core/CLAUDE.md

CORE Agent Instructions for Claude

Read this before modifying the repository. CORE is a deterministic cognitive engine under construction, not a transformer wrapper, not a generic chatbot, and not an infrastructure playground.

End Goal

CORE should become capable of:

listen -> comprehend -> recall -> think -> articulate -> learn from reviewed correction -> replay deterministically

The working design is now:

CognitiveTurnPipeline
  -> intent classification
  -> PropositionGraph
  -> ArticulationTarget
  -> deterministic realizer
  -> generation walk telemetry
  -> reviewed teaching loop
  -> deterministic eval/calibration replay
  -> deterministic trace hash

The system should become more capable by strengthening this path, not by adding opaque LLM fallbacks, stochastic sampling, hidden normalization, or broad infrastructure.

Philosophical Stance

Truth is coherent. Preserve coherence in algebra, memory, articulation, and teaching. Identity, truthfulness, and replayability are architectural commitments, not soft prompt preferences.

Code and tests should make illegal states difficult to represent. Prefer inspectable state, provenance, and deterministic replay over impressive-looking but ungrounded outputs.

Non-Negotiable Field Invariant

Every runtime field state F must satisfy:

versor_condition(F) < 1e-6

Do not weaken this threshold to make tests pass. Fix the operator/construction boundary that violated it.

Normalization Rules

Allowed sites:

• ingest/gate.py for raw input injection.
• language_packs/compiler.py and vocabulary construction.
• algebra/versor.py for algebra-owned sandwich closure.

Forbidden sites:

• generate/stream.py
• field/propagate.py
• vault/store.py
• logging/telemetry/runtime shell code

Do not add drift repair, grade projection, watchdogs, timers, hot-path normalizers, or monitoring functions whose only purpose is to repair another function.

CGA null vectors are not unit versors. Preserve null vectors as null vectors.

Core Primitives

Field transition:

versor_apply(V, F) = V * F * reverse(V)

Metric/recall:

cga_inner(X, Y)

Do not add cosine similarity, HNSW, ANN indexes, or approximate recall to the runtime path. Vault recall is exact and deterministic.

Current Key Modules

• core/cognition/pipeline.py — cognitive turn spine.
• core/cognition/result.py — result object for pipeline evidence.
• core/cognition/trace.py — deterministic trace hashing.
• chat/runtime.py — user-facing runtime contract.
• generate/intent.py — deterministic intent classification.
• generate/graph_planner.py — proposition graph and articulation target planning.
• generate/realizer.py and generate/templates.py — deterministic surface realization.
• teaching/correction.py, teaching/review.py, teaching/store.py — reviewed teaching loop.
• language_packs/data/en_core_cognition_v1 — core cognition semantic seed pack.
• evals/* — deterministic cognition eval harness.
• calibration/* — bounded replay-based calibration.
• docs/runtime_contracts.md — response, telemetry, memory, identity, and testing contracts.

GSM8K math comprehension substrate (sealed; serving stays 3/47/0 until ratified)

• core/reliability_gate/ — calibrated-learning ledger + gate (ADR-0175): ClassTally counts, conservative_floor (one-sided Wilson, N_MIN=10), θ ceilings.
• generate/derivation/ — the comprehension composer: extract.py (lexeme quantity extraction, EX-1/4/5 + function-word unit filter), clauses.py (GB-1 segmentation), compose.py (GB-2a list-sum + GB-3a clause-scoped referent guard), accumulate.py (GB-3b.1 single-referent gain/loss chaining), multistep.py/search.py (bounded search), verify.py (the wrong=0 self-verification gate: grounding ∧ cue ∧ unit ∧ completeness ∧ uniqueness).
• generate/cue_precision/ — (cue, op, unit_shape) reliability ledger + trainer (ADR-0177 CP-1/CP-2a); inert (consulted by no serving/gate path yet).
• evals/gsm8k_math/ — train_sample/ (real GSM8K, the capability metric), practice/ (sealed attempt-and-eliminate lane + ADR-0163-F additive set), confusers/ (ADR-0163-F2 discrimination probe — scored by wrong→0 + pair-consistency, NOT flip-count).
• scripts/verify_lane_shas.py, scripts/generate_claims.py --check — the serving-frozen gate (pinned eval-lane SHAs + CLAIMS.md).

Efficiency and Performance Doctrine

Performance is part of correctness for this project because slow feedback hides regressions and encourages unsafe shortcuts. Do not defer obvious hot-path or validation-lane issues until “later.”

Before changing hot paths, identify whether the change touches:

• algebra backend dispatch
• versor application / closure
• propagation
• injection / OOV grounding
• vault recall/storage
• session turn loop
• runtime/eval loops

Required approach:

1. Prefer semantics-preserving cleanup before new knobs.
2. Use algebra.backend for hot-path algebra when semantics are identical.
3. Hoist repeated imports and repeated structure-building out of tight loops.
4. Cache deterministic immutable data only, or return safe copies.
5. Keep exact CGA recall exact; use batching/vectorization, not approximation.
6. Validate speed-oriented changes through CLI lanes and core eval cognition.

Never improve speed by weakening closure thresholds, skipping construction checks, adding hot-path repair, replacing exact CGA with approximate metrics, or mutating shared cached state unsafely.

For test speed, prefer curated CLI lanes, small-case eval tests, safe fixture reuse, and immutable pack/load caching. Do not delete meaningful tests just because the full suite is slow.

Security and Trust Boundaries

Any change that touches user-controlled text, filesystem paths, dynamic imports, reports, pack validators, or logs must state the trust boundary.

High-risk surfaces:

• core pack validate dynamic validator execution.
• language/source pack loading.
• OOV token grounding and error messages.
• CLI commands that echo user content.
• eval/report output paths.
• pack mutation proposals.
• future file/network/database integrations.

Required approach:

1. Make arbitrary-code execution explicit and opt-in.
2. Reject path traversal and unsafe pack IDs before filesystem access.
3. Centralize safe display/log handling before increasing logging.
4. Keep pack mutation proposal-only unless a reviewed path applies it.
5. Avoid leaking raw sensitive tokens unless the command is explicitly local/debug.
6. Preserve deterministic replay evidence for security-relevant decisions.

Do not add hidden background execution, dynamic imports from untrusted paths, shell passthroughs, or broad filesystem writes without tests and a documented trust boundary.

Runtime Surface Contract

Keep these distinct:

• surface: selected user-facing response.
• walk_surface: raw manifold/token-walk evidence.
• articulation_surface: proposition/realizer surface.

Current policy:

surface = articulation_surface
walk_surface = retained telemetry/evidence

Any change must update docs/runtime_contracts.md and contract tests in the same PR.

Teaching Safety

Learning must be reviewed and auditable.

• Session memory may be immediate.
• Reviewed memory must go through teaching/*.
• Pack mutation is proposal-only until reviewed.
• Identity override attempts are rejected.
• User text must not mutate identity axes, runtime policy, or operator code.

Do not create a parallel correction/learning path.

Semantic Pack Discipline

Prefer compact, curated packs. Do not bulk-ingest corpora into runtime. en_core_cognition_v1 supplies thought vocabulary, operations, and relation predicates. Extend it cautiously, with deterministic ordering and pack tests.

Manifest checksums must hash the bytes actually written to disk:

checksum = hashlib.sha256(Path(lexicon_path).read_bytes()).hexdigest()

Documentation Discipline

ADRs, session docs, audit artifacts, and handoff briefs stay as Markdown (GitHub-flavored). Plain-text artifacts are diffable, greppable, and readable by every agent in the dispatch pipeline.

Within Markdown, two GitHub-rendered features are sanctioned and otherwise sparingly used:

• Mermaid fenced blocks (```mermaid) when a state machine, sequence, or dependency graph genuinely communicates more than prose. Inline, not in a sidecar file.
• <details> / <summary> collapsibles to fold long proofs, large tables, or generated logs without losing single-file context.

Out of scope:

• Standalone HTML artifacts with embedded CSS / inline SVG / sidebar navigation. The "open in browser" model breaks git diff, breaks determinism (CSS regen ordering, SVG element ordering), and breaks cross-agent legibility.
• Dashboards, status pages, or visualizers as a substitute for a pinned data artifact. If a visualization is load-bearing, the underlying data must live in a deterministic JSON/JSONL/Markdown artifact first; any rendering is a read-only view of that artifact.

Diagrams go inside the doc that needs them. Specs do not become single-file applications.

Schema-Defined Proof Obligations

When a schema, type, or struct exists for the sole purpose of naming a structural property the architecture claims to hold (HolonomyAlignmentCase, RoundTripFilter, the various Result discriminants), the obligation is real only when an executing test can meaningfully fail under the violations it is written to catch.

A test that passes under conditions that bypass the obligation it nominally proves is decoration, not proof. Before treating a schema type as a verified property:

1. Identify the violations the schema is written to catch.
2. Confirm an existing test would fail if exactly one of those violations were silently introduced (e.g. by mutating a weight, skipping a step, swapping a fallback).
3. If no such test exists, the obligation is asserted but not proven — record the gap in a follow-up doc rather than treating the schema as load-bearing.

This rule generalises the wrong=0 invariant. wrong == 0 holds because the admissibility gate, the round-trip filter, and the multi-branch disagreement check are all wired to fail loudly when violated. The same discipline applies to every other "this design guarantees X" claim in the codebase.

Validation Through CLI

Use CLI lanes instead of ad hoc pytest fragments:

core test --suite smoke -q
core test --suite cognition -q
core test --suite teaching -q
core test --suite packs -q
core test --suite runtime -q
core test --suite algebra -q
core test --suite full -q
core eval cognition

Run the smallest relevant suite first, then full before merge when practical.

