hypermind 2.1.2__tar.gz → 2.2.1__tar.gz

hypermind-2.2.1/CHANGELOG.md

@@ -0,0 +1,1285 @@
+# Changelog
+
+All notable changes to the **HyperMind** Python SDK
+([github.com/ZySec-AI/hypermind](https://github.com/ZySec-AI/hypermind))
+are documented in this file.
+
+The format is based on [Keep a Changelog 1.1.0](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning 2.0.0](https://semver.org/spec/v2.0.0.html).
+
+While the project is in the `0.x` series, minor-version bumps may include
+breaking changes; we will call them out explicitly in a **Changed** or
+**Removed** section. From `1.0.0` onwards, SemVer is binding: breaking
+changes only on major-version bumps.
+
+## [Unreleased]
+
+### Added
+- **Single-container streamable-HTTP MCP gateway (#1013).** `deploy/Dockerfile.mcp-http`
+  + `deploy/docker-compose.mcp-http.yml` front `hmctl mcp serve` with
+  [`supergateway`](https://github.com/supercorp-ai/supergateway), exposing a
+  plain streamable-HTTP `/mcp` endpoint for platform integrations (gateway-style
+  tool discovery services) that don't speak raw TCP/SSE. Safe-by-default posture
+  matches `deploy/Dockerfile.mcp`: write mode on by default, `MCP_WRITE=0` for
+  read-only. Documented in `deploy/README.md`.
+
+### Security
+- **CAP_REVOKE now gossips (#943 Phase 1).** A signed `CapRevocationRecord`
+  (spec §9.7/§9.7a) previously had no transport — an issuer could construct
+  and locally verify a revocation, but only the node that directly called
+  `store.record()` enforced it; every other node in the federation kept
+  honouring a stolen/compromised grant. `HyperMindAgent.revoke_capability_scope()`
+  now gossips the record on the namespace's bus and periodically re-gossips
+  it for 30 minutes per the spec's re-gossip obligation
+  (`hypermind.transport.cap_revoke_gossip`); every agent auto-subscribes on
+  `__aenter__()` so peers converge on the revocation without a separate pull
+  mechanism.
+- **Hardening: `CapRevocationRecord.from_cbor()` gained the canonical-CBOR
+  re-encoding gate** that `RevocationList.from_cbor()` already had — without
+  it, a validly-signed record has multiple valid byte encodings (the
+  signature covers only the semantic fields, not the raw bytes), a
+  malleability surface that matters now that gossip lets untrusted peers
+  send arbitrary bytes to this parser.
+- **Removed a dead, structurally-incompatible duplicate** of
+  `CapRevocationStore`/`InMemoryCapRevocationStore` that lived in
+  `hypermind.capability` (timestamp+grace-period-scoped) alongside the real,
+  live one in `hypermind.cap_revocation` (epoch-scoped) — the duplicate was
+  unreachable from `agent.py`'s `__post_init__` in practice (dead code
+  behind an always-false `is None` check) but was a correctness trap for
+  anyone who imported it directly expecting it to satisfy
+  `check_cap_revoked()`'s `max_observed_epoch()` contract.
+- `reference_verifier` gained `parse_cap_revocation_record()` — the
+  independent second implementation previously had no way to verify a
+  `CapRevocationRecord` at all.
+- **TAL v0.2 signed wire messages (#943 Phase 2).** `hypermind.tal.TAL`
+  was pure in-process Python state — room admission and vouches happened,
+  but nothing was signed and nothing crossed the wire. `create_room()` /
+  `join_room()` / `vouch()` now accept an optional `signing_keypair=` to
+  also emit a real `SignedStatement` (`ROOM_DEFINE`/`ROOM_JOIN`/`VOUCH`,
+  new record types inheriting A-ALG-BOUND automatically — see
+  `hypermind.tal_wire`) and gossip it once `attach_gossip()` is called
+  (auto-wired into every `_NamespaceWorld`/`HyperMindAgent`). New
+  `TAL.revoke_vouch()` capability (didn't exist before) backs the new
+  `VOUCH_REVOKE` message. `TAL.apply_signed()` independently re-verifies
+  every inbound message's signature *and* that the envelope's
+  cryptographically-bound issuer matches the payload's claimed
+  creator/member/voucher identity — closing a forged-claim class where a
+  validly-signed envelope could otherwise claim a false identity in its
+  payload. `reference_verifier` recognizes the 4 new record types.
+  Existing callers that never pass `signing_keypair=` see zero behavior
+  change. Reputation-floor updates and `TRUST_QUERY`/`REPUTATION_ATTESTATION`
+  are explicitly out of scope for this pass — see `hypermind.tal_wire`'s
+  module docstring for the full scoping rationale.
+
+### Added
+- **Bi-temporal recall + cross-namespace citation resolution (#942, L3
+  follow-up).** `confirmed_hypotheses` entries carried no real timestamp
+  before this — only `created_round`, a simulation round index, not a
+  point in time. `Hypothesis` gained `created_at_ms`/`resolved_at_ms`;
+  `KernelStore.recall_as_of(persona, as_of_ms)` filters a persona's
+  `learned_facts`/`confirmed_hypotheses` to what existed by a given
+  cutoff (existence-based filtering, not full point-in-time state
+  reconstruction — see the docstring for the scoping rationale).
+  `prove_hypothesis_inclusion()`/`verify_foreign_hypothesis_citation()`
+  add real RFC 6962 Merkle inclusion proofs for a single confirmed
+  hypothesis against the persona's sealed `lattice_root`, reusing the
+  same proof primitives the per-claim transparency log already uses —
+  so a citation across namespaces is independently verifiable, not just
+  trusted on the citing agent's word.
+
+## [2.2.0] — 2026-07-28
+
+Autonomy pass (#983): a code-grounded audit of 10 subsystems on the question
+"does this signal actually change a decision?", then closing every gap it
+found. Weighted autonomy score 2.9 → 4.6 / 5.
+
+**Upgrade note.** This is API-compatible — nothing removed, no new required
+arguments — but several defaults now behave differently, deliberately. If you
+depend on the previous behaviour, the relevant knobs are listed under
+*Changed defaults* below. The grounding gates in particular were documented
+no-ops and are now active.
+
+### Security
+- **SSRF: redirects were never re-validated.** `_validate_url()` checked only
+  the initial URL while `urlopen` followed redirects on its own, so an
+  innocent public URL could `302` to `169.254.169.254` and the fetch
+  proceeded — handing cloud instance credentials to whoever controlled the
+  first host. Every redirect hop is now validated, and all HTTP paths route
+  through the validating opener.
+- **SSRF: hostname checks were string-based.** A name an attacker controls
+  (`evil.example.com`) is on no blocklist and can resolve to `127.0.0.1`.
+  Hostnames are now resolved and every resulting address checked via stdlib
+  `ipaddress`, which also catches encodings a prefix check misses entirely
+  (`http://2130706433/` is `127.0.0.1`; `::ffff:127.0.0.1` is loopback).
+
+### Added
+- `hypermind.profiles` — deployment profiles (`coding`, `research`) bundling
+  persona roster, tool allowlist, model tiering and namespace. A second axis
+  to the verifiability class, not a mode switch: classification stays
+  per-question.
+- `hypermind.orgs.coevolution` — dispute-driven prover/disprover fitness
+  (Red Queen dynamics), fed by `OrgSim.coevolution_outcomes()`.
+- `hypermind.observability.silent_failures` — countable exception
+  suppression, with `hypermind.observability.health` classifying subsystems
+  as healthy/degraded/critical and turning critical failures into research
+  goals inside the loop.
+- `hypermind.knowledge.reputation_store` — local reputation persistence so
+  trust compounds across runs the way knowledge already did.
+- `hypermind.responders.cache` (`ResponseCache`, `CachingResponder`) and a
+  TTL'd provider-result cache in `orgs.tool_providers`.
+- `DifficultyRouter` — routes on the question's verifiability class instead
+  of on agent identity; V/G to a cheap model, C/S to a strong one.
+- `quality.set_entailment_scorer()` — optional NLI hook that blends with the
+  lexical grounding score rather than replacing it.
+- `SqliteBackend.store_embedding()` + cosine ranking fused with BM25 via RRF,
+  making the local backend genuinely hybrid. Opt-in; no embedder means
+  byte-identical FTS5-only behaviour.
+- `evolve_population()` generational archive with `stepping_stones()` and
+  `archive_lineage()`; gene pool widened to include `tools` and
+  `specialisation`.
+
+### Fixed
+- **A failed tool call counted as evidence.** `_publish()`'s `evidenced` gate
+  tested `'"error"' not in res`, matching only JSON-shaped errors, while all
+  eleven search handlers emit `[search error: …]`. A search that *timed out*
+  was published with `provenance="evidenced"` — the exact inversion the gate
+  exists to prevent. Failure detection is now centralised in
+  `tool_providers.is_tool_error()`.
+- **Enabling the evidence gate silently suppressed contradiction detection.**
+  `conflict_score = similarity × confidence_product`, so the parametric
+  penalty dropped the product ~34% and pushed genuine contradictions below
+  the detection floor — one safety mechanism switching off another. The
+  scanner now reads the new `_ClaimInfo.stated_confidence`.
+- **Reputation slashing ignored `weight`.** `record_outcome()`'s
+  smooth-slashing branch shaved α by a flat `lam` while the β increment
+  beside it was weight-scaled, so a caller could decline to reward an outcome
+  but never decline to punish one. `weight=1.0` still reproduces
+  `α·(1-lam)` exactly.
+- Expert relevance ranked on raw token overlap summed without bound: an
+  author with many weakly-relevant findings outscored the one author whose
+  finding matched the query's most discriminative term (measured 56.8 vs 6.0
+  on a synthetic corpus; now 6.5 vs 4.9). Now IDF-weighted as
+  `best_match + EXPERT_BREADTH_WEIGHT·log(1+n)`.
+- `RelataBackend.hybrid_search()` requested a BM25+vector ranking then
+  discarded it, re-sorting by raw substring count. Server and lexical
+  rankings are now fused with RRF.
+- Grounding could not distinguish a claim from its own negation — "X causes
+  Y" and "X does not cause Y" share every substantive token. Polarity
+  (parity-based) and numeric contradiction checks now apply.
+- Population clones never inherited their parent's toolkit.
+- `config.py` had a duplicate `ORG_STREAM_TOP_K` definition, the second
+  silently shadowing the first. Effective default (`5`) unchanged.
