended product is a persistent DeepSeek-based assistant/research harness and the team will implement at least the memory loop plus replay/evidence. No if this repository intentionally remains the example-only inference demo and all stateful production behavior lives in external serving frameworks.
What would change this answer: A production target repo with an existing persistence store and test harness would improve the economics; a requirement that every enhancement live only inside this minimal demo with no sidecar/runtime would make the Python/TypeScript bridge less attractive.
Step 7
| Phase | What ships | Why here | Gate to proceed | Effort | Cumulative |
|---|---|---|---|---|---|
| 0–30 days | Replay metadata + Checkpoint Conversion Receipt; persistence interface; Secret Redactor; minimal Cross-Session Persistence; restart tests. | Establish evidence and data contracts before autonomous consolidation. | Stop memory work if restart tests cannot reproduce the exact stored/recalled records or write-time redaction fails adversarial fixtures. | ~8-12 engineer-days (basis: evidence schema + bridge + persistence/redaction + tests) | ~8-12 days |
| 30–60 days | Persistent Dream Memory Loop: Dream Synthesizer + Generation Gate; selective amnesia; context-thread integration. | Only after storage, replay and erase paths are explicit. | Run a fixed evaluation set with and without recall. Stop automatic recall if memory-assisted runs regress more tasks than they improve or stale/incorrect memories repeatedly survive consolidation. | ~10-16 engineer-days (basis: consolidation + recall policy + erasure + evaluation) | ~18-28 days |
| 60–90 days | Dream Lucidity Control; observation-only MoE Routing Drift Monitor; controlled trial of full PRIM-ENG-01 DREAM on a copy of the experience store. | Adaptive depth and full DREAM need real memory/replay telemetry to judge utility. | Stop full DREAM promotion if two controlled cycles fail to yield independently reviewable useful novel patterns or synthesized-memory drift cannot be bounded by the Generation Gate. | ~10-18 engineer-days (basis: metrics + MoE baselines + full-DREAM bridge/port trial) | ~28-46 days |
Licensed Python bridge support; durable persistence adapter; memory schema; test/evaluation harness; cryptographic hash utility; retention/erasure policy; GPU environment metadata for replay/regression work.
Full PRIM-ENG-01 DREAM. It is the highest-ceiling component but the wrong first dependency: without durable input traces, deletion controls and replay evidence, there is nothing trustworthy to consolidate or audit.
Design 2 lands in days 0-30; Design 1 with memory in days 30-60; Design 3 after stable GPU benchmarks; Design 4 only after real consolidation metrics exist.
Build the Persistent Dream Memory Loop after the minimal replay/persistence contract passes.
Capacity assumption: one primary engineer with occasional security/ML review; the roadmap does not assume six parallel workstreams.
Discovery Vault — provisional, kept out of every total
Screened from DISCOVERY-VAULT-INDEX.md rows only. Pre-purchase
the vault source does not ship, so no API surface, LOC figure or verification
verdict is stated here and nothing in this section is carried into any
savings, effort or net total above.
| Discovery id | Capability (index row) | Host concern it may touch | Status | What to read first post-purchase |
|---|---|---|---|---|
| disc-237c53ce Compositional Consolidator | Replays event histories while persistsing state signals across the substrate. Primitive chain: AUDIT -> BLACKOUT -> DREAM -> ECHO -> MEMORY -> SCHOLAR -> TREATY | Closest vault concept to DREAM + MEMORY + replayed history; inspect first after purchase. | PROVISIONAL — index row only, source not read (pre-purchase) | vault/disc-237c53ce-compositional-consolidator.ts |
| disc-29d97d4c Zero-Shot Amplifier | Simulates hypothetical configurations while unifiesing external interfaces across the substrate. Primitive chain: DREAM -> INTEGRATION -> KINETIC -> LEDGER -> MANDATE -> MEMORY -> SCRIBE | DREAM + MEMORY + LEDGER chain may inform a memory-with-receipts composite. | PROVISIONAL — index row only, source not read (pre-purchase) | vault/disc-29d97d4c-zero-shot-amplifier.ts |
| disc-2ebd5c9a Emergent Planner | Persists state signals while simulatesing hypothetical configurations across the substrate. Primitive chain: BRAIN -> CLARITY -> DREAM -> MEMORY -> SCHOLAR -> SYLLOGISM | DREAM + MEMORY chain with planning; relevant only if memory later drives higher-level task planning. | PROVISIONAL — index row only, source not read (pre-purchase) | vault/disc-2ebd5c9a-emergent-planner.ts |
| disc-06dcd2a3 Concept Drift Corrector | Detects and corrects semantic concept drift in long-running cognitive sessions Primitive chain: BRAIN -> VISION -> CORTEX | Could police semantic drift in long-running memory-assisted sessions. | PROVISIONAL — index row only, source not read (pre-purchase) | vault/disc-06dcd2a3-concept-drift-corrector.ts |
| disc-0b461625 Cognitive Flame Graph Generator | Generates flame graphs of cognitive execution paths Primitive chain: OBSERVABILITY -> CORTEX -> BRAIN -> ANALYTICS | Could visualize cognitive/retrieval execution paths if a richer runtime emerges. | PROVISIONAL — index row only, source not read (pre-purchase) | vault/disc-0b461625-cognitive-flame-graph-generator.ts |
| disc-030d599a Adaptive Monitor | Inspects structural patterns while replaysing event histories across the substrate. Primitive chain: ATLAS -> CORTEX -> ECHO -> HERALD -> LINEAGE -> OVERSEER -> ULT_HERALD -> VISION | Replay + lineage + observability chain could complement inference evidence. | PROVISIONAL — index row only, source not read (pre-purchase) | vault/disc-030d599a-adaptive-monitor.ts |
All 6,479 index rows were screened; six provisional rows survived. Their source was not read and their LOC/verification are excluded from every total. If this section is empty, say in one line why the vault produced no candidates rather than deleting the section.
Step 8 — appended, never merged into the body
| Claim above | Re-check result | Estimate change |
|---|---|---|
| The host has chat memory | Confirmed only as process-local history: messages=[] at generate.py:122, append at 139/144, clear at 137. | Corrected terminology: conversational history is not persistent memory. |
| The model has caches | Confirmed K/V/KV/PE buffers are explicitly persistent=False at model.py:439-444. | Confirmed these are inference caches and must not be counted as durable memory. |
| The host is deterministic because it seeds Torch | torch.manual_seed(965) is present, but temperature sampling uses an exponential draw and no environment receipt is stored. | Narrowed claim: seeded run metadata helps replay; no bitwise-determinism claim across hardware/precision. |
| MoE observability is absent | MoE.forward computes expert counts at line 683 but does not persist/export them. | Improved recommendation: instrument an existing signal instead of inventing a new routing probe. |
| Checkpoint conversion lacks provenance | Confirmed no hashlib/sha256 usage; outputs are written directly. FP8->BF16 warns and continues on missing scale_inv. | Raised Checkpoint Conversion Receipt to a first-30-day quick win. |
| Production hardening belongs here | README calls DeepSeek-Infer “example only” and recommends external serving frameworks. | Reduced tracing/service-hardening scope and made the overall buy verdict conditional on a stateful research/assistant goal. |
| Tests may exist outside inference | Full file/path scan found zero host test files. | Confirmed; integration estimate includes creating a harness rather than reusing one from this archive. |
| The kit price language is internally inconsistent | LICENSE and LICENSING.md state $499 one time, perpetual/no recurring fee; an older Phase-1 directive still contains superseded annual/free-door language. | Corrected the report to use the authoritative Perpetual Edition terms and treat only the nonprofit-twin instruction as run-specific. |
The second pass did not invalidate the four composites. It narrowed the scope: this is a reference inference repo, so serving-stack hardening was demoted; memory is explicitly external; full DREAM remains post-persistence. The second pass also caught a kit documentation defect: LICENSE and LICENSING.md are authoritative at $499 one time, while older Phase-1/CTA wording is stale.
Claim ceiling
Extracted from systems the author has run in production since August 2025. Operationally exercised, not independently security-audited.
In pre-purchase evaluation, verification axes quoted in this report come
verbatim from the generated AGENT-HANDOFF-CATALOG.md, which is
the pre-purchase source of truth; in a full delivery,
artifacts/<ID>/artifact-manifest.json is authoritative and
overrides it. Every component
records testCoverage: Not claimed, and no component has had an
independent third-party security review. Static scan results and adversarial
test suites are the author's own.
Verification axes cited in this report: C (compile), I (integration), B (behavior), T (test coverage), S (security). Test coverage is Not claimed for every recommended component. S:scan is an author-run static sweep, not a third-party audit.
1,162 components · 2,135 source files · 446,551 lines. CMPSBL(R) Perpetual Edition License 1.0: US$499 USD one time; no recurring fee and no expiry. The standard grant is perpetual, worldwide, non-exclusive, non-transferable and non-sublicensable; it permits internal use, private modification, and embedding in a product with substantial independent function. For this evaluation, the licensor additionally directs that an identical enhancement package be issued separately to the selected nonprofit, Ai2; the $499 payment is not donated. Offline delivery, no account, no telemetry.