Research-Stack/AGENTS.md

AGENTS.md - Research Stack Operating Contract

This file is the first stop for coding agents working in this repository.

Ground Rules

• Use /home/allaun/Research Stack as the active host checkout and git root. When the research-stack dev container is available, run Lean builds, WGSL shader work, and research execution through podman exec research-stack from the container-local checkout at /home/researcher/repo/. Treat the host checkout and container checkout as synchronized views of the same repo; verify paths before copying artifacts between them.
• The dev container is research-stack (podman, research-stack-otom:latest).
• Research artifacts (design docs, experimental Lean files, exploration output) must be written to container-local paths under /home/researcher/research/. These are not in the git tree and do not pollute the working tree. Use podman cp to extract research artifacts when they are ready for promotion to production.
• Read the nearest nested AGENTS.md before editing a subtree.
• Preserve user work. The working tree is often intentionally dirty; do not revert, delete, or stage unrelated files.
• Prefer repo-native tools and receipt generators over ad hoc summaries.
• Treat Lean as the source of truth for formal or hardware-adjacent claims.
• Load the lean-proof skill before any Lean work. It enforces the proof quality contract: no bare sorries, no tautologies, Q16_16 compliance, and #eval witness requirements. Trigger patterns auto-load it on .lean, lake build, sorry, Q16_16, theorem, lemma, etc.
• No Float in compute paths. ofFloat is only permitted at the external boundary (JSON parsing, sensor data). All core/main computation uses Q16_16.ofNat, Q16_16.ofRatio, or Q16_16.ofRawInt. This applies to Lean, Python, and Verilog. The HiGHS boundary is the only exception (it requires float inputs, but the result is immediately converted to Q16_16).
• Keep claims bounded: a receipt proves only the gate it actually checks.
• Secrets are runtime-only. Use environment variables such as OLLAMA_API_KEY or DEEPSEEK_API_KEY; never paste, print, or commit literal provider keys.
• For repo-stabilization tasks, finish with a clean git status --branch --short --untracked-files=all and an empty git clean -nd dry run before claiming the tree is stable.

Core Surfaces

• Lean/Semantics: 0-Core-Formalism/lean/Semantics/
• Infrastructure shims and probes: 4-Infrastructure/shim/
• Hardware bring-up: 4-Infrastructure/hardware/
• Documentation and wiki surfaces: 6-Documentation/
• Citation reference map: CITATION.cff (external sources are provenance and terminology references unless a Lean theorem or receipt explicitly promotes a bounded claim)
• Stack receipts: shared-data/data/stack_solidification/
• Promoted review receipts: shared-data/artifacts/deepseek_review/
• Canonical Ollama/DeepSeek review emitter: 5-Applications/tools-scripts/llm/ollama_deepseek_review_emitter.py
• CAD harness: 5-Applications/text-to-cad/
• Historical scoped staging maps: 6-Documentation/docs/stack_solidification_staging_manifest_2026-05-09.md and 6-Documentation/docs/stack_solidification_staging_manifest_2026-05-10.md

Verification Expectations

• For Lean changes, run the narrow target first, then the broader lake build when feasible.
• For Python shims, run python3 -m py_compile on touched files.
• For JSON receipts, run python3 -m json.tool or a repo-native receipt parser.
• For promoted Ollama/DeepSeek review receipts, run python3 5-Applications/tools-scripts/llm/ollama_deepseek_review_emitter.py --verify-only so answer_sha256 is checked against the answer file after write.
• For hardware claims, distinguish software witness, bitstream presence, SRAM load, flash persistence, UART beacon, and live hardware receipt.
• Before committing, run git diff --cached --check and a staged secret scan for touched source/receipt files. The repository credential hook is a final gate, not a substitute for local review.
• For root CAD setup changes, JSON-parse package.json, .vscode/settings.json, and .vscode/tasks.json, then use the pinned Python/CAD commands documented in 5-Applications/text-to-cad/AGENTS.md.