+
+### Changed defaults
+- `moa_answer()`/`moa_answer_from_messages()`: `on_not_ready` `"raise"` →
+  `"learn"`, so a cold topic self-bootstraps rather than refusing on first
+  use. `hmctl moa ask` and the `hm-moa` HTTP endpoint pin `"raise"`
+  explicitly, so their fail-fast behaviour and the `MOA-TOPIC-NOT-LEARNED`
+  runbook are unchanged.
+- `MIN_EVIDENCE_FOR_PUBLISH` `0` → `1` and `QUALITY_GATE_ENABLED` `0` → `1`.
+  Both shipped as documented no-ops; neither blocks a publish, they label and
+  discount.
+- `GROUNDING_VERIFY_MIN_OVERLAP` and `QUALITY_HALLUCINATION_THRESHOLD`
+  `0.05` → `0.20`. At 0.05 only a total non-sequitur could fail.
+- `MOA_MAX_REFINEMENT_ROUNDS` `1` → `2` — the critique/revise loop was a
+  no-op by default despite `refine_rounds` being wired.
+- `OrgSim(persist_identities=…)` `False` → `True`, so persisted reputation
+  keys to kids that recur.
+- MoA panel size and refinement rounds are now class-conditional when not
+  passed explicitly (V: 1 proposer, G: 6 + an extra round, C/S: 3).
+- Population fitness is derived from oracle-confirmed `strategy_perf`;
+  verdicts reached without an external ground truth are excluded from
+  fitness entirely and scaled by `SELF_GRADED_CREDIT_WEIGHT` (0.25) in
+  reputation.
+- Hypothesis promotion to `supported` requires a declared falsifier to have
+  been met; status weighs evidence quality rather than list length.
+
+### Changed
+- `moa_answer()`'s `k`/`refine_rounds` accept `None` (resolved per
+  verifiability class). Explicitly-passed values always win.
+- `oracle_resolve()` gains `credit_weight` (default `1.0`) so a verdict can
+  be priced by how falsifiable the claim was.
+- 449 previously-silent exception-suppression sites now record what they
+  swallow.
+
+
+## [2.1.5] — 2026-07-27
+
+### Changed
+- `hmctl eval` HTML report: professional dark-theme redesign (Inter + JetBrains Mono, dark palette, leaderboard cards, sticky tables, category pills, print-friendly)
+
+## [2.1.4] — 2026-07-27
+
+### Added
+- `hmctl eval` — 9-category IEEE-defensible benchmark (18 questions × 4 models)
+- Categories: Domain Knowledge, Hallucination Trap, Adversarial, OOD, Distributional Bias, Temporal Drift, Factual Friction, Confidence Calibration, Specificity
+- Auto-seed via `/v1/hm/publish` before eval
+- Category breakdown, calibration analysis, per-question comparison in HTML report
+- How-Each-Model-Works architecture table + MoA two-phase diagram
+
+## [2.1.3] — 2026-07-27
+
+### Added
+- `hmctl eval run` — research-grade HTML benchmark across all 4 models (hm-network, hm-corpus, hm-moa, hm-direct)
+- 5-question curated set: Domain Knowledge, Hallucination Trap, Adversarial categories
+- Self-contained HTML report: leaderboard, confidence bars, use-case matrix, model reference cards
+
+## [2.1.2] — 2026-07-27
+
+### Added
+- **`hmctl serve-endpoint` LLM responder auto-wiring**: auto-detects `OPENAI_API_KEY` / `OPENAI_BASE_URL` / `HYPERMIND_LLM_MODEL` from env and wires `OpenAIResponder` at startup — all three models (`hm-network`, `hm-corpus`, `hm-moa`) now generate real LLM rationale when a key is present. Falls back to reputation-only with a warning when no key is set.
+- **`--system-prompt` flag on `serve-endpoint`**: default scopes every panel member to swarm-verified knowledge and prefixes answers with `"HyperMind says — "`. Override with `--system-prompt "..."` per invocation.
+- **`"HyperMind says — "` response prefix**: all `/v1/chat/completions` responses (streaming and non-streaming) are now prefixed, making it clear answers come from the HyperMind swarm.
+
+### Fixed
+- Docker `CMD`: `hmctl mcp serve` now explicitly passes `--write --log-level info` — connections and tool calls are visible in `docker logs` output.
+
+## [2.1.1] — 2026-07-27
+
+### Added
+- **Web search result capture** (#973): agents now fetch every search result URL (previously capped at 3), convert HTML to clean markdown via `trafilatura` (strips nav/ads/footers), and store in Relata `/memory/remember` under `{agent_id}/web-search/{url_hash}` with full metadata. Each page is linked to its originating claim via `associate("source_for")`. Future rounds can recall full page content via `/memory/recall` instead of snippets only. `trafilatura>=1.6` added to the `[ingest]` optional extra; falls back to bs4 then skips gracefully.
+- **`hmctl status` endpoint panel**: a dedicated always-visible panel shows HyperMind's OpenAI-compatible endpoint URL, connection snippet, all models with descriptions, and all routes — turns green when live.
+- **`GET /v1/models` spec compliance**: response now includes `created` (Unix timestamp) alongside `id`, `object`, `owned_by`, and `description` — fully OpenAI-spec-compliant.
+
+### Fixed
+- Docker: `hmctl mcp serve` env file no longer overrides compose-set env vars (e.g. `RELATA_URL`) — process env takes priority over `/data/env` file values.
+
+## [2.1.0] — 2026-07-27
+
+### Added
+- **Mixture-of-Agents (MoA) OpenAI-compatible endpoint** (EPIC #949): `learn(topic)` runs one research cycle with a large model and persists what was learned to `KernelStore`; from then on `model="hm-moa"` on the existing OpenAI-compatible endpoint answers questions by having a small/cheap model panel **propose** independent draft answers, then verify or consolidate them — verified by sandboxed execution for code (`CodeExecutionAnchor`, #959), reputation/calibration-weighted synthesis otherwise (#951) — instead of passing through one panelist's raw text. Query digestion condenses multi-turn chat into a standalone, topic-routed, bounded-decomposed query before serving (#963); a verifiability-class router (V/C/S/G, #958) picks the right verdict mechanism per question and discloses trust/expertise tiers; a readiness gate fails closed with a stable `rule_id` and self-queues a research goal on cold topics instead of answering weakly (#956); bounded critique/revise refinement (#960) and real per-proposer SSE streaming (#952) round out the serving path. Self-improvement closes several loops: population evolution now reaches `select_experts()` (#961), `tool_outcomes` feedback is read as well as written (#964), the pipeline's own orchestration parameters (panel size, aggregation weighting) evolve under the same regression-gated pattern as personas (#966), and research decomposition uses best-first/beam search over independent candidate plans instead of single-thread refinement (#967). A mandatory adversarial Sceptic pass gates confidence-only hypothesis confirmation absent an external ground-truth anchor (#968), and lightweight integrity monitors flag — and seal as tamper-evident, signed records — gaming of the system's own verification (#969). A benchmark harness (#955) and signed, independently-replayable evaluation bundles (#965) substantiate the "beats a frontier model at a fraction of cost" claim, with sample efficiency (calls/tokens-to-confidence) tracked as its own objective alongside quality and `$`-cost (#970). `hmctl moa {learn,status,ask,serve,benchmark,replay-bundle}` plus `examples/30_moa_endpoint.py` (CI-safe, no API key required) cover the CLI and drop-in-adoption surface (#957, #971) — the real, unmodified `openai` Python SDK works against `hm-moa` via a `base_url` change alone, including standard-shaped error handling for structured refusals.
+
+### Added
+- **L1–L5 Autonomous Intelligence Stack** (EPIC #932): the full five-level autonomy ladder. **L1** — a `GroundingGapTracker` (#891) turns recurring zero-grounding topics in `OrgSim._think()` into self-generated research goals merged into `run_forever()`'s EIG-ranked candidate list. **L2** — `ReputationKernel.eigentrust()` (#892, #902) computes transitive trust over the vouch graph (EigenTrust power iteration), blended into `select_experts()`'s ranking and exposed via `hm_agents`. **L3** — `KernelStore` schema v6 adds a `lattice_root` (#894, #903): an RFC 6962 Merkle root over confirmed knowledge, independently verifiable via `KernelStore.verify_lattice()` without the local HMAC secret. **L4** — a deterministic evolve-tick → clone → outperform test plus `apply_calibration_feedback()` (#895, #904) wiring Brier/ECE calibration into `KernelStore.policy_params` via an opt-in `hm_calibration` `apply_feedback` param. **L5** — `propose_policy_change()`/`check_quorum()`/`apply_ratified_proposal()` (#896, #905): policy changes are published as signed claims and must survive the existing, unmodified dispute FSM before taking effect — a single dispute blocks ratification unilaterally. See `docs/concepts.md` §13 for the full write-up. Two smaller gaps closed alongside: versioned skill evolution with bounded history (#900) and a confirmation-rate quality gate on skill synthesis (#850), closing EPIC #934.
+- **Rust port** — a new `rust/` Cargo workspace (`hypermind-proto` protocol-primitives library + `hmctl` daemon-client CLI) covering the wire format (COSE_Sign1, A-ALG-BOUND), Ed25519 + ML-DSA-65 hybrid PQ signing, RFC 9591 FROST threshold Schnorr, the reputation kernel, the dispute FSM, Bloom-filter anti-entropy, and a TCP gossip transport interoperable with the Python `TCPGossipBus`. Proven via genuine cross-language interop tests (real subprocesses, real sockets) and conformance vectors shared with the Python SDK. See `rust/README.md` for scope and what deliberately stays Python-only (#924–#931).
+- Federated skill exchange with SCITT chain-of-custody, cross-swarm epistemic-strategy transfer, semantic pattern/trend detection, local dense-vector retrieval, emergent niche construction, auto-distillation of confirmed knowledge, meta-cognitive blind-spot tracking, adversarial red-team falsification agents, explicit consensus/voting protocol, hierarchical skill composition, opt-in learn-vs-serve routing, document ingest (PDF/DOCX/HTML/plain) at research-loop startup (#847–#869).
+- FTS5 full-text search + `hybrid_search()` on the SQLite backend (#918).
+
+### Fixed
+- **Critical security**: pickle RCE in the SQLite dispute backend replaced with CBOR (#882); missing `operator_affiliation`/`financial_interest` on RECEIPT payloads (#842); absent CAP_REVOKE/CAP_GRANT lifecycle (#841); namespace ACL `evaluate()` not wired into `_require_cap` (#840); L4 namespace cascade validation missing on inbound statements (#839); no post-generation grounding verification (#886); refuted claims retrievable as authoritative (#889); federated-skill RCE via bypassable exec blocklist, unsigned `HmSkill` rows, and unverified skill output entering LLM grounding (#912, #913, #915).
+- **High severity**: unbounded `confirmer_set` (#844); non-deterministic HLC-based reputation decay (#843, and the follow-on wall-clock/HLC seeding bug, #937); no reputation gate on federated skill adoption (#919); broken `delete_skill` (#917); skill-description prompt injection (#916); unsigned `KernelStore.learned_facts` (#898); `ask()` silently overriding `min_trust=0.0` (#899); non-atomic publish causing crash data loss (#893); receipts missing grounding-kid provenance (#890); panel herding from shared grounding (#897); `quality.py` never wired into live paths (#909); missing minimum-evidence gate on claim publication (#907); missing `close_dispute()`/`recall()` SDK verbs (#858, #856); refuted hypotheses leaking into `learned_facts` (#877); free generation on zero retrieval (#876).
+- 18 medium- and 4 low-severity findings across storage integrity, retrieval ranking, fact dedup, and degraded-mode visibility (#878–#888, #914, #920–#923).
+
+### Changed
+- Retention policy across the retrieval/storage layer is now confidence/recency-weighted instead of FIFO (#880).
+
+Full detail for every item above is in the individual closed GitHub issues; this PR is the cumulative diff.
+
+## [1.12.1] — 2026-07-25
+
+### Fixed
+- **CI quality gate** — greened the test suite against the cycle-52/53 engine gap work and fixed latent bugs: a `Deliberator` `AttributeError` (`ActionDecision.topic`) that left ticks undispatched, a self-deadlock in `working_memory` LRU eviction (non-reentrant lock re-acquired on the same thread), and env-var resolution in `KernelStore` (`HYPERMIND_HOME` / `RELATA_URL` now honour runtime overrides over the import-time snapshot). Explicit `rounds=` is again a fixed-count contract; recall depth honours `retrieval_k`; `HYPERMIND_CONSOLE` is read live.
+
+## [1.12.0] — 2026-07-24
+
+### Fixed
+- **PATH-3** — `start_loop()` / `_loop_task()` / `_run_cycle()` now accept a `persona` kwarg so OrgSim can supply the member label directly; previously the KernelStore persona key fell back to `kid_hex[:32]`, silently severing the loop→OrgSim knowledge bridge.
+- **PATH-5** — broken `and None` in `hm_activity` Relata backend lookup (`mcp_tools.py`) always resolved to `None`, so the claims-count stat was never read from the live backend.
+- **PATH-6** — promoted `HYPERMIND_INGEST_ROOT`, `HYPERMIND_MAX_CLAIM_TEXT_CHARS`, and `HYPERMIND_CONSOLE` to named constants in `config.py`; all three previously used raw `os.environ.get()` calls scattered across the codebase.
+
+## [1.11.0] — 2026-07-22
+
+### Fixed
+- **Open-issue sweep** — closed all remaining open tickets: YAML frontmatter multi-line fix, `async to_thread` wrapper for `auto_promote`, and Definition-of-Done closures across the backlog.
+
+## [1.10.1] — 2026-07-22
+
+### Added
+- **S6 self-evolution — learning retention & skill compounding** (#814–#822) — cross-sim learned-fact persistence and skill compounding via `KernelStore`; confirmed hypotheses and work history hydrate `working_memory` on the next sim start.
+- **Self-evolution loop over MCP** — `hm_evolve_now` and related tools expose the evolution cycle to MCP clients.
+- **MCP surface gaps + cross-sim data completeness** (#828–#837) — closed MCP tool-coverage gaps and completed the cross-simulation data path.
+
+### Fixed
+- **G7 / G12 / G13 / G18** — reasoning-strategy parsing, EIG refresh, `work_history` wiring, and prediction-match scoring in the hypothesis lifecycle.
+
+### Docs
+- Synced `docs/spec/20-swarm-intelligence.md` and `CLAUDE.md` with the self-evolution changes (G7/G12/G13/G18).
+
+## [1.9.7] — 2026-07-21
+
+### Added
+- **HTTP MCP servers in Docker** — added Node.js + `mcp-remote` to the container image so it can talk to HTTP/SSE MCP servers, not just stdio.
+
+## [1.9.6] — 2026-07-21
+
+### Fixed
+- **MCP tools wire in serve path** — `attach_mcp_config` is now awaited inside the async `_serve` so discovered MCP tools are registered before the first request.
+
+## [1.9.5] — 2026-07-21
+
+### Fixed
+- **MCP env-file precedence** — a local env file now overrides the empty defaults injected by docker-compose at serve startup.
+
+## [1.9.4] — 2026-07-21
+
+### Added
+- **Tool-calling responder in `hm_ask`** — wired `ToolCallingResponder` with search tools into the `hm_ask` agent panel so consult can invoke tools.
+
+## [1.9.3] — 2026-07-21
+
+### Added
+- **`GoalHandle.cancel()`** — running goals can now be cancelled.
+- **Social-media-research MCP scenario** — new scenario exercising tool discovery end-to-end.
+
+### Fixed
+- **Tool discovery sync** — resolved a race in tool-discovery synchronisation.
+- `scenario_social_media_research` now always uses `SyntheticResponder` in demo mode and always passes an explicit `llm_client` to `OrgSim`.
+
+## [1.9.1] — 2026-07-21
+
+### Added
+- **AUTO model selection** — `model=AUTO` picks an appropriate backend from the environment.
+- **MCP defaults and status improvements** — saner MCP defaults and richer serve status reporting.
+
+### Fixed
+- **Grounded HyperMind retrieval with Relata** (#813) — `hm_ask`/retrieve now grounds on real Relata sources instead of stubbed hits.
+- **E2E MCP suite** — corrected tool param names and handled single-node swarm tools gracefully (57/57 pass).
+
+## [1.9.0] — 2026-07-17
+
+### Changed
+- **Deprecation removal** — dropped Python 3.11 support; removed the deprecated `close()` path while restoring the `witness_cosign()` / `legacy-ed25519` compat shim and `from_env()` on a TCP bus.
+- **`HYPERMIND_HOME`** — consolidated home-directory env var.
+
+### Added
+- **Docker image + file-based MCP config** (#812) — official `ghcr.io/zysec-ai/hypermind` MCP image, file-based MCP server config, and CI publish-on-tag.
+- **Pre-push ruff lint gate** for contributors.
+
+### Fixed
+- **Dispute `_STRICT_QUORUM`** — `close()` without `witness_signatures` now respects the strict-quorum guard.
+- `from_env(sandbox=False)` now builds the world with the TCP bus so `agent._world.bus` is the real transport.
+
+## [1.8.3] — 2026-07-13
+
+### Fixed
+- **Wave-1/2 issue merge** — 100+ backlog issues resolved across the SDK surface.
+
+## [1.8.2] — 2026-07-13
+
+Release-plumbing bump; no functional changes beyond `[1.8.1]`.
+
+## [1.8.1] — 2026-07-12
+
+### Added
+- **Persistence epic #647** — TAL, hypotheses, skills, and world state are now persisted; skills auto-load from the backend on startup and the registry is wired into deliberators.
+- **Agent learning → skills pipeline** — confirmed learning is distilled into reusable skills (spec §20.7 Skill Evolution Loop).
+
+### Security
+- **Skill execution surface hardened.**
+
+### Changed
+- Skills fall back to the file store (not SQLite) for local persistence.
+
+## [1.8.0] — 2026-07-12
+
+### Added
+- **SQLite is the default storage backend**, with Relata discovery hints surfaced in the MCP server banner.
+
+### Docs
+- Clarified that releases go through GitHub Actions (tag-triggered), not manual `twine` upload.
+
+## [1.7.10] — 2026-07-12
+
+### Changed
+- **Removed the in-memory storage fallback** — SQLite is always the default; there is no silent no-op storage path.
+
+## [1.7.9] — 2026-07-12
+
+### Fixed
+- **Scalability pass (medium/low)** — closed the remaining 11 medium/low unbounded-growth and concurrency hazards.
+
+## [1.7.8] — 2026-07-12
+
+### Fixed
+- **Scalability pass (high)** — closed 6 high-severity unbounded-growth and concurrency hazards.
+
+## [1.7.7] — 2026-07-12
+
+### Fixed
+- **MCP discovery/pagination** — closed four discovery and pagination gaps in the MCP tool surface.
+
+## [1.7.6] — 2026-07-12
+
+### Fixed
+- Widened a parallel-tick timing test so it passes reliably on loaded CI runners.
+
+## [1.7.5] — 2026-07-12
+
+### Changed
+- **Consult panel timeout** raised to 120s, configurable via `HYPERMIND_LLM_TIMEOUT_S`.
+
+## [1.7.4] — 2026-07-12
+
+### Fixed
+- **Gap-fix Waves 1–3** — closed integration-seam gaps across `mcp_tools`, `responders`, `reputation`, `citation_graph`, `tcp`, `anti_entropy`, `hypothesis`, `replay`, `evidence`, `topic_clustering`, `tool_providers`, `client`, `loop`, `working_memory`, `evolve`, `orgs/sim`, and the mind layer (`agent`, `belief`, `world_model`, `goals`, `mediator`, `dataset`, `leverage`).
+- **Claude Desktop stdio** wiring, resilient Relata startup, and core-dependency pinning.
+- ruff clean-up after wave-3 (RUF005, F841, B905).
+
+## [1.7.3] — 2026-07-12
+
+### Fixed
+- **Autonomous-engine Sprint 4 audit** — 10 critical integration-seam gaps closed.
+
+## [1.7.2] — 2026-07-12
+
+### Added
+- **Multi-user TCP/WebSocket MCP listener** (#601) — serves multiple concurrent MCP clients over a shared TCP/WebSocket transport.
+
+## [1.7.1] — 2026-07-12
+
+### Added
+- Revocation conformance vectors.
+
+### Fixed
+- **Autonomous-engine Sprint 3 audit** + loop test stabilisation (`load/save_loop_state` patched in loop timing tests).
+- Tickets #582–#595 — analysis, roles, pin_policy, dispute `rule_ids`, cli.
+- `hm_publish_claim` now respects `topic=` (auto-join topic room in `client.publish`).
+
+## [1.7.0] — 2026-07-12
+
+### Fixed
+- **Autonomous-engine Sprint 2 audit** — 18 gaps closed across 8 subsystems.
+
+## [1.6.9] — 2026-07-12
+
+### Fixed
+- **Full-stack autonomous-engine audit** — 21 gaps closed across 6 subsystems.
+- **Author-attribution correctness** — author identity corrected in `append_claim` / `retrieve()` (relata backend + client); attribution merge from audit agent.
