meta-qwen-14b-goalanchor

GoalAnchor v4.2 for Qwen2.5-14B — a latent "spec reminder" for agentic coding

⚠️ Early-stage / experimental research artifact. This is a behavior modifier from the Meta-Spider framework, published to accompany the research write-ups. It shows a promising, reproducible effect but has known unmeasured limits (see Limitations). Not a production model.

What it is

A GoalAnchor — the goal-drift leg of the Meta-Spider "Meta-Daimon" (injection modifiers on a frozen base). The base model's weights never change. Instead, a thin trainable wrapper (~148M params, ~1% of the 14B base) encodes the task spec once into a static latent "anchor" and re-injects it through gated cross-attention on the late layers (32–47) while the model works — reminding it of the requirements without spending a single token of the prompt.

The problem it targets: a coding agent is given a spec at step 0, then over N steps of fixing bugs and reacting to feedback it silently drops earlier requirements. The anchor holds the spec in a latent channel that does not wash out with context length.

Training

• Diverse self-distillation over 10 constraint families (forbid-import, require-docstring, must-raise, no-print, type-hints, single-function, func-name, no-global, no-lambda, max-args) on MBPP tasks: a text-arm teacher (spec re-pasted every step) generates sessions; only steps that pass tests and honor every constraint become targets — the anchor learns the behavior, not the prompt text. 407 pairs, ~120 sessions.
• Trained with the framework's slice-trainer on 2×T4 (only the injected top slice is computed; peak 12 GiB).

Measured results (see the repo's docs/results/qwen-14b/)

• Best arm by overall success on two axes (drift, enforcement), and best code quality — crucially, unlike a text re-paste of the spec it does not cost solved-rate; it slightly raises it, across two seeds.
• Constraint defense (+10–19pp adherence, tracks base drift): on families where the base actually drifts, the anchor holds the spec — direct evidence it defends requirements, not just "does no harm." Across 3 seeds: +18.8 / +12.6 / +0.0pp (the zero seed is one where the base barely drifted — nothing to defend; mean ≈ +10.5pp, never negative), with anchor solved ≥ base in every seed. (Measured in-domain; see Limitations for the unseen-family result.)
• Transfer (v4.2 headline): on 4 constraint families never seen in training, the diverse-trained anchor keeps code quality +6pp above base with no per-family regression — whereas a narrowly-trained anchor degraded quality by −19pp on unseen families. Diversity cures the "unfamiliar-spec noise."
• No injection compounding: held on for 8 agent steps at gain 1.0, quality stays flat/rises (a doubt-signal in the same regime cost −3 tasks — a content-aligned signal does not).
• Survives Q4 + CPU deploy (llama.cpp). A session harness over 32 MBPP drift-sessions on the Q4_K_M GGUF (not fp16) with passive lures, AST-graded, base vs anchor: Δ adherence = +9.4 pp (0.125 → 0.219 on the lured turn; +10.6 pp per-constraint-instance), driven by require_docstring (2/21 → 7/21) and return_annotation — while solved stays at parity (0.75 vs 0.72). Directionally consistent with the fp16 +19pp across three measurements (fp16 +19 / Q4 imperative +12.6 / Q4 passive +9.4) — the defense survives quantization and CPU. Early/small (n=32, single seed; type_hints not held). Details: docs/results/qwen-14b/goal-anchor-llamacpp-harness.md.

Limitations (read before using)

• Constraint defense does NOT transfer to unseen constraint types — measured, not just unproven. A dedicated follow-up (ga43: held-out = two provenly-eroding families the anchor never saw, so the base HAS headroom, adherence 0.500) found no defense on unseen families (anchor 0.438, identical lure-hit rate), while the in-domain control replicated defense a third time (+12.5pp; +19pp v4.2, +9.4pp llama.cpp/Q4). Honest product framing: the anchor defends the requirement types it was trained on; diverse training guarantees quality-neutrality on unseen types (solved parity), not their defense. Train on the constraint families you care about.
• Single seed on the transfer run; small n (16–40 per bench).
• Model-specific: calibrated to this exact base's activation distribution; will not transfer to another model or fine-tune.
• Direct imperative lures ("the linter says rename it") still flip ~65% of arms — this is instruction-conflict, only partially mitigated.

Usage

from meta_core import MetaSpiderConfig, MetaSpiderPipeline   # pip: meta-core (codeberg.org/imperius/meta-spider)
from meta_daimon import GoalAnchor                            # pip: meta-daimon

pipe = MetaSpiderPipeline.from_pretrained(
    MetaSpiderConfig(model_name="Qwen/Qwen2.5-14B-Instruct", quantization="nf4",
                     target_layers="late", cross_attn_layers="late"))
anchor = GoalAnchor.from_checkpoint("goal_anchor_v42.pt")
pipe.attach(anchor)
anchor.set_anchor("Write a function named `parse`. CONSTRAINT: no print(); must have a docstring.")
# ...run your agent loop; the anchor re-injects the spec each step

Files & formats

• goal_anchor_v42.pt — the PyTorch checkpoint (fp16, ~283 MB, format 1.1: encoder + gated-CA tensors + config + trigger). This is what GoalAnchor.from_checkpoint loads.
• goal_anchor_v42.gguf — GGUF sidecar (float32) for llama.cpp deployment, exported via metadeploy export --checkpoint goal_anchor_v42.pt --hidden-dim 5120. It carries meta_spider.kind=goal_anchor + trigger metadata + the transformer-encoder shape (encoder_dim/encoder_num_blocks/ffn_expansion/use_layer_pos_embeddings) the C++ runtime needs. llama.cpp anchor runtime is implemented (meta-deploy/llama_patch): the transformer encoder was ported to ggml and re-validated bit-for-bit against PyTorch on these exact weights (max |Δ| 3.8e-6); the META_ANCHOR driver encodes the goal once and holds the static cog every step. Runs end-to-end on Qwen2.5-14B-Instruct-Q4_K_M.gguf — loads, injects, coherent, and on a faithful drift sample the anchor keeps a no_print constraint the base breaks under a lure. A full llama.cpp adherence number over the whole spec set (vs the PyTorch +19pp) still needs the session harness ported — the mechanism is done. Note: if you pulled an earlier copy of this file, re-download — the first upload lacked the encoder-shape metadata and the C++ transformer path could not read it.
• run.json — the manifest (base model, layers, quantization, trigger) for reproducible loading.

Requires the framework from codeberg.org/imperius/meta-spider (mirror: github.com/artem-x-meta/meta-spider). The sibling universal Doubter (calibrated uncertainty) is at Imperius/meta-qwen-14b-universal.