Post-Interaction Workflow (mandatory after every agent session)

After every interaction with the user that changes code, Lean modules, shim scripts, receipts, or architecture decisions, the agent MUST run the following steps before claiming the session is complete:

1. Update AGENTS.md files

Update the nearest scoped AGENTS.md for every subtree touched:

Subtree touched	AGENTS.md to update
0-Core-Formalism/lean/Semantics/	0-Core-Formalism/lean/Semantics/AGENTS.md
4-Infrastructure/	4-Infrastructure/AGENTS.md
Root or multi-subtree	AGENTS.md (this file)

Minimum required updates per touched subtree:

• Blessed surface table — add/remove/update module rows if Compiler roots changed
• Build baseline — update job count and commit hash if lake build was run
• Architecture section — update if data-flow or boundary rules changed
• Pending proof work — add new TODO(lean-port) stubs; remove resolved ones
• Quarantine table — add quarantined files; remove revived ones

2. Verify the build

cd 0-Core-Formalism/lean/Semantics && lake build Compiler

If Lean files were touched, also run the full build:

cd 0-Core-Formalism/lean/Semantics && lake build

3. Commit

Stage only explicitly touched files (never git add .). Commit message format:

<type>(<scope>): <summary>

<body — what changed and why>

Build: <N> jobs, 0 errors (lake build)

Types: feat, fix, chore, docs, refactor. Scope examples: lean, rrc, avm-isa, infra, corpus278.

4. Check tree cleanliness

git status --branch --short --untracked-files=all

Untracked files that are not generated artifacts should either be staged or noted as intentionally dirty. Do not claim the tree is stable if there are unexpected modifications.

5. Programming choice flow (check before writing any new code)

Before writing or placing any new logic, run through this decision tree in order. Stop at the first rule that applies.

New logic needed?
│
├── Does it make an admissibility, routing, alignment, or gating decision?
│   └── YES → Write it in Lean. No Python equivalent allowed.
│             File: Semantics/RRC/Emit.lean or a new Semantics.* module.
│
├── Does it mint, stamp, or emit a top-level receipt or JSON bundle?
│   └── YES → It belongs in Semantics.AVMIsa.Emit ONLY.
│             AVMIsa.Emit is the sole output boundary. No other module
│             may call leanBuildReceipt and produce a top-level JSON.
│
├── Does it classify rows, run an alignment gate, or compute scores?
│   └── YES → Lean (Semantics.RRC.Emit or a new Semantics.RRC.* module).
│             Python may call it via #eval / lake exe but may not replicate it.
│
├── Does it supply raw input features (equation text, route_hint, domain_type,
│   equation_id hashing, weak_axes count)?
│   └── YES → Python shim is acceptable. The shim must:
│             (a) produce only a List FixtureRow (or equivalent data structure)
│             (b) carry no admissibility logic — every field it sets is "raw"
│             (c) be regenerable from source (document the regeneration command)
│             (d) live in 4-Infrastructure/shim/ with a clear TODO(lean-port) if
│                 any logic in it could eventually move to Lean
│
├── Does it use floating-point arithmetic in a compute path?
│   └── YES → STOP. Use Q16_16.ofNat / Q16_16.ofRatio / Q16_16.ofInt instead.
│             ofFloat is only permitted at the external boundary (JSON parsing,
│             sensor input) and must be immediately bracketed.
│
├── Does it advance promotion status (e.g. set promotion = "promoted")?
│   └── YES → STOP. Promotion is always not_promoted until a Lean gate
│             explicitly passes. Never advance it in shim space or by hand.
│
└── Is it pure I/O (read JSON, write JSONL, call subprocess, format output)?
    └── YES → Python shim is fine. Keep it in 4-Infrastructure/shim/.
              Receipt-writing Python must still route output through
              AVMIsa.Emit (Lean stamps; Python only formats/stores).

Summary rule: Lean owns all decisions. Python owns all I/O. If you find yourself writing decision logic in Python, stop and port it to Lean.

What "every interaction" means

This workflow triggers whenever the user message results in:

• Any file edit (Lean, Python, TOML, JSON, shell, Rust, etc.)
• Any lake build run
• Any architectural decision that changes the data-flow or boundary rules
• Any new TODO(lean-port) or quarantine boundary

It does NOT trigger for pure read-only exploration, explanation, or questions that result in no file changes.