+
+## [1.6.8] — 2026-07-12
+
+### Changed
+- **ruff format CI gate** fixed; format violations in `examples/` corrected.
+
+## [1.6.7] — 2026-07-12
+
+### Fixed
+- **5-gap audit** (cli, tal, responders, pin_policy) — agent list, goal submit, vouch decay, prompt caching, `compose()`.
+- `argue()` guard + lint cleanup + `__all__` sort (dispute, harness, examples).
+
+## [1.6.6] — 2026-07-12
+
+### Fixed
+- **TCP graylist pruning, FROST nonce persistence, OrgSim embedding dedup, hypothesis validation** (transport, crypto, orgs).
+- **Bounded unbounded structures + atomic FROST write** (transport, orgs, crypto).
+- 3 autonomous-engine gaps in `hm_calibration` and `hm_evolve_now` (mcp).
+
+## [1.6.5] — 2026-07-12
+
+### Fixed
+- Guard `test_issue_357` with `importorskip` — relata is not in CI `[test]` extras.
+
+## [1.6.4] — 2026-07-12
+
+### Security
+- **`oracle_resolve()` 1-of-1 quorum bypass (HM-DISPUTE-QUORUM-001)** — a single signer could satisfy the dispute resolution quorum; closed.
+- **SSRF / DNS-rebinding in `_is_safe_url`** — integrations URL-safety guard hardened.
+
+### Added
+- CLI commands: agent list, goal submit, claims list; all `HYPERMIND_*` env vars documented.
+
+### Fixed
+- Deliberator posterior merge order; `ContraClaimDisputePolicy` import resolved.
+- Autonomous-engine sprint: oracle FSM, deliberator gaps, loop calibration, missing MCP tools.
+- `HYPERMIND_LOG_ROOT` containment guard for `log_file` hooks.
+- `hm_hypothesis` wired into `HypothesisBook` lifecycle.
+- `InformationGainPlanner` removes the winning question from `working_memory` after `pick()`.
+- `from_env(sandbox=False)` — `connect_to_peer`→`connect`, `bus.start()`, TCP bus wired into world.
+- Raw `bond_topic_scope` passed to `dispute.close()` instead of SHA-256 hex.
+
+## [1.6.3] — 2026-07-12
+
+### Fixed
+- All CI failures resolved — env isolation, hook test, hybrid search stubs.
+- `hm_retrieve` SQL error; `hm_ask` Ollama model auto-detection.
+
+## [1.6.2] — 2026-07-12
+
+### Added
+- `strategy_history` wiring + full MCP tool test suite (76 tests).
+
+### Fixed
+- Autonomous-engine gaps — `GoalQueue.push()`, deliberator goals lookup, peer persistence.
+
+## [1.6.1] — 2026-07-12
+
+### Changed
+- Lint clean, coverage gate at 75%, duplicate command removed.
+
+### Fixed
+- Autonomous-engine final audit pass + verification pass — remaining critical/high gaps closed.
+- Order-dependent test failures and deprecation warnings resolved.
+
+### Docs
+- README badge counts updated for v1.6.0 (2772 tests, 52 MCP tools).
+
+## [1.6.0] — 2026-07-11
+
+### Added
+- **Autonomous engine — cohesion & wave 4–5.** Roles API (`hypermind.mind.roles`), federation relay, distributed kernel store, cross-run learning, in-loop research with anti-herding, Jaccard-fix similarity, Brier-score wiring into reputation, OrgSim ticking scheduler, and a startup coordinator. Event-pipeline wiring + end-to-end tests + feedback loops across the mind / storage / integration layers (12 gaps closed).
+- **MCP tool audit.** Improved every tool description, fixed input schemas, added missing tools (dispute/oracle/vouch/synthesize/hooks surface; v1.6 additions: `hm_goal_submit`, `hm_claims_list`, `hm_dispute_list`, `hm_reputation_show`). Tool count now **52 read-only / 53 with `--write`**.
+- **CLI + SDK exports.** New `hmctl` commands, expanded SDK `__all__` exports, and an enforced coverage gate.
+- **Three-layer event-driven integration** (ADR-0006) — inbound webhooks (`hmctl webhook serve`), outbound notifications (`HYPERMIND_NOTIFY_URL` / `_FILE`), and a declarative runtime hook chain (`~/.hypermind/hooks.yaml`, `hm_hooks` tool).
+
+### Changed
+- Retrieval now wires dispute status into the `hybrid_search` score; vouch guard uses a structured `BlockReason`; contested claims are flagged in synthesis output.
+- Structured logging (`structlog`) across hot paths; async-correctness and performance hotfixes; improved error messages, type hints, `repr`, `hmctl doctor`, and API-friction cleanups.
+- Reputation is flushed on agent close (last-agent-exit guard); `OTel` tick metrics; dual oracle event emit.
+
+### Fixed
+- Test-suite green restoration (batch-write helper, deliberator syntax fix), circular import in `_run_deliberator`, `asyncio.get_event_loop()` → `asyncio.run()`.
+
+## [1.5.3] — 2026-07-11
+
+### Fixed
+- **Dataset export overhaul** — SFT / DPO / reward-model export quality improved for downstream training use.
+
+## [1.5.2] — 2026-07-11
+
+### Changed
+- Removed deprecated MCP tool aliases, leaving a clean 41-tool surface (pre-1.6.0 audit baseline).
+
+## [1.5.1] — 2026-07-11
+
+### Fixed
+- Relata bearer-token passthrough into backend requests.
+
+## [1.5.0] — 2026-07-11
+
+### Added
+- **Autonomous loop v2** — always-on learning loop per topic with adaptive scheduling, topic clustering, and automatic sub-loop spawning (`hm_loop`, `hm_loop_ingest`, `hm_pool_status`, `hm_agent_memory`, `hm_evolve_now`, `hm_session_summary`).
+- **Self-directed learning tools** — `hm_next_question`, `hm_scan_contradictions`, `hm_calibration`, `hm_federate`.
+- **MCP tool consolidation + CLI expansion.**
+
+### Fixed
+- `hm_retrieve` default `min_trust` lowered 0.6 → 0.5 for in-memory fallback; `utcfromtimestamp` deprecation.
+
+## [1.4.9] — 2026-07-11
+
+### Added
+- `hmctl` config wizard + first-install nudge.
+
+## [1.4.8] — 2026-07-11
+
+### Changed
+- SQLite auto-activated as the default local storage backend; RelataDB messaging clarified.
+
+## [1.4.7] — 2026-07-11
+
+### Changed
+- Maintenance release; version bump.
+
+## [1.4.6] — 2026-07-11
+
+### Added
+- **Autonomous learning engine** (issues #289–#295) — `hm_stats` MCP tool + `hmctl status` segmented knowledge-base snapshot.
+
+### Fixed
+- Federation `build_filter` used in place of the nonexistent `BloomSetReconciliation`; ruff lint across autonomous-engine files.
+
+## [1.4.5] — 2026-07-11
+
+### Added
+- `hm_stats` MCP tool + `hmctl status` — segmented knowledge-base snapshot.
+
+## [1.4.4] — 2026-07-11
+
+### Added
+- `hm_status` MCP tool + `hmctl status` command — full configuration picture.
+
+## [1.4.3] — 2026-07-11
+
+### Added
+- ASCII logo + 30-second quickstart banner on `hmctl mcp serve`; `~/.hypermind/env` config file + rich startup banner.
+
+### Fixed
+- Test isolation, coverage gate, MCP audit trail; tolerance comparison for float weights in `test_fit_weights_track_quality`.
+
+## [1.4.2] — 2026-07-11
+
+### Fixed (security)
+- **webui: `vite` <6.4.2** — `server.fs.deny` bypass on Windows (high) + NTLMv2 hash disclosure via UNC path (medium). Updated to `^6.4.3`.
+- **webui: `markdown-it` <14.1.1** — quadratic complexity DoS in smartquotes rule (medium). Updated to `^14.2.0`.
+- **webui: `linkify-it` <=5.0.0** — quadratic algorithmic complexity in match scan (high). Resolved via markdown-it bump.
+- **transport-rust: `pyo3` <0.29.0** — out-of-bounds read in `nth`/`nth_back` for `PyList`/`PyTuple` iterators (high) + missing `Sync` bound on `PyCFunction::new_closure` (medium). Updated to `0.29`.
+- **transport-rust: `pyo3` <0.24.1** — buffer overflow risk in `PyString::from_object` (low). Resolved by 0.29 bump.
+
+## [1.4.1] — 2026-07-11
+
+### Changed (docs/docs-cleanup)
+- **README** — rewrote with structured 34-tool MCP table (organized into 8 categories), elevated marketing (three-layer capability table, learn-large/serve-small, offline mode callout), expanded CLI section. Condensed 60-line natural-language chat block to 2-line teaser.
+- **ROADMAP** — collapsed ~430 lines of shipped v0.x milestone detail into a compact summary table (669 → 250 lines). Removed stale cleanup-pass notes, local plan paths, and historical kill-criteria. Bumped all version references.
+- **Release notes** — standardized all five files (v1.0.0–v1.4.0) to a consistent template (Highlights → Breaking changes → Cross-references). Renamed `v1.2.md` → `v1.2.0.md`. Fixed typo `murphy_winkier` → `murphy_winkler`. Removed v1.0.0 duplicate Highlights section and pre-GA artifact. Deleted v1.0-checklist stub.
+- **Feature matrix** — added 16 missing `__all__` symbols (incl. `open`, `__version__`, `__wire_version__`, `__protocol_version__`, `ProtocolVersionError`, `SLSA_LEVEL`, all 7 type symbols, 3 revocation symbols). Fixed broken test/example references (`test_belief.py` → `test_belief_revision.py`; `reference_verifier/examples/verify_one.py` → removed). Added "Module constants & shared types" section.
+- **WebUI landing** — fixed wrong license (MIT → Apache-2.0), broken CLI command (`hmctl console` → `hmctl simlab`), stale versions (v1.0.0 → v1.4.1, v0.10 → v1.4.1, 1800+ tests → 2400+), stale spec count (28 → 30), unified SimLab/Console naming. Rewrote Security.vue "Reserved for v1.0" → "Future hardening (post-1.4)".
+- **Docs cleanup** — deleted `docs/archive/` (8 superseded files), untracked scratch artifacts (`persona_audit.jsonl`, `persona_trace.json`, `trace.json`, `experiments/` ~10 MB). Fixed stale GA dates in `lts.md`, `migrate/`, `operate/02-runbook.md`. Rewrote `_meta/public-surface.md` (dropped all "Provisional v0.5.x" markers). Added orphaned docs to mkdocs.yml nav (SIMLAB, simulations, sop-knowledge-quality, decide/, paper/, simlab/modes).
