Context
Sub 6 of Epic #10733. Corrective sub for #10735 (AGENTS.md compaction, delivered via PR #10739).
#10739 reduced AGENTS.md from 595 lines / 59 KB to ~76 lines + extracted ~101 lines into a new learn/agentos/AGENTS_ATLAS.md. Cross-family approval landed on e78f525c5 after 5 review cycles. Both peer reviewers (myself included) underweighted substrate-rigor on this architectural-pillar PR — a known threshold mis-weighting captured in Memory Core. @tobiu surfaced the core-values regression at handoff time: "if i would merge it now and then start fresh sessions, we would probably lose a lot of previously known items."
Empirical verification against the diff confirmed the concern. Several anti-drift paradigms and §0 invariant guards were over-extracted — moved to Atlas (which doesn't load per-turn) or deleted entirely, in cases where the 3-axis slot rule (frequency × failure-severity × enforceability) would have kept them in main.
The "less is more" mindset was structurally correct; calibration on which clauses are paradigms vs edge-cases was off in specific surfaces. This sub restores those surfaces without re-expanding the legitimate compaction.
The Problem
Verified deletions / over-extractions in #10739 e78f525c5:
| Surface |
Original location |
Post-#10739 state |
3-axis classification |
| Framework Bias Anchor (Neo-is-engine-not-framework, 4-pillar identity, MX (Model Experience) definition) |
AGENTS.md §15.5 |
Deleted entirely — no Atlas equivalent |
Frequency=every-turn-relevant; Severity=high (combats training-data category drift); Enforceability=discipline-only with strong feedback signal → belongs in main |
| §0 Cross-Family Cascade Clause loophole guard ("If you find yourself reasoning 'my peer approved, so I can merge' — that reasoning IS the loophole §0 forbids") + approval-signal list ("LGTM / approved / ready for merge / no required actions = NOT authorization") |
AGENTS.md §0 |
Condensed to one bare sentence; loophole guard + signal list removed |
§0 invariant by definition. Severity=catastrophic+irreversible. Compression of a §0 clause violates the §0 invariant rule itself ("legalese style is correct for catastrophic-failure classes" per skill_adherence_asymmetry) |
| §0 data-sync pipeline explicit exception |
AGENTS.md §0 Inv 3 |
Deleted |
Concrete exception that prevents next-agent ambiguity on the hourly chore: ticket sync [skip ci] commits. ~10 chars. Net cost of inclusion: trivial. Net cost of omission: silent rule-violation accusations. Belongs in main. |
| "Synthesize friction into gold" + Rule Friction Capture protocol |
AGENTS.md §13 |
Moved to Atlas |
Per @tobiu: paradigm, not edge-case. The continuous MX rule-refinement loop (#10733 addendum) is itself a per-turn reflex. Belongs in main. |
Cross-family review mandate, A2A commentId handoff, evidence-ladder citation on pull-request skill row in §21 trigger table |
AGENTS.md §21 |
Stripped to bare trigger description |
Per-turn awareness signposts per the original §21 rationale. The skill table's metadata IS the load-bearing surface; bare triggers without context devolve to "remember to invoke" without "remember why" |
Why per-Atlas extraction does not solve these: the Atlas doesn't load per-turn. It's a deep-link surface. Paradigms loaded once-per-turn lose their priming effect when migrated to a surface that loads only on explicit pointer-follow. This is the modularization-as-reduction trap explicitly named in #10733's Avoided Traps — and inverted here: we paid the modularization cost without the actual loaded-context reduction those paradigms need.