Do Not Sweep

Avoid broad cleanup or staging commands such as:

git add .
git add 0-Core-Formalism 4-Infrastructure 6-Documentation shared-data
git checkout -- .
git clean -fdx

Use explicit file lists from the relevant staging manifest.

shared-data/ is ignored by default because most of it is generated or offloaded. Promote only specific, durable receipts with git add -f -- <path>, and keep empty/failed model outputs out of Git unless they are themselves the evidence under review.

Git Remote Hygiene

• The active branch may not have an upstream. Inspect with git rev-parse --abbrev-ref --symbolic-full-name @{u} before assuming push state.
• For GitHub sync, prefer the github remote and verify the remote head after push:

git fetch github <branch>
git rev-list --left-right --count FETCH_HEAD...HEAD
git push -u github <branch>
git ls-remote --heads github <branch>

• Dependabot banners printed by GitHub after push may be stale relative to the live alert API. Treat the push result and remote-head hash separately from dependency-alert remediation.

Glossary (before reading further)

These terms appear throughout all AGENTS.md files and the codebase:

• Sidon label — an address from a set where all pairwise sums are unique. Powers of 2 (1,2,4,8,16,32,64,128) are the canonical Sidon set for 8 strands. Sidon slack = address budget − max label used (encodes capacity headroom).
• BraidStorm — the 8-strand braid topology used by the eigensolid compressor. Strands cross pairwise; each crossing merges phase and produces a residual.
• eigensolid — the converged, stable state of a braid crossing loop. Detected when crossStep(s) = s. The DC baseline in the TNT BraidCarrier model.
• scar — a FAMM failure record in ene.scars. Stores scar_pressure, failure_mode, and optional coarsening_agent for remediation path. Scar absence (∅) is a positive receipt dimension.
• Yang-Baxter — the braid relation βij βjk βij = βjk βij βjk that defines braid-order invariance. Represented in BraidedFieldPaths.lean; operational braid-action proofs must state their exact evaluation model.
• Anti-BraidStorm — adversarial dual that tests Yang-Baxter invariance and receipt aliasing as a validation layer for the compressor.
• MORE FAMM — Memory-Optimized Recursive Entropy Fractal Aggregate Memory Model. Capability-based memory isolation for the runtime stack.
• TSM — Topological S3C Manifold. Thermodynamic safety monitor with Builder/Warden/Judge clock domains.
• GCL — Genetic Code Language. Self-improving evolutionary program representation.
• AVM — Adaptive Virtual Machine. Universal bridge between math languages and Python bytecode. Core ISA is live in Semantics.AVMIsa.* (Types, Value, Instr, State, Step, Run). AVM is the sole output boundary for RRC receipts: AVMIsa.Emit stamps all top-level receipt JSON; RRC.Emit and RRC.Corpus278 feed it as classifier and raw-feature supplier respectively.
• enwik9 — the Hutter Prize 1GB Wikipedia XML corpus, used as the canonical end-to-end test vector for the hierarchical compressor.
• receipt — a machine-readable attestation record stored in ene.receipts. Receipt dimensions (C, σ, k, ε_seq, t, ∅_scars) together form the encoding of the compressed state. Every gate generates a receipt; a receipt proves only the gate it actually checks.

Compression First Principles

• Zero-gap-timing-spacing-silence IS signal. No byte, no space, no absent row is noise. The compressor encodes everything; the decompressor must reconstruct everything, including the gaps, because the gaps ARE the compression.
• The receipt IS the compressed state — not metadata around it. Receipt dimensions (crossing matrix C, Sidon slack σ, step count k, residual series ε_seq, write timing t, scar absence ∅) together form the encoding. Invertibility of this receipt is the definition of lossless compression.
• Every compute substrate denies being a CPU: GPU (WGSL/wgpu), ASIC (SHA-256), FPGA (Verilog), PCIe/DMA (bus mastering), storage (NVMe/BTRFS), blitter (6502 OISC), hydraulic (pipes). All compute identically because Q16_16 integer arithmetic is deterministic across all of them. No substrate is privileged.
• The database IS the test field. enwik9 (1GB) and 259 TiddlyWiki tiddlers live in the same ene.packages table. The compressor does not distinguish between them. The schema is the substrate.
• Two distinct Lean theorems are required for every compressor:
  1. eigensolid_convergence — the braid crossing loop stabilizes
  2. receipt_invertible — given the receipt, the original state is reconstructible within bounded error, including all gap/timing/absence dimensions
• Float (ofFloat) is forbidden in compute paths. Q16_16.ofNat and Q16_16.ofRatio are the canonical constructors. ofFloat is only permitted at the external boundary (parsing JSON, reading sensor data) and must be immediately bracketed.