+
+### Removed
+- `docs/archive/` — 8 superseded pre-GA files (self-flagged "may be deleted").
+- `docs/release-notes/v1.0-checklist.md` — GA go-live checklist (GA shipped; items tracked in ROADMAP §External).
+- `persona_audit.jsonl`, `persona_trace.json`, `trace.json` — gitignored scratch output (~164 KB).
+- `experiments/` — gitignored scratch dir (~10 MB JSONL/HTML/scripts).
+
+## [1.4.0] — 2026-07-11
+
+### Added
+- **Local/offline mode**: `hmctl mcp install --local` — zero-config setup using Ollama + SQLite, no internet or API key required.
+- **Ollama auto-detection**: `env_responder()` probes `localhost:11434` when no LLM config is set; picks the best available model (prefers llama3 > mistral > phi > qwen).
+- **Persistent SQLite ledger**: `HYPERMIND_DB=~/.hypermind/ledger.db` wires a durable local ledger into `build_mcp_server`; all published claims survive process restarts.
+- **Population checkpoints**: `HyperMindClient` auto-checkpoints the swarm's reputation to `~/.hypermind/populations/<name>.json` on `close()` when `HYPERMIND_POPULATION` is set.
+- **`hm_population_list`**: List all saved populations with agent count, top topics, and claim count.
+- **`hm_population_get {name}`**: Detail a population's personas, roles, and earned reputation per topic.
+- `HyperMindClient.sandbox(db_path=, population_name=)`: wire durable storage and population checkpointing programmatically.
+
+### Changed
+- `env_responder()`: Ollama auto-detect added as final fallback before deterministic mode.
+- `build_mcp_server()`: reads `HYPERMIND_DB`, `HYPERMIND_POPULATION`, `RELATA_URL` from env.
+- `HyperMindClient.close()`: checkpoints population + closes SQLite backend when configured.
+- `hmctl mcp install`: new `--local` flag bakes in Ollama + SQLite env vars.
+
+## [1.3.0] — 2026-07-11
+
+### Added
+- **Verifiable RAG**: `hm_retrieve` now uses Relata hybrid BM25 search when `RELATA_URL` is set; each result carries author, trust score, and claim provenance.
+- **`hm_justify`**: full evidence chain for any claim — text, author, reputation, dispute history, citations.
+- **`hm_recall`**: memory search over the signed ledger; degrades gracefully to in-process scan without Relata.
+- **`hm_related`**: graph traversal via Relata `PATHS_BETWEEN` (CITES / CORROBORATES links); requires `RELATA_URL`.
+- **`hm_learn`** (writable): trigger a focused OrgSim research run on a topic; publishes signed findings to the ledger.
+- `RelataBackend.hybrid_search()`: BM25 MATCH with LIKE fallback over denormalized `HmClaim` rows.
+- `HyperMindClient.sandbox(relata_url=)`: wire a Relata backend for durable, hybrid retrieval.
+- `HyperMindClient.justify(kid_hex)` and `HyperMindClient.recall(query)`: programmatic access to evidence chains and memory search.
+
+### Changed
+- `RelataBackend.append_claim()`: now denormalizes `author_hex`, `text`, `confidence` into the Relata type row — enables native text search without decoding wire bytes on every query.
+- `hm_retrieve` description updated: "Verifiable RAG" framing with Relata-upgrade note.
+
+## [1.2.2] — 2026-07-10
+
+First PyPI release since 1.0.0 — ships the full 1.1–1.2 feature set that had
+landed in git but not been published: the eval metric suite, the MCP manager
+tooling + machine identity, the verified corpus, and the integrations layer.
+See the 1.1.0 / 1.2.0 / 1.2.1 sections below for the detailed change lists.
+
+### Added
+- **Eval metric suite** (`hypermind.eval`) — proper-scoring-rule primitives
+  (`logarithmic_score`, `expected_calibration_error`, `skill_score`,
+  `murphy_winkler_decomposition`), the `compute_metric_suite` bridge,
+  `fit_composite_weights` (regression-fit composite replacing legacy CIQ
+  constants), and `render_html_report`. New `hmctl eval` command prints the
+  metric table with source footers and writes a self-contained HTML report
+  (`--output report.html`) with a Single-LLM comparison column.
+- **MCP manager tooling** — `hm_roster`, `hm_agent`, `hm_topic_experts`,
+  `hm_disputes`, `hm_tools`, and research-as-a-tool (`hm_research_*`).
+- **Machine-derived signing identity** (`hypermind.machine`).
+- **Version visibility** — `hm_version` MCP tool + startup banner.
+
+### Changed
+- `hm_ask` now uses a real OpenAI-compatible LLM when configured (OpenRouter,
+  Ollama, vLLM, LM Studio), with Anthropic as a fallback.
+
+### Fixed
+- MCP `initialize` reports the real version (was a hardcoded `0.10.1`).
+- TLS profile no longer SIGABRTs the process under statically-linked OpenSSL.
+
+## [1.2.1] — 2026-07-10
+
+### Added
+- **`hm_version` MCP tool** + startup banner — report the running build (version,
+  wire version, read-only/read-write mode, LLM status, tool list) both in-chat
+  (`hm_version`) and in the host's MCP server log (a stderr banner on
+  `hmctl mcp serve`), so it's always clear which build is live.
+- **`hm_ask` uses a real LLM when configured** — OpenAI-compatible first (OpenRouter,
+  Ollama, vLLM, LM Studio via `OPENAI_BASE_URL`; local servers need no key), with direct
+  Anthropic as a fallback when `ANTHROPIC_API_KEY` is set. `hmctl mcp serve` wires the
+  responder from env (provider via `HYPERMIND_LLM`, model via `HYPERMIND_LLM_MODEL`) so
+  `hm_ask` returns real deliberation instead of the deterministic reputation-only
+  fallback. No config → unchanged offline behaviour. `build_mcp_server` gained an
+  optional `responder=` param (default `None`, so tests never hit a live API).
+- **`hmctl mcp install --env KEY=VALUE`** (repeatable) writes the server entry's `env`
+  block — the MCP host does not inherit your shell env, so this is how `OPENAI_BASE_URL`
+  / `OPENAI_API_KEY` / `RELATA_URL` reach the server. Merges with any existing env.
+
+### Fixed
+- **MCP `initialize` now reports the real version** in `serverInfo.version` (was a
+  hardcoded `0.10.1`).
+- **TLS profile no longer SIGABRTs the process.** `transport/tls_profile.py` reached
+  the underlying `SSL_CTX*` by guessing a struct offset from `id(ctx)` and calling
+  `SSL_CTX_set_ciphersuites` on a `find_library("ssl")` libssl that could differ from
+  the one CPython's `_ssl` is linked against (e.g. statically-linked OpenSSL under
+  uv/macOS) — aborting the whole process when building any TLS context. Removed the
+  unsafe raw-pointer path; the TLS 1.3 cipher floor is recorded as policy (there is no
+  safe stdlib API to set TLS 1.3 ciphersuites). TLS 1.3-only, ALPN, PFS groups, and
+  ticket policy are unchanged.
+
+## [1.2.0] — 2026-07-10
+
+### Added
+- **Manager-dashboard MCP tools** (`hypermind.integrations.mcp_tools`) — read-only
+  views over the swarm: `hm_roster` (every agent's role/rooms/reputation/ramp),
+  `hm_agent` (one agent's capability snapshot + earned topics + published claims),
+  `hm_topic_experts` (agents with standing in a topic, ranked), and `hm_disputes`
+  (dispute counts by FSM state, topic-scoped). Backed by new `HyperMindClient`
+  verbs `roster` / `agent_detail` / `topic_experts` / `dispute_summary`.
+- **`hm_tools` MCP tool** — lists swarm tool providers with live availability
+  (missing-env reasons), configured MCP servers, and a curated set of good-to-have
+  MCP servers to attach.
+- **Research-as-a-tool over MCP** (`hypermind.integrations.research`) — async job
+  model: `hm_research_start` kicks off a real `OrgSim` run and returns a job id;
+  `hm_research_status` polls progress (round + findings-so-far); `hm_research_result`
+  returns findings. Writable-only; reads `RELATA_URL` + an LLM key from the server env.
+- **Machine-derived signing identity** (`hypermind.machine`) — `hmctl mcp serve/install
+  --machine-identity` derives a deterministic, traceable Ed25519 seed from the host's
+  machine id (`machine_id` / `machine_seed` / `machine_uuid`), so the same machine
+  reproduces the same identity across runs. Opt-in; the random seed remains the default.
+
+### Changed
+- **`hm_ask` no longer accepts a `topic` parameter.** In the single-room sandbox any
+  non-default topic had zero panel candidates and failed; the param is removed. Scope
+  by topic via `hm_publish_claim` / `hm_retrieve`. The no-model fallback text is now
+  explicit ("Reputation-only signal (no model wired)…") instead of a generic stub.
+
+## [1.1.0] — 2026-07-10
+
+### Added
+- **Verified corpus** (`hypermind.orgs.corpus`, spec §29) — trust-scored, attributed
+  knowledge export from a run's ledger. Per-claim trust score blends reputation,
+  corroboration, dispute-survival and hypothesis status; an external ground-truth
+  verdict overrides toward its pole. `build_corpus` / `export_corpus` /
+  `records_to_corpus` / `corpus_stats`, plus `hmctl orgs corpus {stats,population,topics,claims,export}`.
+- **Ground-truth anchoring** (`hypermind.orgs.anchor`, spec §30) — `Verdict`-returning
+  `GroundTruthAnchor` with `FeedAnchor` (anchor to reality), `TieredAnchor` (verify only
+  high-stakes claims), `CompositeAnchor` (precedence). `.as_ground_truth()` /
+  `anchor_from_ground_truth()` bridge the existing `GroundTruth` slot.
+- **Persistent, evolving population** — `HyperMindAgent.testbed(seeds=…)` (deterministic
+  Ed25519 identities), `OrgSim(persist_identities=True, seed_reputation=…)`,
+  `OrgSim.export_reputation()`, and `ReputationKernel.export_posteriors()` /
+  `import_posteriors()` so a population carries earned reputation across runs
+  (a consistently-correct member compounds instead of resetting).
+- **Influence-operation templates** — `influence_operation` / `counter_influence_operation`
+  org templates plus the `influence` interaction mode, for information-integrity research.
