Research-Stack/REBASE_RULES.md

14 KiB
Raw Blame History

REBASE RULES — Research Stack Agent / Rules Manifest

Purpose: This file is a rebase seed. When starting a fresh project map, copy only the files and rules listed here. Do not migrate runtime state, stale audits, or completed assignment boards.

Principle: Port the contracts and the rituals, not the state.


1. Core Operating Contracts (MUST PORT)

These files are the source of truth for every agent, skill, and editor hook. They are actively maintained and cross-referenced throughout the stack.

File Scope
AGENTS.md (root) Master operating contract: post-interaction workflow, programming-choice flow, Q16_16 doctrine, Git hygiene, Do-Not-Sweep rules, glossary, claim boundaries.
0-Core-Formalism/lean/Semantics/AGENTS.md Lean/Semantics local contract: blessed Compiler surface, build baselines, quarantine boundaries, proof-tactic lessons, current solidification anchors.
4-Infrastructure/AGENTS.md Infrastructure and shim contract: storage stack, hardware-witness boundaries, ENE schema, compute dispatch pattern.
6-Documentation/docs/AGENTS.md Strict LLM operating rules: anti-drift evidence standards, bind primitive, naming conventions, sigma levels, SI units, deletion criteria, PQC policy.
5-Applications/text-to-cad/AGENTS.md CAD harness contract: generated CAD/URDF source-of-truth, viewer handoff, repo policies.
scripts/qc-flag/AGENTS.md Lean QC protocol: 7-check inspection protocol.
shared-data/artifacts/lean_expert_agent/AGENTS.md Lean expert agent /inspect protocol and output format.
.github/RRC_OPERATING_CONTRACT.md Rainbow Raccoon Compiler gate definition: ACCEPT / HOLD / QUARANTINE.

Non-negotiable rules extracted from the contracts

  1. Lean is the source of truth. Python, Rust, and Verilog are extraction targets only.
  2. Programming choice flow — before writing new logic, check this decision tree in order:
    • Admissibility / routing / alignment / gating decision? → Lean only.
    • Emitting a top-level receipt or JSON bundle? → Semantics.AVMIsa.Emit only.
    • Classifying rows / computing alignment scores? → Lean (Semantics.RRC.*).
    • Supplying raw input features (equation text, IDs, weak_axes)? → Python shim OK, no admissibility logic, regenerable from source.
    • Float arithmetic in a compute path? → STOP. Use Q16_16.ofNat, Q16_16.ofRatio, or Q16_16.ofRawInt.
    • Advancing promotion status in shim space? → STOP. not_promoted until a Lean gate passes.
    • Pure I/O (read/write JSON, subprocess, format output)? → Python shim OK. Receipt output routes through AVMIsa.Emit.
  3. No Float in compute paths. ofFloat is permitted only at JSON/sensor boundaries and must be immediately bracketed.
  4. Post-interaction workflow (mandatory after any code/receipt/architecture change):
    • Update the nearest scoped AGENTS.md for every touched subtree.
    • Run lake build Compiler (and full lake build if Lean files touched).
    • Commit with an explicit file list; never git add ..
    • Run git status --branch --short --untracked-files=all.
  5. Do Not Sweep: never run git add ., git checkout -- ., or git clean -fdx.
  6. Secrets are runtime-only: use environment variables (OLLAMA_API_KEY, DEEPSEEK_API_KEY, etc.); never commit literal keys.
  7. Legacy recovery is explicit: use RECOVER LEGACY INFORMATION: <path/commit/concept> only. Current cornfield ref: backup/distilled-with-vcd-history-2026-05-11.

2. Editor / Agent Configs (MUST PORT)

These files inject the contracts into editor/agent runtimes.

File Purpose
.cursorrules Cursor: ContextStream init()/context()/search() rules.
.devin/config.json Devin: programming-choice reminder hook.
.devin/hooks.json Devin: ContextStream hooks.
.devin/hooks/programming-choice-reminder.sh Injects programming-choice flow + post-interaction checklist before every user message.
.claude/settings.json Claude Code: ContextStream hooks.
.cursor/hooks.json Cursor: ContextStream hooks.
.roo/mcp.json Roo: MCP server definitions.
.mcp.json (root) Minimal MCP config: filesystem, git, fetch, contextstream, lean, remote-lean-proof, sympy, wolfram, aws, github, headroom.
.contextstream/config.json ContextStream workspace/project IDs.
.kilo/rules/contextstream.md Kilo Code manual ContextStream rules (no hooks available).
.roo/rules/contextstream.md Roo Code ContextStream rules.
.windsurf/rules/contextstream.md Windsurf ContextStream rules.

