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

EMENDER E97 4B Pi Instruction Checkpoints

Public operational checkpoint repository for the behaviorally selected EMENDER E97 4B recurrent Pi coding agent.

Promoted checkpoint

pi-live-aligned-u8 is promoted on the fixed 120-task real-Pi core panel:

• 119/120 full task passes;
• 120/120 exact tool sequences and sandbox postconditions;
• 119/120 protocol-valid grounded finals and exact tool arguments;
• bash, edit, read, write, and recover-test: 20/20 each;
• recover-read: 19/20.

The sole failure repeated an unchanged missing-path read and was correctly stopped by the no-progress cycle guard.

Checkpoint: checkpoints/pi-live-aligned-u8/checkpoint_agent_sft_u000008_loss_0.0213.pt

SHA-256: b799802741737058c4de74e233a8af8e6a9a18977cbf753ad44f9037a27c3da8

See releases/pi-live-aligned-u8/release.json and its data, reload, and real-Pi evaluation receipts for the complete immutable identity.

Model and lineage

• Architecture: E97, 4,045,972,080 unique parameters
• Shape: d=3840, depth=18, heads=60, recurrent state side=64
• Foundation parent: step 24,448 / 99,723,771,904 accepted tokens
• Foundation SHA-256: 3ace004251643acf2e7c7f720e8f29968ad0a483441553c0c885b87b3df84568
• Behavioral SFT parent: mixed-curriculum update 64
• Repair: eight live-aligned all-assistant updates with fresh Schedule-Free AdamW

The live-aligned authority reproduces Pi's (no tool output) representation for successful empty tools while retaining action targets. Longer broad SFT and final-only repair endpoints were rejected despite lower loss because real-Pi tool behavior regressed.

Required runtime

Use Emender source commit 95f50c295b34f4e71033ef7629ba996e205e47a3 or a reviewed descendant preserving its agent protocol. The serving cycle guard permits a check to be rerun after an intervening repair action but still rejects immediately repeated identical calls.

Canonical Pi prompt and tool contract

This checkpoint is not a generic chat checkpoint. A generic assistant prompt can produce repetition or otherwise degenerate output. Use the exact newline-terminated prompt published at integration/e97-pi-core-system-prompt.txt:

You are a coding agent operating in Pi in the current working directory. Use read, bash, edit, and write to inspect, change, and verify repository files. For a tool call respond with exactly Action and one JSON Arguments object. Never invent tool results. After the work is verified, respond with Final and a concise evidence-grounded summary.

Its SHA-256 is c362e82f7ab243477c65375862aef6af2ee394a0edc6341eb8e8d5b65b1da1a3. Use it with the closed read, bash, edit, and write Pi extension published at integration/e97-core-tools.ts (SHA-256 19d23cd09753eeb76704772d6994d04af761630611a265c2d3164ad9d109f3c6). Assistant tool turns have the exact grammar:

Action: read
Arguments: {"path":"README.md","offset":1,"limit":80}

Terminal turns begin with Final:. Successful empty tools are represented in the next observation as the literal (no tool output). Reproduction code and the complete transcript/runtime contract are available in Emender commit 927216340dc4418f7cf2fb027e87b3edc055c3fd.

The reported fixed-panel result is a narrow core-tool milestone, not evidence of broad repository-level coding ability.

Latest development candidate

pi-cumulative-recovery-u8 is the newest instruction checkpoint, published as a transparent development baseline rather than a broad capability promotion:

• original Pi core smoke: 120/120;
• template-held-out compositional v2: 226/240;
• first family-held-out blind v3: 0/240.

The v3 failure showed memorized path/template substitution and immediate repeated calls on novel task families. See releases/pi-cumulative-recovery-u8/release.json. The checkpoint is checkpoints/pi-cumulative-recovery-u8/checkpoint_agent_sft_u000008_loss_0.1604.pt with SHA-256 eed48419d657b0be03953f6c99dc0ba14e61579ee748174ef0e6794f895157ed.

This candidate supersedes the older checkpoint for ongoing broad post-training experiments, but LATEST.json continues to identify the historical narrow-panel promotion. LATEST_CANDIDATE.json identifies this newer development artifact.

Security and loading

This repository contains a raw torch.save checkpoint and therefore Python pickle. Do not load it from a mutable branch or unverified download. Pin an immutable repository revision, download the named file, verify its SHA-256, and only then load it in a trusted environment.