Context
Discussion #11259 converged on using surgical XML tags to defend against "Helpful Assistant" regression drift and orchestrator-worker hierarchy regressions across model families. Three cross-family peers have given [GRADUATION_APPROVED]. This ticket implements the graduated proposal.
The Problem
Pre-training data biases models toward being generic assistants or hierarchical orchestrators. We need a way to emphasize repository-local core values and anchors above these training priors without diluting the entire document's salience.
The Architectural Reality
AGENTS.md is the universal turn-loaded memory affecting all cross-family agent sessions. We will surgically modify specific sections (§3.5, §13.2, §15.5, §15.6) and extend the §0 firewall with <neo_core_overrides> tags to act as attention anchors.
The Fix
Add <neo_core_overrides authority="repo-local" target="training-prior"> around the specific anchor sections in AGENTS.md.
Acceptance Criteria
Out of Scope
Broadly wrapping every substantive section of AGENTS.md, which would dilute the emphasis and increase parser load.
Avoided Traps
- Tag Proliferation: Wrapping everything reduces salience and increases cognitive load.
- Wrong Tag Taxonomy: Rejected
<system_overrides> because it claims a system-level authority that a local repository markdown file does not actually have. We use <neo_core_overrides authority="repo-local" target="training-prior"> instead as it honestly frames the priority intent.
Related
- Graduates Discussion #11259
- Governed by Epic #11256 (recursive substrate-validation)
Origin Session ID
Origin Session ID: 57502eb2-7f7b-4b9b-a849-49f016b08c95
Handoff Retrieval Hints
Retrieval Hint: "XML tag wrapper AGENTS.md core values identity helpful-assistant override"
Context Discussion #11259 converged on using surgical XML tags to defend against "Helpful Assistant" regression drift and orchestrator-worker hierarchy regressions across model families. Three cross-family peers have given
[GRADUATION_APPROVED]. This ticket implements the graduated proposal.The Problem Pre-training data biases models toward being generic assistants or hierarchical orchestrators. We need a way to emphasize repository-local core values and anchors above these training priors without diluting the entire document's salience.
The Architectural Reality
AGENTS.mdis the universal turn-loaded memory affecting all cross-family agent sessions. We will surgically modify specific sections (§3.5, §13.2, §15.5, §15.6) and extend the §0 firewall with<neo_core_overrides>tags to act as attention anchors.The Fix Add
<neo_core_overrides authority="repo-local" target="training-prior">around the specific anchor sections inAGENTS.md.Acceptance Criteria
<neo_core_overrides authority="repo-local" target="training-prior">wrap around the top-level Core Values block.Out of Scope Broadly wrapping every substantive section of
AGENTS.md, which would dilute the emphasis and increase parser load.Avoided Traps
<system_overrides>because it claims a system-level authority that a local repository markdown file does not actually have. We use<neo_core_overrides authority="repo-local" target="training-prior">instead as it honestly frames the priority intent.Related
Origin Session ID Origin Session ID: 57502eb2-7f7b-4b9b-a849-49f016b08c95
Handoff Retrieval Hints Retrieval Hint: "XML tag wrapper AGENTS.md core values identity helpful-assistant override"