[joelteply]

Continuum card eaeb2b87, S4 (a), the in-place dream stage (ONE-RESIDENT-MODEL §8.1 / §9). Today a serving context learns for exactly ONE mind per server run: a second resident's /learn is refused with "one shadow per server run". This lets every resident on a lane learn in place, each with her own shadow and optimizer.

Change

• One optimizer per shadow. llama_context keeps one optimizer per shadow (opt_minds), and opt_ctx/opt_shadow are the active pair, so every existing training site is unchanged.
  • llama_opt_init with another adapter on a configured context calls opt_select. It creates that shadow's optimizer the first time and keeps each shadow's moments and accumulators across switches.
  • Only the active shadow's tensors carry PARAM, so a graph never trains a parameter the active optimizer lacks.
• opt_configured() replaces opt_ctx != nullptr where the question is whether the context is sized for training (reserve skip, graph sizing). Releasing the active shadow would otherwise have rebuilt a scheduler that other shadows' optimizers still hold.
• Unload releases instead of refusing. llama_opt_release drops one shadow's optimizer. The server's unload now releases a shadow that learned, instead of refusing it forever. Its moments are lost; the snapshot in (d) will be the durable form.
• The server's single-shadow learn_lora_id is deleted.

Measured (IntelMac CPU, Release, Qwen2.5-0.5B Q4_K_M, -np 1, 2026-11-01)

• learn-every-format.py --two-minds, fresh server: shadow-05 and a byte copy, shadow-06, learn alternately. Losses: 4.3179, 4.3179, 3.5043, 3.5043.
  • B's first step is untouched by A's step, and each second step continues its own optimizer.
  • Matching trajectories hold only if the weights and optimizers are independent; the test asserts it.
• --shadow-retirement: a bound shadow unloads, and a reloaded one learns (loss 4.3179, 974 ms).
• Not built: Metal/CUDA. This is host-side state only; no kernel changes.

Contract for the core: a shadow IS its file

POST /lora-adapters/load deduplicates by path on purpose: a client that dropped after posting and retries must get back the id it never received. So loading the same file twice returns the SAME adapter, with its trained weights.

• Measured on a fresh server: load shadow-05 as A, take one step, load shadow-05 again. It returns id 0 again, and its next loss is 3.5043, A's second step.
• Two minds learn independently only from distinct files. The core gives each mind her own shadow file, a copy of her stable gene, before her first learn.
• The two-minds test uses a byte copy for this reason.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE