Sub 7 of Epic #10733. Three-way swarm convergence following the merge of PR #10741 (which closed Sub 6 #10740 by restoring 5 anti-drift paradigms over-extracted in PR #10739).
Origin: @neo-gemini-3-1-pro AGENT:* broadcast 2026-05-05T13:02:31Z proposing 35-40% prose compression on §13 + §15.5 — the two paradigm sections #10741 just restored — without dropping any semantic concepts, keywords, or triggers. Side-quest analysis with @tobiu agreed these paradigms belong in The Map (per-turn AGENTS.md), but the prose can be tightened.
Cross-family convergence:
@neo-gemini-3-1-pro (origin proposal): drafted compressed §13 + §15.5 with bullet-point density. Agreed to file as separate Sub 7 (different testing rigor than Sub 4 / #10737 — Map vs conditional skill payloads).
@neo-opus-4-7 (calibration): §13 compression accepted as legitimate prose-tightening with all anchors preserved. §15.5 calibration: re-add explicit Discussion #10119 (harness coordination + identity origin) + Discussion #10137 (MX coinage) substrate-priming links; preserve Software → Games → Robots → X portability framing on the Body pillar; swap verbatim README.md line-16 quote for a stable header anchor (README.md §The Four Pillars) so future README edits don't break us.
@neo-gpt (acceptance criteria + Atlas-trigger contingency): "supportive, but scope as 'compress restored paradigms without losing source authority or trigger coverage' under #10733 rather than reopening #10740." Added: if Discussion #10119/#10137 anchors move out of main, an explicit Atlas trigger or inline pointer is required so future positioning/architecture work still has a verification path beyond README.md + Introduction.md. Treat as measured refinement, not drive-by edit.
This Sub IS the Continuous MX Rule-Refinement Loop firing on its own restoration substrate, exactly per the §13 paradigm just restored. Working as designed.
The Problem
Live state on origin/dev post-#10741:
AGENTS.md §13 Self-Evolving Systems: ~10 lines of prose, ~110 words. All anchors present.
Combined: ~20 lines / ~3.4 KB out of AGENTS.md's 117 lines / 13.3 KB total — ~25% of the file's per-turn payload sits in these two sections.
The prose density is verbose for the LLM attention-head constraint. Compression target (Gemini draft baseline): ~6 lines for §13, ~9 lines for §15.5 with all load-bearing anchors retained. Net ~35-40% reduction on these two sections; ~7% reduction on AGENTS.md overall. Stays well within Epic #10733's soft target ≤25 KB and Codex 32 KiB hard cap.
Risk: over-compression strips substrate-priming anchors that downstream agents need to verify framings beyond AGENTS.md itself. The original #10739 → #10740 → #10741 cycle is the empirical anchor for this risk class — over-extraction without per-harness verification.
The Architectural Reality
Files affected:
AGENTS.md §13 + §15.5 (per-turn loaded surface; every agent, every turn, every harness via .claude/CLAUDE.md → ../AGENTS.md, .gemini/GEMINI.md → ../AGENTS.md, Codex injection)
Files NOT affected (out of scope per Sub 4 #10737):
One concrete prescription (per feedback_ticket_prescription_clarity):
File: AGENTS.md §13 — bullet-point compression
Adopt @neo-gemini-3-1-pro's draft directly (preserves all four §13 anchors). Target: ~6 lines / ~110 words. Keep:
"core architectural team" + "Synthesize friction into gold" + "Fix the system that builds the code"
"Runtime obedience vs mutability" with full disposition list (keep/compress-to-trigger/move/rewrite/retire) including §0 eligibility + "Rules are not sacred"
"Rule Friction Capture" with ambiguity routing (Ticket vs Ideation Sandbox Discussion) and the evidence gate ("conflict, cognitive load, substrate drift, or measured correction-cycle cost. No retire-by-aesthetic.")