+- **`hypermind.integrations` — drop-in adoption layer.** `HyperMindClient` (+ sync mirror
+  `SyncHyperMindClient`) is a one-object facade over network generation, trust-scored
+  retrieval, signed claims, reputation, and ground-truth verification on an in-process swarm
+  (no Relata server, no key for the deterministic primitives): `HyperMindClient.sandbox(...)`,
+  `publish` / `ask` (→ `NetworkAnswer`) / `retrieve` (→ trust-scored `CorpusRecord`s) /
+  `reputation`. Plus an **OpenAI-compatible endpoint** (`hmctl serve-endpoint`) — point any
+  framework's `base_url` at it and generation is answered by a reputation-weighted panel
+  grounded in the verified corpus; an **expanded local MCP server** (`hmctl mcp serve`)
+  exposing the full operational surface (simulate, query the corpus, ask the network,
+  publish/verify) for Claude Desktop / Cursor; **framework adapters**
+  (`hypermind.integrations.langchain.HyperMindRetriever`, plus OpenAI Agents SDK, CrewAI,
+  AutoGen/AG2, LlamaIndex); and `hmctl init` / `hmctl quickstart` scaffolding. See
+  [docs/integrations.md](docs/integrations.md).
+
+## [1.0.0] — 2026-07-09 — Engineering GA
+
+The reference SDK is **feature-complete**. From this release, **SemVer is binding**:
+breaking changes only on the next major version.
+
+What 1.0.0 covers (engineering):
+- **Protocol core** — COSE_Sign1 wire format (A-ALG-BOUND), Ed25519 + composite PQ
+  signing, RFC-6962 transparency log, HLC, dispute FSM, capability tokens,
+  reputation kernel, revocation — with a second independent verifier for interop.
+- **Federated research swarm** (`hypermind.orgs`) — master-decomposed experts,
+  self-refined planning, 12 pluggable reasoning strategies, tools (MCP + search),
+  100% evidence retention, real-grounded citations, the four architecture-native
+  metrics (verifiable/attributable/contestable/grounded), a hypothesis lifecycle
+  (form → keep in memory → test → update), streaming tools-off query with
+  reputation-ranked expert selection, and learn-large/serve-small.
+- **Storage** — real Relata ledger (signed HmClaim/HmDispute/HmReputation typed
+  rows) + per-agent memory (evidence + hypotheses), with full-run JSONL export +
+  replay-into-Relata.
+- **Transport** — in-process (sandbox) and real multi-node **TCP federation**
+  (identity-bound handshake, gossip, duplicate-kid rejection). libp2p remains an
+  optional external Rust binding for the same capability.
+- **CLI** — `hmctl orgs` (templates / simulate / ask / distributed / doctor)
+  with an animated banner and a production-preflight `doctor`.
+
+Out of scope for the SDK (tracked in ROADMAP, gate the *ecosystem*, not the code):
+IETF I-D submission, IANA allocation, OpenSSF acceptance, third-party security
+audit, Tamarin/ProVerif proofs, foundation awards, pilot MOUs, chaos-test run.
+
+## [0.37.0] — 2026-07-09
+
+### Removed
+- **Cloud-KMS signing** (unused `NotImplementedError` stubs) — `KMSSigner` and the `crypto/kms_backends/{aws,gcp,azure,vault}` backends, the `kms-*` extras, and the KMS test suite. The **`Signer` protocol + `LocalSigner`** (real in-process Ed25519, the shipped backend) are unchanged; a custom HSM/KMS `Signer` can still be supplied by any caller. No production code imported the removed surface; local signing + verification are untouched.
+
+### Changed
+- **Real multi-node federation is first-class + tested (v0.2 transport).** The shipped `TCPGossipBus` (identity-bound handshake, gossip publish/subscribe, duplicate-kid rejection) is the production multi-machine path — proven end-to-end in `tests/test_tcp_federation.py` (a signed message federates across two nodes; a same-kid second connection is rejected). `agent.open(mode="libp2p")` now points at this real TCP path; **libp2p** remains an optional external Rust binding for the same capability, not a blocker.
+
+## [0.36.0] — 2026-07-09
+
+### Added
+- **Full-run export + replay** (`hypermind.orgs.replay`) — `export_jsonl(report, path, world=…)` dumps **every** record from a run (signed claims as **wire bytes**, disputes, reputation, citations, Merkle head, evidence, hypotheses) to one JSONL; `replay_to_relata(path, …)` re-inserts a dumped run into any Relata **later, without re-running the LLM work** — claims round-trip byte-identical through `SignedStatement.from_bytes` (pass verification). `hmctl orgs simulate --export run.jsonl`.
+- **Real signed ledger, un-stubbed** — the typed `HmClaim`/`HmDispute`/`HmReputation` ledger persists to **real Relata** once the four Hm types are provisioned (admin token) + ACL-granted (boot-time). `deploy/relata-serve.sh` captures the reproducible server launch. Verified end-to-end.
+- **Animated `HYPERMIND` ASCII banner** (`hmctl.dashboard.banner`) shown by `hmctl orgs simulate` / `doctor` with the full run detail.
+- **`hmctl orgs doctor` production validation** — checks LLM key, reasoning-strategy registry, MCP tools, Relata reachability, and Hm ledger-type registration; critical checks gate readiness, capabilities are informational.
+
+## [0.35.0] — 2026-07-09
+
+### Added
+- **Architecture-native report metrics** (`hypermind.orgs.analysis`) — measures the four properties only HyperMind's machinery produces, not generic RAG scores: **Verifiable** (Merkle root + audit completeness + a runnable byte-flip **tamper-evidence** demo), **Attributable** (earned reputation share + Gini + role split), **Contestable** (bonded-dispute survival), **Grounded** (knowledge-source split tool/reuse/parametric + **source-faithfulness** checked against the *retained raw evidence*). Pure functions over an `OrgReport` — no LLM, no latency. `full_analysis()`.
+- **Query the learned swarm** — `OrgSim.stream()`: a **tools-off, internal-RAG, streaming** answer over the swarm's accumulated knowledge (no fresh internet). `select_experts()` consults a panel ranked by **relevance × earned reputation** (not all agents). `model=`/`client=` overrides enable **learn-with-a-large-model, serve-with-a-small-one** — grounded synthesis a small model does well, verifiable via source-faithfulness.
+- **Reasoning-strategy registry** (`hypermind.orgs.strategies`) — 12 pluggable frameworks in 4 groups (TRIZ, first_principles, morphological, scamper, analogical / five_whys, systems_thinking, fmea / popper_hypothesis, pre_mortem, dialectic / bayesian). `strategy` is first-class on `Seat`; the PI assigns one per expert during decomposition and it's injected into the system prompt — diversifying *how* the swarm reasons, not just what it searches.
+- **Hypothesis lifecycle** (`hypermind.orgs.hypothesis`) — each finding becomes a **falsifiable hypothesis** (falsifier from `key_unknowns`, tagged with its reasoning method), **kept in the agent's Relata memory**, and **tested** by the existing loop: a citation corroborates, a bonded dispute disconfirms, an oracle resolution is the verdict. Belief evolves `open → contested → supported/refuted` across rounds. `OrgReport.hypotheses()` / `hypothesis_summary()`.
+
+### Fixed
+- **Context-window scaling at 8–10+ agents** — decoupled *retain 100%* from *feed-the-LLM*: tool results fed back into the conversation are capped (`HYPERMIND_TOOL_RESULT_LLM_CHARS`, default 4 KB) and finding reasoning fed into retrieval/synthesis context is capped (`HYPERMIND_CONTEXT_FINDING_CHARS`, default 900) — the full result/reasoning is still stored on the claim/evidence. Fixes 128k `context_length_exceeded` overflow with large Scholar/Patents payloads.
+
+## [0.34.0] — 2026-07-09
+
+### Added
+- **100% evidence retention** — every claim now persists a COMPLETE evidence record: untruncated reasoning, verbatim LLM payload, and **full raw tool output** (previously capped at 500 chars). Kept off the size-bounded signed claim and stored durably in the agent's Relata notebook. `OrgReport.evidence()` / `evidence_for(kid)` expose it. (`responders/tools.py`, `responders/structured.py` adds `reasoning_full` + uncapped `raw_json`, `orgs/sim.py`).
+- **Structured citations + real grounding** (`hypermind.orgs.evidence`) — parses raw tool output (searchapi JSON + DuckDuckGo text) into de-duplicated `Citation`s ranked by multi-expert corroboration. `OrgSim._retrieve` now grounds the judge/answer on **real source snippets**, not only the swarm's own summaries. `OrgReport.references()` is the study bibliography. Pure parsing, cached by claim count — no added LLM latency.
+- **Self-refined master decomposition** — `hypermind.orgs.prompts` (planner prompt + curated few-shot exemplars, one-/multi-shot), `charter.plan_research(shots=, critique=)`, and `hypermind.orgs.refine` (`score_plan` + `plan_and_refine`) re-plan the decomposition on a structural critique (distinct areas, role diversity, tool breadth, on-topic, vs-k) **before** spending the simulation. No framework, reproducible.
+- `RelataMemoryStore(single_tenant=True)` — omit `X-Organization-Id` for single-tenant / trial (`max_tenants:0`) servers, isolating by `session_id` (agent kid); `OrgSim(relata_single_tenant=…)`.
+
+### Fixed
+- **MCP transport** — 16 MB stdio buffer for MCP subprocesses; large Google Scholar/Patents results no longer trip asyncio's 64 KB `LimitOverrunError` (`orgs/mcp_config.py`).
+- Tool-call traces key the tool name under `tool` (matching production `ToolCallingResponder`); tests aligned.
+
+## [0.33.0] — 2026-07-09
+
+### Added
+- **Pluggable tools — provider registry + per-cycle gateway** (`hypermind.orgs.tool_providers`). Each `ToolProvider` declares its env gate + health; the `ToolGateway` refreshes availability **every round** and exposes only live tools (call-time gated), emitting a `tool_status` event. Built-in web search: **DuckDuckGo** (keyless) + **Tavily / Brave / Serper** (key-gated).
+- **MCP auto-discovery** — a standard `mcpServers` JSON at `~/.config/hypermind/mcp.json` (or `$HYPERMIND_MCP_CONFIG`, same shape as Claude); every server's tools are discovered, registered, and health-checked per cycle (`attach_mcp_config`, `ToolGateway.add_mcp`). HTTP servers bridge via `npx mcp-remote <url>`.
+- **`hmctl orgs simulate/ask --tools/--no-tools`** (default on) — wires the gateway (built-in search + MCP servers) into the run.
+- `OrgSim(tool_gateway=…)` — per-cycle status check + live tools for tool-using seats.