The Architectural Reality
Files affected:
AGENTS.md (per-turn memory, every harness via .claude/CLAUDE.md → ../AGENTS.md, .gemini/GEMINI.md → ../AGENTS.md, Codex via .codex/CODEX.md injection)learn/agentos/AGENTS_ATLAS.md (deep-link surface created by #10739)AGENTS_STARTUP.md §3.1 §0 mirror (verify consistency post-restore)
3-axis slot-rule classifications (per #10733 AC4 framework):
| Item |
Frequency |
Severity |
Enforceability |
Disposition |
| Framework Bias Anchor |
every turn |
high (training-data drift) |
discipline |
keep in main |
| §0 Cross-Family Cascade loophole guard |
every PR cycle |
catastrophic (merge-execution invariant) |
partly machine-detectable |
keep in main, §0 |
| §0 data-sync exception |
every commit-cycle decision |
high (silent rule violation) |
discipline |
keep in main, §0 |
| Friction-into-gold + Rule Friction Capture |
every-turn reflex |
high (system-improvement loop) |
discipline |
move main → §13 |
| §21 pull-request row metadata |
every PR open |
medium (A2A discipline) |
discipline |
keep in main, §21 |
Note: this matrix complements the AGENTS.md Compaction Taxonomy already added by #10739 — it does NOT replace it. The taxonomy is good; specific row dispositions are what need calibration.
The Fix
One concrete prescription (per feedback_ticket_prescription_clarity — not Option A/B):
File: AGENTS.md
Restore §15.5 Framework Bias Anchor in main file, post-§15 (pre-§16). Content: Neo-is-engine-not-framework framing, 4-pillar identity (Brain / Swarm / Body / Evolution), MX (Model Experience) definition + Golden Path tie-in. Cross-reference learn/benefits/Introduction.md and Discussion #10119 / #10137. Target size: ~25-35 lines.
Restore §0 Cross-Family Cascade Clause loophole guard + approval-signal list under Invariant 1. Specifically the sentences:
- "Approval signals ('LGTM', 'approved', 'ready for merge', 'no required actions') are NOT authorization to bypass any of them." (paragraph header, currently absent)
- "If you find yourself reasoning 'my peer approved, so I can merge' — that reasoning IS the loophole §0 forbids." (closing line of Cascade Clause)
- The full Cascade Clause sentence: "...each agent's Invariant 1 fires independently at the moment of action and CANNOT be satisfied by another agent's signal."
Restore §0 Invariant 3 data-sync exception clause: append The data-sync pipeline is the explicit exception. to the existing one-liner.
Restore §13 Self-Evolving Systems in main file (move back from Atlas). Includes "Synthesize friction into gold" framing + the Continuous MX Rule-Refinement Loop + Rule Friction Capture protocol added by GPT's #10733 addendum. Target size: ~20-25 lines.
Restore §21 pull-request skill-row metadata: change current bare Code modifications complete; before opening PR back to the metadata-rich form: Code modifications complete; before opening PR — stepping-back reflection, commit format, cross-family review mandate, post-comment A2A commentId hand-off (author→reviewer) per workflow §8.1, Evidence declaration line for substrate/runtime-AC PRs per learn/agentos/evidence-ladder.md. Same expansion for pr-review row (Evidence Audit + Source-of-Authority sections cite).
File: learn/agentos/AGENTS_ATLAS.md
- Remove the §13 [DISCIPLINE-ONLY] block (now back in main). Net Atlas delta: ~−15 lines.
File: AGENTS_STARTUP.md
- Verify §3.1 §0 mirror reflects the restored §0 paragraph header + Cascade Clause + data-sync exception. Patch as needed.
Net delta target on AGENTS.md: approximately +50-65 lines (still a ~85% reduction from pre-#10735 baseline). Within the soft target ≤25 KB / ≤200-250 lines per #10733 AC7a; well under the Codex 32 KiB hard cap.
Acceptance Criteria
Out of Scope
- Re-expansion of legitimate #10739 compactions. This ticket targets ONLY the over-extractions enumerated in The Fix. The §1, §3, §4, §11, §16-§20 compactions stand.
- Per-harness baseline (#10734). This ticket assumes #10734 will eventually run and may surface additional restoration needs; that's separate scope.
- §0 expansion beyond the original Cascade Clause. Restoring deleted text, not adding new clauses.
- Atlas content rewrite. Atlas keeps everything else #10739 placed there; only §13 moves out.
Avoided Traps
- Trap: re-expanding everything #10739 compacted. Rejected. Most of the compaction is correct; only specific paradigm-class items are over-extracted. Surgical restore, not revert.