ContextStream-first workflow (enforced or self-enforced in every editor)

  • First message: init(folder_path=...)context(user_message=...).
  • Subsequent messages: context(user_message=...) FIRST.
  • Before any file search: search(mode="auto") BEFORE Glob/Grep/Read/Explore/Task/EnterPlanMode.
  • After edits: project(action="index") in editors without auto-indexing.
  • Save transcripts: save_exchange=true in editors without auto-saving.

3. Skills (MUST PORT)

Skill Path Purpose
lean-proof .devin/skills/lean-proof/SKILL.md Mandatory for all Lean work: proof quality, Q16_16 contract, #eval witnesses, promotion gate.
contextstream-workflow .github/skills/contextstream-workflow/SKILL.md Persistent memory: plans, tasks, decisions, lessons.
research-stack-workflow .github/skills/research-stack-workflow/SKILL.md Token-efficient proof workflow; local inference first on neon.
headroom .opencode/skills/headroom/SKILL.md Context compression and cross-agent memory.
cad 5-Applications/text-to-cad/skills/cad/SKILL.md Script-driven STEP/STL/DXF/GLB generation.
urdf 5-Applications/text-to-cad/skills/urdf/SKILL.md URDF generation and validation.
rrc-compile .claude/skills/rrc-compile/SKILL.md RRC pipeline orchestrator.
rrc-scale-analysis .claude/skills/rrc-scale-analysis/SKILL.md RRC pre-flight Q16_16 scale-band analysis.

4. MCP Servers (MUST PORT)

From .mcp.json / .roo/mcp.json:

Server Role
contextstream Persistent memory / context / search
filesystem Read/write proof artifacts and receipts
git History, blame, diff
fetch HTTP docs and RFCs
lean Lean 4 / Mathlib typecheck bridge
remote-lean-proof Hermes / OpenCode / Codex proof backend
sympy Symbolic arithmetic verification
wolfram-alpha Formula verification (needs WOLFRAM_APP_ID)
aws AWS API access
github PRs, issues, CI
headroom Token / context compression

5. Conditional Items (PORT ONLY IF DOMAIN REMAINS ACTIVE)

File Condition
.windsurf/lifeguard.yaml Keep only if genetics / GCCL / Mass Number abstraction work continues. Contains safety doctrine for biological-equivalence claims, reverse-collapse, bounded recursion, and claim-state hygiene.
.opencode/agents/*.md Keep only if the target sorry/theorem is still open in the new rebase. These are task-specific subagent prompts.
.claude/workflows/ground-disproved-theorems.js Keep only if Claude Code still uses this workflow.

6. DROP — Do Not Port

Item Reason
.consolidation-manifests/ Historical SHA256 manifests from 2026-05-04. Not load-bearing.
.headroom/*.db Runtime memory databases; regenerate on first use.
.agents/plugins/marketplace.json References a dormant ./plugins/substack-connector. Not load-bearing.
.github/copilot-instructions.md Core rules are valid, but the "Current Project State" block is stale (2026-05-28). Port rules separately, not this file as authority.
.windsurf/SORRY_AUDIT.md Stale audit (2026-05-28). Root AGENTS.md now reports 0 sorries in the active Compiler surface. Archive only.
.windsurf/ASSIGNMENTS.md Dated 2026-04-19; all assignments marked COMPLETE and references old paths. Historical residue.
Staging manifests (stack_solidification_staging_manifest_2026-05-09.md, etc.) Historical archives; referenced but not authoritative.
Cornfield / legacy branches Do not merge or base work on them. Use only via explicit recovery trigger.

7. Minimal Rebase Checklist

Use this when standing up the new project map.

  • Copy all files in Section 1 (Core Operating Contracts).
  • Copy all files in Section 2 (Editor / Agent Configs).
  • Copy all files in Section 3 (Skills).
  • Copy/adapt .mcp.json and .roo/mcp.json with current server paths.
  • Update .contextstream/config.json with new workspace/project IDs if they change.
  • Review Section 5 conditional items and keep only active-domain rules.
  • Verify Section 6 dropped items are not carried over.
  • Run lake build Compiler and confirm job count matches the new baseline.
  • Update build baseline and commit hash in 0-Core-Formalism/lean/Semantics/AGENTS.md.

8. Design Note for the New Map

When designing the new project layout, preserve these boundary rules:

  • Lean owns decisions; Python owns I/O.
  • Semantics.AVMIsa.Emit is the sole top-level receipt boundary.
  • shared-data/ is ignored by default. Promote only durable receipts with git add -f -- <path>.
  • Research artifacts go outside the git tree (e.g. /home/researcher/research/ in the container model).
  • Every new compression path needs two Lean theorems: eigensolid_convergence and receipt_invertible.


