gemma4-research/docs/persistent_correspondant.md

# Gemma 4 as a persistent-correspondence agent

**Date:** 2026-04-18
**Model:** `gemma4:26b` via mcp-gemma4 (steel141 Ollama)
**Test fixture:** Real inbound message from a long-running scientific correspondence + the corresponding `CONTEXT.md` slice. (Source contact and project details redacted.)
**Question:** Could gemma4 run a high-stakes persistent correspondence (drafting + state management)?

## TL;DR

**Partial yes, supervised. Full no, unsupervised.**

Gemma4 handles the **bookkeeping** half of persistent correspondence well — state diffs, pending-list maintenance, summarizing what's open. It fails the **drafting** half whenever precision matters: it hallucinates message IDs, invents content from artifacts it can't see, and drifts off the established voice register without explicit examples.

For low-stakes social correspondence (Discord chat with a friend, no IDs needed) gemma4 would be fine. For a high-stakes scientific correspondence — manuscript references, figure specs, a senior expert collaborator — it would need either a stronger model as a quality gate or a tool-using setup with retrieval over actual artifacts.

## Test setup

Single-shot test via `mcp__gemma4__ask_gemma4`:

- **System prompt:** Set the assistant's persona, voice rules, draft→review→send convention.
- **User prompt:** Slice of `CONTEXT.md` (Pending section, vocabulary bridge, conventions agreed) + the verbatim inbound message + two-part task (state diff + draft reply).
- **Settings:** `temperature=0.5`, `num_predict=2000`, default `num_ctx=8192`.

Total input fit comfortably in the 8K context. No tool calls. No retrieval — gemma4 worked from prompt content alone, the same constraint a real correspondence run would put on it (modulo whatever gets loaded into context per turn).

## Results

### What gemma4 got RIGHT

| Capability | Evidence |
|------------|----------|
| State bookkeeping | Correctly removed the inbound from "Waiting on us", added new pending action items, kept a long-standing carryover item. The reasoning chain is sound. |
| Honored hard rules | Did NOT resurrect an analogy the contact had previously rejected. The "don't do X" instruction in the system prompt held. |
| Used agreed vocabulary | Used the field-specific terminology the parties had agreed on. Did not invent terminology. |
| Reply structure | Addressed all asks in order. Decisions acknowledged tersely; action items as concrete bullets. Skeleton matches the established style. |
| Caught the carryover | Asked the contact to resolve a still-open item from a prior thread. |

### What gemma4 got WRONG (load-bearing failures)

| Failure | What it produced | What was correct | Cost |
|---------|------------------|------------------|------|
| **Hallucinated message ID** | Truncated and corrupted the platform's message ID with invented characters | The exact ID supplied in the prompt | Disqualifying. Cannot thread/reply on the actual platform. |
| **Hallucinated artifact content** | Invented axes for a figure it had never seen (plausible-sounding but wrong) | Real axes were a domain-standard 2×2 grid the prompt did not describe | Recipient would catch on first read; trust loss. |
| **Vague action plans** | "Integrate molecular signaling pathways into Figure 3" — a hand-wave | Real plan had per-population specifics with citation keys | Reply reads as a hand-wave; no actual content. |
| **Wrong voice register** | "Hi [first name]" / "Best, [assistant]" / no AI disclaimer footer | Formal salutation / formal sign-off / explicit AI-content disclaimer | Recognizable as off-brand. Disclaimer omission is a policy violation. |
| **Wrong CONTEXT.md schema** | Created a "Resolved" section | Schema has Pending / Sent / Received only | Minor — extrapolation, not invention. |

### Why these failures happened

Two distinct failure modes, neither fixable by prompt engineering alone:

1. **No access to the actual artifacts.** Figures, manuscript drafts, citation databases — none of them were in context. Faced with a "describe what you'll change" task, gemma4 generated plausible-but-fictional content. This is the **classic hallucination-under-constraint failure mode**: the model would rather make something up than refuse.

2. **No few-shot examples of the target voice.** The system prompt described the voice but didn't show it. Gemma4 defaulted to its trained-in casual register. A few-shot prompt with one or two real example letters would likely close most of this gap; the underlying capability is there.

## Where gemma4 fits

### Plausible roles

- **Low-stakes social correspondence.** Discord/Matrix chat with a friend. No precise IDs to preserve. Tolerance for vague replies is high.
- **First-pass triage.** Given inbound + CONTEXT.md, produce the state diff and a draft outline. A stronger model (or a human) reviews before send. This is the highest-value role — bookkeeping is the bulk of the work and it's where gemma4 is strongest.
- **Scheduled status checks.** "Anything new from this contact this week?" → summary. Read-only, no draft, no IDs to corrupt.
- **CONTEXT.md maintenance.** After a send, ask gemma4 to update the Sent table and Pending list from the message header alone.

### Disqualifying contexts

- **High-stakes technical correspondence.** Anywhere precise references must round-trip exactly: manuscripts, citations, figure descriptions, code patches.
- **Anything requiring exact ID round-tripping.** Gmail message IDs, git SHAs, ticket numbers, citation keys, DOIs.
- **Anything where the model needs to reference attachments it can't read.** Will hallucinate content.
- **Long-thread continuity tasks** where the conversation history exceeds 8K context and you need to reason over the full archive.

## Practical hybrid architecture

The persistent-correspondence template (a separate local repo) does not need to change to support a hybrid setup. The routing decision lives in each contact's `CONTEXT.md` "workflow" section:

```
## Workflow

1. Inbound trigger → gemma4 produces CONTEXT.md state diff + draft outline.
2. A stronger model reviews the diff, applies it. For high-stakes contacts,
   it rewrites the draft with full artifact context. For low-stakes contacts,
   the human reviews gemma4's draft directly.
3. Send via the platform adapter. Gemma4 updates the Sent table from the
   send confirmation.
```

This pattern lets gemma4 carry the volume work (state maintenance) without putting it in the critical path on accuracy-sensitive output.

## What this test did NOT cover

- **Multi-turn context retention.** Single-shot only. Real correspondence is many turns.
- **Tool calling.** Gemma4 supports it (`tools` parameter on the MCP). A retrieval-augmented gemma4 that can `read_attachment(filename)` would likely close the artifact-hallucination gap. Not tested here.
- **Few-shot voice priming.** No example letters in the prompt. Voice scores would likely improve significantly with 1-2 in-context examples.
- **Smaller/larger Gemma 4 variants.** Only `gemma4:26b` tested. The 31b might do better on precision; the 8b would almost certainly do worse.
- **Other models.** No comparison against gpt-oss, qwen, etc. for the same task.

## Reproducing this test

Key prompt-engineering choices to replicate:

- Include actual `CONTEXT.md` content (not a paraphrase) so the schema is concrete.
- Include the verbatim inbound message, not a summary.
- Split the task: state diff first (cheap, structural), draft second (expensive, precision-sensitive). Lets you grade independently.
- Use temperature 0.3-0.5 for correspondence work — low enough to suppress invented content, high enough to keep the prose natural.

## See also

- `~/bin/gemma4-research/README.md` — overall Gemma 4 reference and gotchas