ivd

Changelog

All notable changes to IVD are documented here. Full reasoning behind each change lives in DECISIONS.md (FDR entries — Framework Decision Records).

This project does not yet strictly follow Semantic Versioning or Keep a Changelog format conventions, but is moving toward both. See ROADMAP.md.

[Unreleased]

• ComplyEdge checks wait 15 s by default (was 5). A cold ComplyEdge instance answers in about 8 s; with 5 s the check was skipped, so the caller lost the verdict and the Log ID of a record ComplyEdge had written. COMPLYEDGE_TIMEOUT_S still overrides it.

• ComplyEdge Log IDs reach the caller. Every checked tool call returns _meta["complyedge"] = {input_event_id, output_event_id}, outside the result text, which is returned exactly as checked. With the caller’s copy of the text, each Log ID verifies through ComplyEdge POST /v1/audit/verify. Privacy Policy and DPA describe the keyed fingerprint, pseudonymous IDs and 184-day retention.

• Every MCP tool call is now screened by ComplyEdge when the server has a key. call_tool checks the arguments (prompt) before the tool runs and the result (output) before it returns, sending agent_id, jurisdiction and the audit attribution fields (user_id as a hash of the caller’s IVD key, user_role, session_id). A block returns the rule instead of the result; an unreachable ComplyEdge lets the call through. Off without COMPLYEDGE_API_KEY. Privacy Policy and DPA list ComplyEdge as a sub-processor (v1.1). See FDR-027.
• Added ivd_attest — the process-attestation gate (33 tools, 19 core). ivd_validate checks the artifact; nothing checked whether the agent actually followed the method. An agent could skip the Rule 4 stress test, single-pass nine constraints against Rule 1, never re-read from disk per Rule 2, and still emit an artifact that validates clean — the read-acknowledge-violate failure mode turned on IVD’s own rules. ivd_attest mechanically checks Rule 1 (segmentation mandatory at 3+ constraints), Rule 2 (re-read, per-constraint coverage, joint satisfaction), Rule 3 (provenance cross-check), and Rule 4 probe coverage, returning COMPLIANT | INCOMPLETE | VIOLATION. The load-bearing check is the provenance cross-check: it is derived from the artifact rather than the attestation, so a claimed PASS on a constraint whose only evidence is an ai_generated test is downgraded to NEEDS_EXTERNAL_ORACLE regardless of what the agent asserts — the one assertion an agent cannot self-grade. Structure-only (no LLM, no network, no disk write). Governed by mcp_server/intents/attest_intent.yaml; 24 new tests. Honest bound documented in LEGAL.md §3.11: the attestation is agent-supplied, so this converts a silent skip into a visible claim — it raises the floor, it does not close the gap.
• ivd_validate now warns when a workflow- or system-level intent declares no termination contract (evaluation.cycle with max_iterations + a stop condition). Unbounded iteration and unclear completion are among the largest multi-agent failure modes; task/module intents are exempt. Warning only — never an error, so legacy intents keep validating.
• Fixed a stale claim in LEGAL.md §3.7 (R-006): it stated ivd_validate “does not warn on constraint count.” It does — the 7-constraint budget has been enforced at validate.py. Corrected to describe the advisory warning accurately.
• Security: bumped mcp 1.23.0 → 1.28.1 (requirements.txt, mcp_server/requirements.txt, pyproject.toml floor), clearing all 6 open Dependabot alerts (3 advisories: HTTP transport principal-verification, experimental task-handler isolation, WS Host/Origin validation). Validated in an isolated venv: server imports clean and the full transport/registry/ auth surface passes.
• Judgment phase — closed two loop gaps (see judgment_layer.md v1.2):
  • Added ivd_judgment_resolve MCP tool — records an entry’s resolution (outcome, optional held / fix_applied, auto resolved_at) and transitions codified | paired → resolved. Step 9 of the loop previously had no tool, so outcomes were never logged and future runs re-derived settled diagnoses. Tool count: 31 → 32 (10 judgment).
  • Added a 4th ivd_judgment_inject_context layer, ruled_out — surfaces injection_status: rejected comparison-pair hypotheses as a do-not-retry veto (negative knowledge), so the loop does not re-propose a falsified theory.
  • Added authored never / related_files fields to Pattern (do/never/related craft guidance), preserved across re-detection, excluded from detection_hash.
  • Additive-only: existing .judgment/ folders keep working; new fields default empty. 9 new tests in test_judgment.py (incl. a re-detection regression guard: authored never/related_files survive re-detect and stay out of detection_hash).
• Added ivd_import_spec MCP tool — parses a GitHub Spec Kit or OpenSpec spec.md into an IVD constraint scaffold (User Story / Requirement + Given/When/Then scenarios), read-only, no LLM call. Neither source format binds a scenario to an executable test by default; this tool exists to add that binding without replacing either tool’s own planning workflow. New imported_from: optional intent-artifact field for traceability. New recipes: import-spec-kit.yaml, import-openspec.yaml. Tool count: 30 → 31 (18 core).
• Added pyproject.toml — pip/editable installable, ivd-mcp console entry point, optional extras ([search], [compliance], [dev]).
• Added examples/intent_demo/ — offline, runnable demonstration of the core Intent → Implementation → Verification loop (catching a hallucinated implementation against a structured intent artifact).
• Added docs/positioning.md, ROADMAP.md, DEVELOPMENT.md.
• Removed unsubstantiated “zero hallucinations” marketing language from README and index.html; replaced with claims the constraint-check mechanism can actually back.
• Fixed tool-count drift across README, server.json, .env.example (actual count is 30 tools: 17 core + 9 Judgment + 4 Canon; two core tools — ivd_review_intent and ivd_run_constraint_tests — were previously undocumented in the count).
• Fixed recipe-count drift in README (table now lists all 18 recipes, including the 3 Judgment recipes that were previously omitted).
• CI compliance job (TrustLint) is now non-blocking (continue-on-error) so a ComplyEdge-side release fetch failure can’t fail unrelated contributor PRs.
• Fixed broken Quick Start MCP client config ("command": "python" pointed at system Python; dependencies are installed into .venv/ by setup.sh) — now points at the venv’s Python explicitly.

Note on versioning: the last tagged GitHub release is v2.4.0 (2026-03-20). The v3.0 (Judgment phase) and v3.1 (Canon phase) changes below shipped to main but were not tagged/released on GitHub. See ROADMAP.md for the plan to close that gap.

[3.1.0] — 2026-04-23 (shipped to main, not yet tagged)

Added

• Canon — Human Translation Layer (Phase 0): canon_render, canon_check, canon_diff, canon_check_rules_installed MCP tools. Enforces five communication invariants (Setting Phase, Confidence Calibration, Verification Beat, Folk Theory Management, Anthropomorphism Ceiling) on top of any LLM output. See canon_layer.md.
• ivd_judgment_check_installed — read-only detection of .judgment/ activation state; does not write to disk.
• recipes/canon-rules.yaml — pasteable Canon Phase 0a rules block.

[3.0.0] — 2026-03-21 (shipped to main, not yet tagged)

Added

• Judgment phase (Principle 9, IVD’s 4th phase): ivd_judgment_init, ivd_judgment_capture, ivd_judgment_codify, ivd_judgment_save_codified, ivd_judgment_pair, ivd_judgment_detect_patterns, ivd_judgment_inject_context, ivd_judgment_propose_recommendation — capture, codify, cluster, and inject real-world corrections as compounding contextual knowledge. Opt-in via .judgment/. See judgment_layer.md. (FDR-015, FDR-016)
• recipes/capture-correction.yaml, recipes/comparison-pair.yaml, recipes/distill-pattern.yaml.

[2.4.0] — 2026-03-20

Initial public release (tagged GitHub release).

Fixed (post-launch health audit, 2026-05-10 — FDR-017 through FDR-024)

• ivd_list_features returning duplicate artifacts (FDR-017)
• ivd_validate rejecting valid recipe artifacts (FDR-018)
• ivd_init guard not detecting existing named system intents (FDR-019)
• ivd_list_recipes returning opaque metadata for 8 of 17 recipes (FDR-020)
• ivd_discover_goal ignoring the user_hint parameter (FDR-021)
• ivd_propose_inversions returning a silent empty scaffold with no agent instruction (FDR-022)
• IVD rules block in .cursorrules missing detection fence markers (FDR-023)
• Published LEGAL.md — public disclosure of AI limitations, hosted-server data transmission, marketing-claim scope, and deployer obligations (FDR-024)

Added (pre-launch foundation, 2026-02-09 — FDR-001 through FDR-014)

• Core IVD methodology: Intent → Implementation → Verification loop, 15 core MCP tools
• Intent Stress Test (Rule 4), Post-Implementation Verification Protocol (Rule 2)
• Constraint-Segmented Implementation for 3+ constraints (Rule/Principle basis in read-acknowledge-violate and lost-in-the-middle research)
• Interpretive Entropy Spectrum for non-code constraint design (FDR-010)
• Constraint Satisfiability checks and constraint_satisfiability block (FDR-011)
• Contextual Knowledge Source Hierarchy + 2-attempt enrichment trigger (FDR-013)

This site is open source. Improve this page.