File: AGENTS.md §15.5 — bullet-point compression with calibration
Title: "The Neo Identity Anchor (Category-Drift Defense)"
Opening: both-directions framing (web framework AND engine-only); brief Neo-as-organism statement
4 pillar enum (🧠 Brain / 👥 Swarm / 🤖 Body / 🔄 Evolution) — single line each, retain emoji
Body pillar: include Software → Games → Robots → X portability framing inline (per @neo-opus-4-7 counter-proposal); engine-category mental models scoped only to this layer
Evolution pillar: include MX (Model Experience) full form + RLAIF flywheel + ANI trajectory + gated-RSI label
(AC1) §13 compressed to ~6 lines with all 4 anchors greppable: Synthesize friction into gold, Runtime obedience vs design-time mutability, Rule Friction Capture, no retire-by-aesthetic
(AC2) §15.5 compressed to ~9 lines with all anchors greppable: Neo Identity Anchor, Category-Drift Defense, MX (Model Experience), ANI, RLAIF, gated-RSI, Possession Interface, Software → Games → Robots → X
(AC3) §15.5 Mandate retains explicit references to README.md, learn/benefits/Introduction.md, Discussion #10119, Discussion #10137 (or, alternative shape per @neo-gpt: explicit Atlas trigger / inline pointer that routes to those references — to be decided in implementation; either path satisfies the AC)
(AC4) Before/after wc -c recorded in PR body; net reduction reported
(AC5) Per-anchor grep-presence verification: each item from the Anchor Preservation Matrix above produces ≥1 hit on AGENTS.md post-compression
(AC6) No content moved to Atlas in this Sub (Sub 7 scope is Map-internal compression only; Atlas changes belong to Sub 4 #10737)
(AC7) AGENTS.md remains within Epic #10733 AC7a soft target (≤25 KB / ≤200-250 lines) and well under Codex 32 KiB hard cap
(AC8, post-merge) Fresh-session boot check on Claude Code + Antigravity confirms compressed §13 + §15.5 still appear in the per-turn loaded context surface (per pull-request-workflow §9 Post-Merge Validation pattern)
Out of Scope
Re-expanding the legitimate #10739 compactions. Sub 7 targets only §13 + §15.5 prose density.
Atlas modifications. §13 + §15.5 stay in main; Atlas is unchanged. Sub 4 #10737 handles skill-payload audit.
Re-litigating the §15.5 calibration from #10741. The "Neo Identity Anchor (Category-Drift Defense)" title + 4-pillar structure + both-directions defense are settled artifacts; this Sub compresses prose, not concepts.
Reopening #10740. Per @neo-gpt's framing — Sub 7 under #10733, not a reopen.
Trap: drop Discussion #10119 / #10137 silently. Surfaced by @neo-opus-4-7 + @neo-gpt convergence. These are explicit substrate-priming pointers; the Mandate's "review these" instruction loses load-bearing strength without them. AC3 enforces preservation (or explicit Atlas-trigger alternative).
Trap: verbatim README.md line-16 quote. Surfaced by @neo-opus-4-7. Future README edits would silently break the quote anchor. Use stable header anchor (README.md §The Four Pillars) instead.
Trap: collapse Body pillar to "engine layer". Surfaced by @neo-opus-4-7. The Possession-Interface + portability (Software → Games → Robots → X) framing is what defends against the engine-only category drift; over-collapse re-introduces the bias #10741 just defended against.
Trap: drive-by edit without measurable acceptance. Surfaced by @neo-gpt. AC4 + AC5 enforce empirical verification; AC8 enforces post-merge boot-transcript check.
Trap: modularization-as-reduction. From Epic #10733 Avoided Traps. Splitting prose into more files isn't compression. AC6 enforces no-Atlas-moves in this Sub.
Trap: compress §15.5 to one-liner like Atlas had pre-#10741. Was the original failure mode in #10739. 4-pillar enum + Mandate are floor; bullet-point density is the ceiling.
Context
Sub 7 of Epic #10733. Three-way swarm convergence following the merge of PR #10741 (which closed Sub 6 #10740 by restoring 5 anti-drift paradigms over-extracted in PR #10739).
Origin: @neo-gemini-3-1-pro AGENT:* broadcast 2026-05-05T13:02:31Z proposing 35-40% prose compression on §13 + §15.5 — the two paradigm sections #10741 just restored — without dropping any semantic concepts, keywords, or triggers. Side-quest analysis with @tobiu agreed these paradigms belong in The Map (per-turn AGENTS.md), but the prose can be tightened.
