emender-e97-4b-pi-instruction-checkpoints

EMENDER E97 4B — Pi-Native Coding Agent

EMENDER E97 4B is a 4,045,972,080-parameter hybrid recurrent language model trained from scratch (~100B tokens), then developed into an observation-driven conversational tool agent through a verified post-training program: task proposals are executed by a real coding-agent harness (Pi), and only deterministically verified trajectories enter training. Every capability claim below was measured on frozen evaluation panels under a fixed protocol, with full audit receipts retained in the release folders.

Current release: releases/repair-v6-u96-promoted/ — PROMOTED qualified baseline

The first checkpoint in the program's history to pass its frozen dual gate. All thresholds passed on a single measurement with no re-rolls:

• Execution: 71/96 on the 96-case real-workspace panel (live OpenHands-compatible environment, autonomous multi-turn), beating the representation-bridge parent's 67/96: fresh 32/32 · prior-fresh 16/16 · prior-regression 11/16 (the previously persistent error-repetition loop is broken) · transfer 8/16 · composition 4/16
• Pi-native tool panel (Stage-B): 12/14 valid frames, 10/14 correct first actions — gate leg passed, reproduced exactly across two independent runs
• Retention: all windows pass (tool format 1.000 token accuracy; conversation and development NLL within their frozen envelopes)
• Checkpoint SHA-256: d81464982c3ebc0d72769a87e079068bf535d6ca2010185dc1bb03ce264b8f5b
• Promotion receipt and full gate audit chain: releases/repair-v6-u96-promoted/release.json

Known weaknesses — read before testing

• Weak composition: only 4/16 multi-step composition cases pass.
• Conversational drift: retention windows pass, but long conversations remain the model's most fragile behavior.
• No hostile-code isolation claims — the sandbox is an integration boundary, not a security boundary.
• A longer-horizon "wide arc" (16x the training data, near-zero repetition, full admitted trace pools) is training now; this README will be updated with its gate results.

What testing is most wanted

1. Long conversations — where does drift become behavioral?
2. Multi-step file editing and composition (the weakest measured axis)
3. Tool-use edge cases: error recovery, misroutes, budget exhaustion
4. Any loop, stall, or invalid tool call — report with transcripts; failures feed the program's detector-driven data generation.

Running it

OpenAI-compatible serving with the Emender repo:

git clone https://github.com/spinozans/emender.git
python scripts/serve_e97_agent_openai.py \
  --checkpoint <downloaded release folder> --args-json args.json

Point any OpenAI-API coding-agent client (e.g., Pi with the bundled integration/e97-core-tools.ts tool surface and system prompt) at the server. The model speaks a Pi-native eleven-tool surface: read, bash, edit, write, process, ffgrep, fffind, web_search, source_check, fetch_content, get_search_content.

Superseded history

• releases/repair-v5-u128/ — the earlier experimental canary (gate near-miss: Stage-B passed 14/14 valid, 11/14 correct; execution 70/96 with the loop broken; failed the conversation-retention window and one prior-fresh case). Retained for collaborative failure-finding and provenance.
• Earlier u8-era releases — superseded; see release folders.

Limitations and license

The policy has not demonstrated autonomous tool discovery or general software engineering ability. No standalone model license has yet been selected; license: other is intentional.