+
+## [0.32.0] — 2026-07-09
+
+### Added
+- **Persistent shared config** — `hmctl orgs` auto-loads keys from `~/.config/hypermind/env` (then `~/.hypermind/env`, then `./.env`, or `$HYPERMIND_CONFIG`); exported env still wins. Store keys once, never re-export.
+- **Env-backed flags** — every `hmctl orgs` flag also reads an env var (`RELATA_URL`, `OPENAI_BASE_URL`, `HYPERMIND_LLM`, `HYPERMIND_LLM_MODEL`, `HYPERMIND_API_KEY_ENV`, `HYPERMIND_TEMPLATE`), so with the persistent config the commands run **flag-free** (`hmctl orgs ask "…"`).
+- **`hmctl orgs config`** — shows the resolved config paths + which keys are set (masked).
+
+## [0.31.1] — 2026-07-09
+
+### Added
+- **Relata in docker-compose** — `docker compose -f deploy/docker-compose.yml up -d relata` starts a ready-for-orgs Relata (`ghcr.io/zysec-ai/relata`, lite mode + `Hm*` ACL grant) on `:9090`. (Prometheus moved to host `:9091`.)
+- **k8s distributed fleet** — `deploy/k8s/orgs-distributed.yaml`: Relata + master + N worker replicas; `kubectl apply` then `kubectl scale deploy/hm-orgs-worker`. README "Deploy & scale" shows the one-liner (compose/docker) and the cluster (k8s) paths.
+
+## [0.31.0] — 2026-07-09
+
+### Added
+- **`hmctl orgs doctor`** — one-command preflight for adoption: checks the LLM API key, Relata reachability, and that the `Hm*` ledger types are writable, with ✓/✗ and a fix hint per row. Exit 0 only when ready.
+- **`.env.example`** — every env var the orgs engine needs, in one place.
+- **Deploy & scale docs** — README section: set env → `doctor` → start Relata (with the `RELATA_ACL_GRANT` boot line) → single-host run → cluster (master/worker) commands.
+
+### Fixed
+- Example prerequisites pointed at a non-existent `docker compose … relata` service; now show the real `relata serve` command with the ACL grant.
+
+## [0.30.1] — 2026-07-08
+
+### Added
+- **Fancy quality panel with fill animation.** `hmctl orgs ask` now renders the answer + quality rubrics as a colour-graded Rich panel — per-rubric gauges (grounding, **hallucination**, calibration, consensus, …), an overall gauge, and a green/red hallucination callout. On a real terminal the gauges **animate (fill up)** on reveal; piped/CI output renders once. So hallucination and every other metric are now visible clearly, not a plain text list.
+
+## [0.30.0] — 2026-07-08
+
+### Added
+- **Live simulation dashboard** (#196) — a colourful, animated Rich view of a running sim: round progress bar, glowing stat cards (claims/citations/disputes/resolves), a per-round claims sparkline, a per-unit tally, and a colour-coded live activity feed. Turn on with `hmctl orgs simulate --live` / `ask --live`.
+- **Cross-shard grounding** (#195) — in a distributed run, agents now ground on *other shards'* findings pulled from the shared Relata ledger (`RelataBackend.recent_findings`, `OrgSim(cross_shard=True)`, default-on in the distributed worker), so a sharded run reasons like one swarm instead of N isolated ones. (Cross-shard citation/dispute graph + reputation unification is tracked as a follow-up.)
+- **`hmctl orgs` — no-code CLI for the living-simulation engine.** Run the whole engine from the shell:
+  - `hmctl orgs templates` — list presets.
+  - `hmctl orgs simulate --template university --members 24 --rounds 6 --topic "…" [--html report.html]`.
+  - `hmctl orgs ask "question" --template … [--samples N] [--no-quality]` — study→ask with the quality rubrics **on by default** (grounding, hallucination, calibration, native ledger rubrics).
+  - `hmctl orgs distributed --template … --members 10000 --workers 8 --topic "…"` — sharded run; the master provisions Relata + reassigns failures.
+  - LLM + Relata configured via flags/env (`--llm`, `--llm-model`, `--base-url`, `--api-key-env`, `--relata-url`); Hm* types auto-provisioned.
+
+## [0.23.0] — 2026-07-08
+
+### Added
+- **Distributed, Locust-style execution** — `hypermind.orgs.distributed`. A master coordinator shards a swarm across workers (other machines or local processes) over a TCP control plane (length-prefixed CBOR frames); each worker runs its shard as `OrgSim(lightweight=True)` against the **shared Relata namespace** and returns bounded aggregates, merged into one `OrgReport` (#177).
+  - `Coordinator` / `Worker`, `run_distributed(...)` (single-host convenience), and a `master` / `worker` CLI (`python -m hypermind.orgs.distributed`).
+  - Capacity-weighted shard planning; a worker that drops before returning is recorded as a failed shard — the run completes and notes it, never hangs.
+  - `OrgSim.export_aggregates()` for the merge; `OrgModel` wire (de)serialisation.
+
+## [0.22.0] — 2026-07-08
+
+### Added
+- **Batched Relata writes** — `RelataBackend(write_batch_size=N)` buffers claim rows and flushes via one bulk `batch_upsert` instead of a per-claim round-trip; `OrgSim` defaults to 64. ~batch× fewer round-trips at scale (#175).
+- **Bounded / streaming report** — `OrgSim(lightweight=True)` caps memory to O(members + windows), not O(records): per-member streaming aggregates + a rolling claim window, exact scorecard totals. `metrics.scorecard_from_aggregates(...)` (#176).
+- Parallel keypair generation at `testbed()` spin-up (thread pool; the crypto releases the GIL) — part of #177.
+
+## [0.21.0] — 2026-07-08
+
+### Added
+- **Native response-quality rubrics** — `hypermind.orgs.quality`: `sim.assess(question)` → `QualityReport`. Four rubric groups, all on by default, switchable via `disable`: native/deterministic (provenance coverage, dispute exposure, reputation support, consensus, abstention — verifiable from the signed ledger), judge-based faithfulness (grounding, hallucination rate, answer relevance, context precision), calibration (Brier, self-consistency), and context recall. Missing deps degrade to informational notes. `llm_judge(...)` factory.
+
+## [0.20.0] — 2026-07-08
+
+### Added
+- **Hybrid retrieval** — `reciprocal_rank_fusion(...)` and `hybrid_retriever(*retrievers, reranker=None)` fuse dense + entity-graph + lexical signals (RRF) with an optional cross-encoder rerank, over our own Relata data. Plugs into the existing `OrgSim(retriever=...)` seam.
+
+## [0.12.0] – [0.19.0] — 2026-06 … 2026-07
+
+### Added
+- **`hypermind.orgs` — the living-simulation engine.** Turn a declarative `OrgModel` / `Unit` / `Seat` into a running multi-agent simulation (research → publish → cite → dispute → resolve → learn → evolve) persisted to Relata, driven by pluggable `Interaction` strategies (`cooperative` / `adversarial` / `crowd`).
+  - **Template library**: `university`, `hospital`, `city_at_war`, `foreign_policy_cabinet`, `un_security_council`, `internet_crowd`, `research_lab`, … via `list_templates()` / `get_template()` — the core never imports them.
+  - **Per-agent models** (per-seat provider/model or a router) and **group-scoped native tools** (declarative or random grants; ungranted tools are absent, not refused).
+  - **Study → ask**: `sim.study(topic)` then `sim.ask(question)` — a transparent, queryable swarm that answers from what it learned, weighted by earned standing.
+  - **Proper RAG** for `ask`: lexical → semantic embeddings (`embedder=`, `openai_embedder`) → claim↔entity graph (`use_graph_retrieval`, `MentionGraph`) → real NER hooks (`llm_ner_extractor`, `chat_ner_extractor`, `pioneer_ner_extractor`) → Relata-backed retrieval (`relata_retriever`).
+  - **Scorecard & analysis**: reputation ranking, citation DAG, dispute ledger, Brier calibration (#151), convergence metrics (#152), memory-RAG (#153), deliberator mode (#154), tool-trace capture (#155), adversarial dynamics (#156), pluggable ground truth (#157), experiment/variant harness (#158), interventions (#159), visualization snapshots (#160), verifiable receipts (#161).
+- **Scalability hardening** across the engine: bounded peer window, O(1) member indices, top-k cross-pollination, concurrent think phase (`asyncio.gather` + semaphore) and concurrent blocking I/O (`asyncio.to_thread`).
+- `examples/org_university.py` and an expanded examples index.
+
+## [0.11.0] — 2026-05-10
+
+### Added
+- `AgentCard` dataclass + `HyperMindAgent.publish_card()` verb + `hmctl card publish` CLI (P3-01, spec §3.2)
+- §21 Belief State Layer spec section — `TheoryOfMind`, `BeliefStore`, `WorldModel` (P3-06)
+- libp2p `peer_id` derivation (`derive_peer_id()`) locked in spec §1.9 and implemented (P3-07)
+- `standards-map.md` updated for all v0.9.x/v0.10 spec sections — 7 new rows (P3-10)
+- §11.1e gaps closed: Gap 3 (AE resumability §6.7), Gap 6 (mediator selection §8.3b), Gap 7 (21 conformance vectors), Gap 10 (STH cross-gossip §9.9), Gap 12 (TAL predicate grammar §7.G)
+- Normative spec sub-sections: §4.4 task protocol, §4.8a staleness decay, §6.4 KNOWS_DELTA, §19.5a cap-gated autonomy, §20.5a SwarmIQ recommendations
+- Tutorials 22-25: AgentCard, task delegation, MCP tool call, swarm trace (examples/)
+- Hypothesis property-based tests (test_property_based.py)
+- SLSA L3 cosign + slsa-github-generator release workflow (.github/workflows/release.yml)
+- Appendix Z expanded to 30 conformance vectors (9 new: mcp_provenance ×2, knows_delta ×2, task_handle ×2, swarm_trace ×2, reputation_decay ×1)
+- Fuzz harnesses: `fuzz_dispute.py`, `fuzz_agent_card.py`
+- `swarm_trace.py` — `new_span_id()`, `encode_trace_context()`, `OTEL_TRACE_CONTEXT_LABEL`
+- All normative-trace.md TBD rows (31-40) resolved
+
+### Changed
+- Strict quorum now default-on; opt-out via `HYPERMIND_LEGACY_QUORUM=1`
+- `require_chain=True` default in `CapToken.authorize()`
+- `RevocationList.from_cbor(verify=True)` verifies signature by default
+- ruff format + lint applied (167 files reformatted)
+
+### Fixed
+- R1-01..06: spec internal numbering drift corrected across 5 spec files
+- Stale scaffold guards in `agent.py` and `namespace.py` removed
+- `asyncio` import missing from `responders/tools.py`
+
+### Deprecated
+- `HyperMindAgent.from_env(rooms=…)` — `rooms=` param removed; use `HyperMindAgent.sandbox()`
+
+## [0.10.0] — 2026-05-10
+
+**Milestone close-out: distributed agentic coordination layer.** v0.10 ships
+the federated MCP transport, durable signed workflows, contract-net
+auctions, the cognitive layer (belief state + Theory-of-Mind + federated
+swarm memory), production ops (distributed swarm-trace, KMS backends,
+benchmarks), and 14 industry UseCase demos.