9. Cornfield Recovery Manifest (SilverSight Port)

What is the cornfield? Concepts developed during Research Stack sessions that were archived, quarantined, or left as probes — not in the active Lean Compiler surface, but recoverable. The canonical record lives in extraction/cornfield_concepts.json. The databases are authoritative.

Database state (2026-06-21):

  • arxiv.math_objects WHERE id LIKE 'cf_%': 19 rows
  • ene.packages WHERE package_type = 'cornfield_concept': 19 packages
  • ene.relations (cornfield edges): 23 relates_to edges
  • Gremlin concepts graph, cohort=cornfield: 19 vertices, 38 edges

To add a new cornfield concept: append to extraction/cornfield_concepts.json, run uv run scripts/load_cornfield.py.

9a. Port — Carry to SilverSight

ID Concept Action Priority
cf_phi_universal Φ = Σ wᵢ/ln(Nᵢ) — universal field equation promote_to_foundation P0 CRITICAL
cf_codebase_memory FAMM 7-domain bank + thermal + scar + dual-map port_as_stub HIGH — skeleton for SilverSight memory model
cf_codebase_fsdu ΔS = S_ahead S_behind scar differential gate port_as_stub HIGH — discrete form of NKHodgeFAMM scar
cf_codebase_receipt Receipt-gated memory access port_as_stub HIGH — completes AVMIsa.Emit → Hermes chain
cf_hachimoji_shell N=2^m base alphabet, k=⌊n^(1/m)⌋ shell promote_to_active HIGH — physics grounding for SpectralSignature
cf_w_axis_omega Ordinal/cost/bridge pressures for AngrySphinx formalize_in_lean HIGH — FORMALIZATION_DRAFT, missing from synthesis
cf_ms3c_nested_gear 4-weight balanced shear scoring {0.25×4} extend_s3c_shell MEDIUM — softer gate than binary Q0_2 threshold
cf_gap_space_padic p-adic ultrametric generalizes Sidon gap extract_padic_gap MEDIUM — unifies CrossDomainOneOverN family
cf_padic_calculus p-adic integration, Haar measure extend_cross_domain_1_over_n MEDIUM
cf_eit_invariant PHI = non-Euclidean correction to Landauer link_phi_to_landauer MEDIUM — one import away
cf_coherence_kernel R_O = 1/7 invariant formalize_as_constant MEDIUM — 1/7 in two independent contexts
cf_info_thermo Landauer + PHI_NE correction merge_into_landauer MEDIUM
cf_landauer_genetic_clock E_gen = G·ε·E_r; T_clock formula merge_into_landauer MEDIUM — bridges Hachimoji to HutterPrize
cf_genus3_topology β₂(Σ₃)=1 → NKHodgeFAMM fails link_to_nk_hodge MEDIUM — explicit counterexample class
cf_adiabatic_limit g_ε=g_B/ε²+g_F, ε→0 collapse link_to_nk_hodge MEDIUM — formal justification for Q0_2 threshold
cf_morphic_dsp Morphic scalar field + neural network bridge extract_missing_formulations LOW — active module exists, probe has more
cf_rrf_ranking rrf_score = Σ 1/(k+rᵢ), k=60 add_to_graph_rank LOW — already used in shim layer
cf_semantic_basin_overflow 100×bandwidth × 50M×compression = 5×10⁹ link_to_phantom LOW — context for phantom coupling scale

9b. Drop — Do Not Port

ID Reason
cf_snap_citations INFRA GAP, not a concept — ingest cit-HepPh/HepTh to arxiv.citations before port (2 COPY commands from neon-64gb:/home/allaun/snap_data/)

9c. Cornfield Query Patterns

-- All cornfield concepts by priority
SELECT pkg, archetype, provenance->>'lean_status'
FROM ene.packages WHERE package_type='cornfield_concept'
ORDER BY archetype;

-- Cornfield → active module connections (via Gremlin)
-- g.V().has('cohort','cornfield').out('depends_on').values('id')

-- Concepts needing Lean formalization
SELECT id, name FROM arxiv.math_objects
WHERE id LIKE 'cf_%' AND lean_status IN ('ARCHIVED','UNDOCUMENTED','FORMALIZATION_DRAFT');

9d. Invariant for the new map

Every new cornfield concept recovered into SilverSight must satisfy:

  • Either imports an existing active Lean module (inherits its proof constraints), OR
  • Has a silversight_action of formalize_in_lean with a named stub theorem in the build baseline.

No floating concepts. If it's not reachable from Semantics.AVMIsa.Emit via imports or depends_on, it is not on the map.


Generated from active Research Stack agent rules and contracts. Principle: port contracts and rituals, not state. Cornfield manifest last updated: 2026-06-21 — 19 concepts.