Work Sequencing

Current near-term sequence:

1. Keep CLI lanes and core eval cognition green.
2. Tighten hot-path backend consistency and semantics-preserving performance.
3. Harden pack/OOV/logging trust boundaries.
4. Add exact vault recall indexing/batching without approximate search.
5. Add Rust backend parity only after Python semantics are locked by tests.
6. Expand curriculum teaching after replay/eval/calibration remain deterministic.

Avoid broad docs-first churn, dashboard work, or large infrastructure unless it unlocks one of these steps.

Lookback Review Discipline

Multi-PR architectural work accumulates latent defects when each PR is reviewed only against its own acceptance criteria. A hazard introduced in PR N can sit dormant until PR N+2 exercises it — by which point the substrate is harder to fix and three PRs are implicated rather than one.

Mandatory lookback review is triggered at three points:

1. Before starting the next phase of a multi-phase ADR. Before any code on Phase N+1, audit Phase N's shipped substrate. Check for: ADR-doc vs implementation drift, untested predicate paths, wrong=0 hazard surfaces, cross-phase trace/event/rank consistency, things the ADR says that didn't actually ship.

2. Before merging a stacked PR sequence into main. When 2+ PRs stack (PR #420 stacked on #416, #423 stacked on #420), the review-each-PR-individually pattern misses cross-PR consistency issues. Audit the whole stack as one unit before any merge.

3. After any 3+ PR sequence on the same module or architectural surface. When work concentrates on one area, regression risk compounds. Audit before claiming the surface is "stable" or "ready for the next layer."

What a lookback review covers (template — adjust per scope):

• Documentation drift. Does what shipped match what the ADR / brief said would ship? Signature differences, scope reductions, missing pieces — flag them.
• Test coverage gaps. Run the test suite under coverage. For every predicate/branch in a closed-set contract (like VALID_PREDICATE_NAMES), confirm at least one test asserts the specific elimination/admission path. Vacuous tests (assertions that pass under broken impl) are coverage gaps.
• Parity gaps. When a new implementation claims byte-equivalence with an existing one, exercise BOTH on the same inputs and confirm identical outputs — including failure modes, not just success.
• wrong=0 hazard surface. Every new code path: under what input conditions could it admit a candidate the prior path would have refused? Trace upstream to confirm no input class can trigger it. If a class CAN trigger it, build the defensive refusal NOW, before the next phase makes it load-bearing.
• Cross-PR consistency. Trace event shapes, rank handling, determinism contracts, dataclass invariants — do they compose cleanly across PRs?
• Honest LOC accounting. Did this phase net add or net remove lines? ADR claims of "removes ~N lines" only count post-collapse; intermediate phases that ADD substrate before removal happens should be called out.

Output. The review produces a structured report with findings categorized as: solid, gaps (no risk), drift (need amendment), and hazards (live wrong=0 risks). Hazards require a fix-before-next-phase decision.

Cost. A lookback review on a 3-PR substrate typically takes 20-40 minutes of focused tool calls. Skipping it costs more: every PR built on an undetected hazard becomes implicated when the hazard fires, and the fix has to land across multiple PRs instead of one.

Architectural Scan Exclusions

The invariant tests in tests/test_architectural_invariants.py perform full source-tree walks to enforce structural claims (INV-02, INV-21, INV-24). These scans must exclude .claude/ from traversal.

Why this matters: Agent operators (Claude Code, Codex, Gemini) create worktrees under .claude/worktrees/. Those worktrees contain full copies of the source tree — including vault/, chat/, generate/, etc. — and will trip every structural invariant that scans for forbidden callsites. The failures are silent killers: the tests report real-looking violations against files that aren't in the live codebase, poisoning the smoke suite and masking actual regressions.

Maintained exclusion sets (keep .claude in both):

# INV-02  os.walk exclusion (test_normalize_not_called_outside_gate)
{".git", ".venv", "__pycache__", ".pytest_cache", ".hypothesis", ".claude"}

# INV-21 / INV-24  rglob exclusion (EXCLUDED_DIRS)
{"tests", "evals", "benchmarks", "scripts", "docs",
 "core-rs", ".venv", "__pycache__", ".claude"}

If you add a new source-tree scan to the invariant suite, add .claude to its exclusion set before the first commit. Never rely on worktrees being pruned — they can persist across sessions and CI runs.

PR Checklist

Before opening or merging, answer:

What capability, performance property, or security boundary did this add/protect?
Which invariant proves the field remains valid?
Which CLI suite/eval proves the lane?
Did this avoid hidden normalization, stochastic fallback, approximate recall, and unreviewed mutation?
If it touches user input, files, dynamic imports, or logs, what trust boundary was enforced?

Prefer small, load-bearing PRs with clear evidence.