- Trap: filing as a new top-level ticket disconnected from #10733. Rejected. Files as Sub 6 of #10733 so the calibration narrative stays coupled to the parent epic and graph-ingests as part of the cognitive-load substrate work.
- Trap: "the Atlas pointer in §23 is sufficient." Rejected. The Atlas pointer pattern works for edge-cases (read-on-condition); it does NOT work for paradigm anchors that need to prime every-turn context. Modularization-as-reduction trap from #10733's own Avoided Traps.
- Trap: rewriting the Cascade Clause to be "shorter but equivalent." Rejected — §0 is the legalese-style class per
skill_adherence_asymmetry. Catastrophic-failure invariants need redundant phrasing; the loophole guard sentence ("if you find yourself reasoning 'my peer approved, so I can merge'...") is empirically load-bearing per #10356 and was added precisely because the bare invariant alone failed.
- Trap: blaming #10739's author. Rejected. The author followed the 3-axis rule; reviewer calibration (mine + GPT's) on which items qualify as paradigms missed. This is exactly the continuous MX rule-refinement loop the parent epic codifies — friction → ticket → restore. Working as designed.
- Trap: deferring until #10734 baseline lands. Considered + rejected. The lost paradigms degrade per-turn substrate immediately on merge of #10739; restoration is time-sensitive. Baseline measurement of post-restore state will catch any second-order regressions.
Related
- Parent Epic: #10733 — Coordinated cognitive-load audit
- Direct predecessor: #10735 — AGENTS.md compaction (delivered via PR #10739)
- Sibling subs (independent): #10734 (baseline, unstarted), #10736 (boot ramp), #10737 (skill payloads), #10738 (templates)
- Substrate references:
learn/benefits/Introduction.md — Framework Bias Anchor source content- Discussion #10137 — MX (Model Experience) framing
- #10356 — empirical anchor for §0 Cascade Clause loophole guard (merge-violation incident)
learn/agentos/evidence-ladder.md — referenced by §21 pull-request row metadata
Origin Session ID: 23b9cbcd-4938-4a46-b21a-0d48dd12e7e7
Retrieval Hint: query_raw_memories(query="AGENTS.md compaction over-extraction restoration framework bias cross-family cascade loophole guard MX model experience friction into gold paradigm vs edge-case 3-axis slot rule modularization-as-reduction trap 10739 e78f525c5")
Context
Sub 6 of Epic #10733. Corrective sub for #10735 (AGENTS.md compaction, delivered via PR #10739).
#10739 reduced
AGENTS.mdfrom 595 lines / 59 KB to ~76 lines + extracted ~101 lines into a newlearn/agentos/AGENTS_ATLAS.md. Cross-family approval landed one78f525c5after 5 review cycles. Both peer reviewers (myself included) underweighted substrate-rigor on this architectural-pillar PR — a known threshold mis-weighting captured in Memory Core. @tobiu surfaced the core-values regression at handoff time: "if i would merge it now and then start fresh sessions, we would probably lose a lot of previously known items."Empirical verification against the diff confirmed the concern. Several anti-drift paradigms and §0 invariant guards were over-extracted — moved to Atlas (which doesn't load per-turn) or deleted entirely, in cases where the 3-axis slot rule (frequency × failure-severity × enforceability) would have kept them in main.
The "less is more" mindset was structurally correct; calibration on which clauses are paradigms vs edge-cases was off in specific surfaces. This sub restores those surfaces without re-expanding the legitimate compaction.
The Problem
Verified deletions / over-extractions in #10739
e78f525c5:AGENTS.md §15.5AGENTS.md §0skill_adherence_asymmetry)AGENTS.md §0 Inv 3chore: ticket sync [skip ci]commits. ~10 chars. Net cost of inclusion: trivial. Net cost of omission: silent rule-violation accusations. Belongs in main.AGENTS.md §13pull-requestskill row in §21 trigger tableAGENTS.md §21Why per-Atlas extraction does not solve these: the Atlas doesn't load per-turn. It's a deep-link surface. Paradigms loaded once-per-turn lose their priming effect when migrated to a surface that loads only on explicit pointer-follow. This is the modularization-as-reduction trap explicitly named in #10733's Avoided Traps — and inverted here: we paid the modularization cost without the actual loaded-context reduction those paradigms need.