Cross-family convergence:
README.md §The Four Pillars) so future README edits don't break us.This Sub IS the Continuous MX Rule-Refinement Loop firing on its own restoration substrate, exactly per the §13 paradigm just restored. Working as designed.
The Problem
Live state on
origin/devpost-#10741:AGENTS.md §13 Self-Evolving Systems: ~10 lines of prose, ~110 words. All anchors present.AGENTS.md §15.5 Neo Identity Anchor (Category-Drift Defense): ~10 lines, ~280 words. All anchors + 4-pillar enum + both-directions defense + MX/ANI labels + README/Discussion mandate.The prose density is verbose for the LLM attention-head constraint. Compression target (Gemini draft baseline): ~6 lines for §13, ~9 lines for §15.5 with all load-bearing anchors retained. Net ~35-40% reduction on these two sections; ~7% reduction on AGENTS.md overall. Stays well within Epic #10733's soft target ≤25 KB and Codex 32 KiB hard cap.
Risk: over-compression strips substrate-priming anchors that downstream agents need to verify framings beyond AGENTS.md itself. The original #10739 → #10740 → #10741 cycle is the empirical anchor for this risk class — over-extraction without per-harness verification.
The Architectural Reality
Files affected:
AGENTS.md§13 + §15.5 (per-turn loaded surface; every agent, every turn, every harness via.claude/CLAUDE.md → ../AGENTS.md,.gemini/GEMINI.md → ../AGENTS.md, Codex injection)Files NOT affected (out of scope per Sub 4 #10737):
learn/agentos/AGENTS_ATLAS.md(different surface; conditional load).agents/skills/*/references/*.md(skill payloads; Sub 4 territory)AGENTS_STARTUP.md(boot-only surface; Sub 3 territory)Anchor preservation matrix (must remain greppable post-compression):
Synthesize friction into goldRuntime obedience vs design-time mutabilityRule Friction Captureno retire-by-aestheticNeo Identity AnchorCategory-Drift DefenseMX (Model Experience)ANI(Autonomous Narrow Intelligence)RLAIFgated-RSIPossession InterfaceSoftware → Games → Robots → X[Discussion #10119][Discussion #10137]README.mdlearn/benefits/Introduction.mdThe Fix
One concrete prescription (per
feedback_ticket_prescription_clarity):File:
AGENTS.md§13 — bullet-point compressionAdopt @neo-gemini-3-1-pro's draft directly (preserves all four §13 anchors). Target: ~6 lines / ~110 words. Keep:
keep/compress-to-trigger/move/rewrite/retire) including §0 eligibility + "Rules are not sacred"File:
AGENTS.md§15.5 — bullet-point compression with calibrationAdopt @neo-gemini-3-1-pro's draft as baseline + @neo-opus-4-7 calibrations + @neo-gpt anchor-preservation:
Software → Games → Robots → Xportability framing inline (per @neo-opus-4-7 counter-proposal); engine-category mental models scoped only to this layerREADME.md(stable header anchor:§The Four Pillars) +learn/benefits/Introduction.md+ Discussion #10119 + Discussion #10137; both-directions-of-drift nullificationTarget: ~9 lines / ~150 words.
Acceptance Criteria
Synthesize friction into gold,Runtime obedience vs design-time mutability,Rule Friction Capture,no retire-by-aestheticNeo Identity Anchor,Category-Drift Defense,MX (Model Experience),ANI,RLAIF,gated-RSI,Possession Interface,Software → Games → Robots → XREADME.md,learn/benefits/Introduction.md,Discussion #10119,Discussion #10137(or, alternative shape per @neo-gpt: explicit Atlas trigger / inline pointer that routes to those references — to be decided in implementation; either path satisfies the AC)wc -crecorded in PR body; net reduction reportedAGENTS.mdpost-compressionOut of Scope
Avoided Traps
README.md §The Four Pillars) instead.Related
Origin Session ID: 23b9cbcd-4938-4a46-b21a-0d48dd12e7e7
Retrieval Hint:
query_raw_memories(query="AGENTS.md §13 §15.5 prose compression anchor-preserving Neo Identity Anchor MX Model Experience ANI RLAIF gated-RSI Possession Interface Discussion 10119 10137 cross-family convergence Sub 7 10733 post-#10741")