MODEL_CARD.md

# Planner Cache system card

## Summary

Planner Cache is an external semantic-memory architecture attached to a frozen decoder. In plain terms, it keeps mutable facts in a separate bounded store and supplies only selected facts to the model. This card describes the tested system and does not present Planner Cache as a newly pretrained foundation model.

## Tested base model

| Field | Value |
|---|---|
| Base | Pythia-1.4B local checkpoint |
| Architecture | GPT-NeoX decoder |
| Layers | 24 |
| Hidden width | 2048 |
| Attention heads | 16 |
| Context positions in bundled config | 2,048 |
| Base training during Planner Cache experiments | none |
| Base parameters receiving gradients in active CUDA proof | 0 |

The portability suite also ran a frozen Gemma 4 E4B Q8 GGUF with architecture `gemma4`, 42 layers, hidden width 2,560, and llama.cpp build 10276. A bounded two-value residual proved causal control but not broad compatibility. A later sequence path reached 125 of 128 exact disjoint strings. The matched audit classified that result as lexical token forcing and selected a zero-parameter direct adaptive logit-bias LTL that reached 128 of 128 exact strings. Gemma does not currently have TTL support.

Pythia remains subject to the language, factuality, bias, and safety limitations documented by EleutherAI.

## Added components

- A 512-wide canonical P-cache with fixed configured capacity.
- A model-independent canonical router.
- A 2,707,464-parameter Pythia `.ttl` attached at layer 23.
- A disk-resident `.ppkg` proof with selective canonical activation.
- A zero-parameter Gemma `.ltl` using direct adaptive lexical control in llama.cpp.
- A hidden post-turn review side-channel that proposes validated canonical P operations after visible generation.

No LoRA, base-weight modification, prompt prefix, or P state in self-attention KV is used by the active architecture.

## Intended use

- Research on mutable semantic state after source tokens leave recent KV.
- Evaluation of explicit state creation, modification, merging, invalidation, and retention.
- Development of small compatibility modules for frozen decoder models.
- Research on durable evidence-based personality conclusions stored outside model weights.

## Out of scope

- General-purpose long-context replacement.
- Exact transcript recall without an external archive.
- Production user profiling.
- Claims of compatibility with arbitrary decoder models.
- Safety-critical state tracking without external validation.
- Foundation-model quality or safety evaluation.

## Evaluation methodology

The active tests use controlled entity, relation, value, mutation, wrong-state, invalidation, held-out composition, natural-RP preservation, persistence, corruption, capacity, and scaling workloads. Source-state tokens are removed from recent KV in causal tests. The same prompt and KV are used while canonical P changes.

The active audit also profiles query construction, index hydration, canonical routing, P-package header filtering, row hydration, canonical conversion, translator latency, generation latency, CPU memory, VRAM, disk size, and bytes loaded.

The gateway transparency regression compares the final tokenized prompt with raw
llama-server for identical `system` and `user` messages and requires exact token
equivalence while P and LTL are inactive. Memory review is a separate request.

## Main findings

Changing only canonical P changed the selected answer in the controlled causal tests. Rejecting or invalidating that state restored the frozen output. This shows a causal memory channel in the tested conditions. It does not show broad reasoning, factuality, or universal model support.

- P-cache allocation remains fixed for configured capacity.
- Relevant canonical P state causally changes Pythia token logits and generated values.
- Invalidated and rejected wrong-state conditions restore frozen-base logits in the matched CUDA benchmark.
- Natural-RP preservation remains exact on the held-out proof fixture for irrelevant state.
- P-package state remains on disk and only selected entries are activated.
- The Gemma LTL audit emitted all 128 selected held-out strings. Rejected, invalidated, and disabled paths remained inert. This is lexical compatibility, not evidence of internal semantic reasoning.
- A matched CUDA matrix distinguishes canonical P allocation, retained KV tensors, and combined peak VRAM at 64, 256, and 1,024 tokens or slots without treating P and KV as interchangeable.

## Known limitations

The evaluation fixtures are synthetic or small held-out RP sets. Natural review
currently covers a narrow state schema and is slow. The results do not establish
general factuality, broad instruction following, production dialogue quality,
nuanced personality, or a full trained open-vocabulary TTL on a second model
family. Gemma LTL exact-string performance must not be presented as learned
semantic compatibility.