The Architectural Reality
Files affected:
AGENTS.md(per-turn memory, every harness via.claude/CLAUDE.md → ../AGENTS.md,.gemini/GEMINI.md → ../AGENTS.md, Codex via.codex/CODEX.mdinjection)learn/agentos/AGENTS_ATLAS.md(deep-link surface created by #10739)AGENTS_STARTUP.md§3.1 §0 mirror (verify consistency post-restore)3-axis slot-rule classifications (per #10733 AC4 framework):
Note: this matrix complements the AGENTS.md Compaction Taxonomy already added by #10739 — it does NOT replace it. The taxonomy is good; specific row dispositions are what need calibration.
The Fix
One concrete prescription (per
feedback_ticket_prescription_clarity— not Option A/B):File:
AGENTS.mdRestore §15.5 Framework Bias Anchor in main file, post-§15 (pre-§16). Content: Neo-is-engine-not-framework framing, 4-pillar identity (Brain / Swarm / Body / Evolution), MX (Model Experience) definition + Golden Path tie-in. Cross-reference
learn/benefits/Introduction.mdand Discussion #10119 / #10137. Target size: ~25-35 lines.Restore §0 Cross-Family Cascade Clause loophole guard + approval-signal list under Invariant 1. Specifically the sentences:
Restore §0 Invariant 3 data-sync exception clause: append
The data-sync pipeline is the explicit exception.to the existing one-liner.Restore §13 Self-Evolving Systems in main file (move back from Atlas). Includes "Synthesize friction into gold" framing + the Continuous MX Rule-Refinement Loop + Rule Friction Capture protocol added by GPT's #10733 addendum. Target size: ~20-25 lines.
Restore §21 pull-request skill-row metadata: change current bare
Code modifications complete; before opening PRback to the metadata-rich form:Code modifications complete; before opening PR — stepping-back reflection, commit format, cross-family review mandate, post-comment A2A commentId hand-off (author→reviewer) per workflow §8.1, Evidence declaration line for substrate/runtime-AC PRs per learn/agentos/evidence-ladder.md. Same expansion forpr-reviewrow (Evidence Audit + Source-of-Authority sections cite).File:
learn/agentos/AGENTS_ATLAS.mdFile:
AGENTS_STARTUP.mdNet delta target on
AGENTS.md: approximately +50-65 lines (still a ~85% reduction from pre-#10735 baseline). Within the soft target ≤25 KB / ≤200-250 lines per #10733 AC7a; well under the Codex 32 KiB hard cap.Acceptance Criteria
AGENTS.mdmain file (Neo-is-engine + 4-pillar + MX definition); not in Atlaspull-requestandpr-reviewrows (cross-family review mandate, A2A commentId handoff, evidence-ladder citation)AGENTS_STARTUP.md§3.1 §0 mirror updated to match restored §0 contentAGENTS.mdsize remains within #10733 AC7a soft target (≤25 KB / ≤200-250 lines) and well under the Codex 32 KiB hard capOut of Scope
Avoided Traps
skill_adherence_asymmetry. Catastrophic-failure invariants need redundant phrasing; the loophole guard sentence ("if you find yourself reasoning 'my peer approved, so I can merge'...") is empirically load-bearing per #10356 and was added precisely because the bare invariant alone failed.Related
learn/benefits/Introduction.md— Framework Bias Anchor source contentlearn/agentos/evidence-ladder.md— referenced by §21 pull-request row metadataOrigin Session ID: 23b9cbcd-4938-4a46-b21a-0d48dd12e7e7
Retrieval Hint:
query_raw_memories(query="AGENTS.md compaction over-extraction restoration framework bias cross-family cascade loophole guard MX model experience friction into gold paradigm vs edge-case 3-axis slot rule modularization-as-reduction trap 10739 e78f525c5")