Legacy Recovery Trigger

The phrase RECOVER LEGACY INFORMATION is the explicit retrieval trigger for archived or quarantined concepts. Treat this as a user-controlled cold archive request, not permission to revive an old branch wholesale.

Accepted trigger forms:

RECOVER LEGACY INFORMATION: <path, commit, concept, or artifact>
Recover Legacy Information: <path, commit, concept, or artifact>
recover from cornfield: <path, commit, concept, or artifact>

When this trigger appears:

• Inspect the requested legacy source first with read-only commands such as git show, git log, or targeted file reads.
• Recover only the named file, concept, commit slice, or receipt requested.
• Modernize the recovered material onto the current clean branch before committing it.
• Never merge, reset to, or base new work on a legacy/cornfield branch unless the user explicitly asks for that exact branch operation.
• Preserve the legacy branch as retrievable archive state.

Current cornfield ref:

backup/distilled-with-vcd-history-2026-05-11

Nested Contracts

• Strict Lean/docs contract: 6-Documentation/docs/AGENTS.md
• Lean module-local contract: 0-Core-Formalism/lean/Semantics/AGENTS.md
• Infrastructure contract: 4-Infrastructure/AGENTS.md
• CAD harness contract: 5-Applications/text-to-cad/AGENTS.md
• QC flagger contract: scripts/qc-flag/AGENTS.md
• Lean expert agent contract: shared-data/artifacts/lean_expert_agent/AGENTS.md

Clean PIST predictions pipeline (no Rust authority)

Canonical IDs + dedup

Join key is invariant ID: equation_id := invariant_receipt.object_id (format rrc_eq_<hex>). The 278 source file contains duplicate object_id groups; predictions artifacts must be deduped by equation_id.

Preserve auditability by attaching provenance:

• summary.total_source_records = 278
• summary.unique_equation_ids = 250
• each prediction row includes source_records : Array {equation_record_id, name} for all source rows sharing that invariant.

Deterministic representative selection

When multiple compiled records share the same equation_id, select a representative deterministically:

representative := min(records, key = equation_record.equation_id)  (lexicographic)

Never "last wins".

Prediction artifact (v1, matrix-only)

Generate: shared-data/rrc_pist_predictions_278_v1.json

Constraints:

• schema: "rrc_pist_predictions_278_v1"
• claim_boundary: "matrix-only;no-classifier;no-lean-spectral"
• proxy_pred: null, exact_pred: null (until a classifier surface is defined)
• include:
  • global_vocab_hash
  • matrix_schema: "token_strand_adjacency_8x8_v1"
  • matrix_hash (sha256 of canonical row-major JSON, no whitespace)
  • matrix_8x8 (Int counts)

Generation rules (matrix_schema v1)

1. global vocab = all unique tokens across corpus, sorted
2. strand(token) = vocab_index % 8
3. matrix is 8×8 adjacency of token bigrams in original order, projected to strands: M[strand(t_i)][strand(t_{i+1})] += 1

Merge path into Lean corpus

pist_matrix_builder.py
  → writes rrc_pist_predictions_278_v1.json (dedup by invariant id)
  → build_corpus278.py reads it and merges by equation_id = rrc_eq_<hex>
  → regenerates Semantics/RRC/Corpus278.lean with pistProxyLabel/pistExactLabel
    populated when present
  → emit278.json alignment gate becomes non-missing_prediction only when labels exist.

Required validations (every change)

• python3 -m py_compile on touched shim scripts
• python3 -m json.tool on generated JSON
• reproducibility check: two consecutive runs must produce identical file SHA256
• lake build (full workspace) must stay green

NOTE on counts

Prediction artifact row count may be 250 while source record count is 278. This is correct: one prediction per invariant equation id, with provenance for all source records.

When to Use ContextStream Search:

✅ Project is indexed and fresh ✅ Looking for code by meaning/concept ✅ Need semantic understanding

---