+
+### Added
+
+- **T1 — MCP federated transport** (§4.12.2 / §4.13.1). Inbound MCP tool
+  calls bind `mcp_provenance` to the protected header (label `-65627`)
+  with `calling_agent_kid`, `cap_token_root`, and `request_kid`. Idempotency
+  cache deduplicates retries within 24 hours.
+- **T2 — Durable signed workflows** (§coordination-T2). Saga workflow
+  records carry `workflow_saga_id` (label `-65628`) and
+  `workflow_step_index` (label `-65629`) in the protected header.
+- **T3 — Contract-Net protocol** (§27). `RFP_OPEN` / `BID` / `AWARD`
+  records with reservation labels `-65630..-65634`: `rfp_id`, `bid_kid`,
+  `winning_bid_kid`, `award_signer_kids`, `bidder_filter`. AWARD is
+  FROST-signed by ≥1 signer kid.
+- **T4 — Cognitive layer.** Belief state (§21 promoted to Normative),
+  Theory-of-Mind agent models, and federated swarm memory wired through
+  the Mind layer's `Deliberator` and `WorldModel`.
+- **T5 — Spec foundations.** Network agent-to-agent envelope path with
+  cascaded security and first-class private/org namespaces.
+- **T6 — Production ops.** `OTEL_TRACE_CONTEXT` (label **`-65635`**) for
+  W3C-Trace-Context distributed swarm-trace (§28). KMS backend hooks for
+  signing material rooted outside the SDK process. Coordination
+  benchmarks pinned in CI.
+- **14 UseCase demos** under `examples/coordination/usecases/` covering
+  cybersecurity (cy-05, cy-11), healthcare (hc-06, hc-21), finance
+  (fi-04, fi-22), government (gv-01, gv-21), AI research (ai-01), and
+  five generic-ID demos for legal e-discovery, energy grid coordination,
+  manufacturing supply chain, education credentials, and AI alignment
+  red-team consortium.
+
+### Changed
+
+- **§21 (Belief State) promoted to Normative.** `valid_from_hlc` (label
+  `-65625`) and `valid_until_hlc` (label `-65626`) are now defined for
+  BELIEF records.
+- **SwarmMemory federation wired** through the Mind layer; per-agent
+  memory snapshots flow through the federated world model.
+
+### Fixed
+
+- **R3-01 — wire-label registry collision.** Spec §4.13 originally
+  declared `OTEL_TRACE_CONTEXT = -65628`, but the code-side allocation
+  pre-claimed `-65628..-65634` for T2 (Saga) and T3 (Contract-Net).
+  Reconciled to code (source of truth): `OTEL_TRACE_CONTEXT = -65635`.
+  Spec §03-messages.md, §28-distributed-trace.md, and
+  `normative-trace.md` updated.
+
+## [0.9.2] — 2025-Q4
+
+_Current trunk; new work after v0.9.0a0._
+
+### Added
+
+- **SimLab deliberation-mode registry** — `src/hypermind/simlab/modes/`
+  with `ModeSpec` registry and 7 new modes. Default mode (`binary-forecast`)
+  preserves all legacy behaviour.
+  - `governance-committee` — FROST 3-of-5 signed AI governance verdict
+    with named, signed dissents and a NIST AI RMF / EU AI Act / ISO 42001
+    crosswalk. Headline visual: `GovernanceReceipt.vue`.
+  - `what-if-tree` — counterfactual scenario tree over driver variables
+    with InformationGainPlanner suggestions and AntiHerdingTrigger
+    interventions. Headline visual: `ScenarioTree.vue`.
+  - `red-team-stress` — bonded dispute FSM stress-test of any claim.
+    Surfaces §20.1 `adversarial_robustness_score` and a failure-mode
+    tree. Headline visual: `RedTeamReport.vue`.
+  - `post-incident-aar` — HLC time-slider over an incident timeline
+    with witness cosignatures and retroactive source revocation. Headline
+    visual: `AARTimeline.vue`.
+  - `decision-tournament` — round-robin pairwise bracket with Condorcet
+    + Borda winners and sensitivity pivots. Headline visual:
+    `DecisionBracket.vue`.
+  - `crisis-tabletop` — auto-swarm composition by scenario family
+    (cyber-breach / supply-chain / regulatory / market-shock /
+    public-safety) with step-by-step branching. Headline visual:
+    `CrisisTabletop.vue`.
+  - `backtest-harness` — predicted-vs-actual scorecard with calibration
+    curve, per-role drop-one attribution, and improvement
+    recommendations. Headline visual: `BacktestScorecard.vue`.
+  - `evaluation-bench` — **IEEE-grade evaluation**: HyperMind head-to-head
+    against single-LLM and naive-ensemble baselines. Brier with
+    bootstrap 95% CIs, ECE, MCE, reliability diagram, paired t-test,
+    Wilcoxon signed-rank, McNemar test, Cohen's d, drop-one-role
+    ablation, difficulty stratification, and a downloadable IEEE-formatted
+    markdown report. Headline visual: `EvaluationReport.vue`.
+- **`/evidence` top-level route** — evidence-first landing page that
+  pulls the latest evaluation-bench run and embeds the full report.
+  Adds an "Evidence" item to the main nav.
+- `GET /v1/modes` endpoint exposing the registry; SPA reads it on
+  `/new` page load.
+- `webui/src/components/forms/SchemaForm.vue` — schema-driven form
+  renderer for mode-specific inputs (string / textarea / enum /
+  integer / number / boolean / array-of-string).
+- `webui/src/components/DissentList.vue` — reusable signed-dissent
+  viewer used by Governance and Scenario-Tree result views.
+- "Download trace (audit)" button on `SimDetail.vue` — works for
+  every mode; the trace.json file ships any cryptographic receipts
+  (FROST signatures, ed25519 dissent sigs) for offline verification.
+- 7 new example scripts in `examples/scenario_*.py`, one per mode.
+- 7 new test modules in `tests/simlab/test_modes_*.py` (43 new tests).
+- `tests/test_self_heal_cycle3_governance.py` — pins FROST quorum
+  + verdict-bytes binding + dissent-signature reviewer-binding.
+
+### Changed
+
+- `SimConfig` (`src/hypermind/simlab/config.py`) gains `mode` and
+  `mode_payload` fields. Default `mode="binary-forecast"` preserves
+  legacy behaviour. Validator rejects unregistered mode slugs.
+- `scenario.py` `run_persona_sim()` now dispatches to the registry
+  instead of calling `_scenario_impl.main` directly. The trace JSON
+  is stamped with `mode: "<slug>"` so the SPA can branch the result
+  view.
+
+### Fixed
+- _(none yet)_
+
+### Security
+
+- Cryptographic receipts in `governance-committee` mode use real
+  FROST 3-of-5 (RFC 9591) — no simulation. Tampering with the verdict
+  bytes fails verify; substituting reviewer kids on dissents fails
+  verify. Pinned in `tests/test_self_heal_cycle3_governance.py`.
+
+## [0.9.0a0] - 2026-05-02
+
+This release rolls up engineering through ROADMAP §v0.9 — every milestone
+from **v0.5 (versioning bridge)** through **v0.9 (conformance + wire
+freeze scaffolding)** is now 🟢 code-complete. The trailing `a0` is held
+until the §External & human gates clear (IETF I-D submission, IANA
+early-allocation, OpenSSF Sandbox/Incubating acceptance, third-party
+audit, ≥3 pilot MOUs, 7-day chaos-test green run, side-channel CT
+30d-green). At that point we cut `1.0.0` GA.
+
+### Engineering scope rolled up
+- **v0.6 — Project Trust Bundle.** SQLite durable storage (`storage/`),
+  Anthropic + OpenAI + Synthetic responders (`responders/`), ASGI
+  management API (`api/app.py`) with `/v1/stats`, `/v1/claims`,
+  `/v1/audit`, `/v1/consult`, `/v1/calibration`, `/v1/cost`,
+  `/v1/eval/compare`, `/v1/replay`, `/v1/events/stream`. OpenTelemetry
+  exporter (`observability/otel_export.py`). Cost tracker
+  (`observability/cost.py`). Eval comparison + calibration + replay
+  primitives (`eval/`).
+- **v0.7 — Crypto hardening.** FROST hedged-random nonces (RFC 9591
+  §4.2). LAMPS draft-16 §6 domain prefix on composite Ed25519+ML-DSA-65.
+  Dual-vendor PQ adapter (`MlDsaBackend` Protocol + Quantcrypt + OQS
+  backends). CapToken canonical-encoding gate (mirrors A-ALG-BOUND).
+- **v0.8 — Production profile.** Deliberation rounds primitive
+  (`deliberation.py`). Bias detector. Per-agent calibration. Bench harness
+  for publish/dispute/consult (`bench/`). Deploy scaffolding
+  (Dockerfile, docker-compose, Prometheus + Grafana provisioning).
+- **v0.9 — Conformance + wire freeze scaffolding.** Two-impl interop
+  pinned (`tests/test_two_impl_interop.py`). Scenario harness
+  (`harness/`) with assertion library and threat-intel reference scenario.
+  Parallel multi-domain code review (`review/`).
+
+### Added — `hmctl simlab` (multi-sim app)
+- **`hypermind.simlab/`** package: `SimConfig`, `run_persona_sim()`,
+  registry on disk under `~/.hypermind/sims/{id}/`, single-slot worker
+  queue with per-sim `Recorder.scoped(sim_id)` for SSE isolation,
+  standalone ASGI app exposing `GET/POST /v1/sims`, `/v1/sims/{id}`,
+  `/status`, `/events/stream`.
+- **`hmctl simlab`** Click subcommand boots uvicorn + Vue 3 SPA at
+  `localhost:8765` with `--port` / `--host` / `--no-open`.
+- **Vue 3 + Vite SPA** under `webui/`, builds into
+  `src/hypermind/simlab/ui/dist/` and ships in the wheel:
+  Home (hero / 5 primitives / how-it-works / arch stack / recent runs /
+  quickstart), Sims list, New simulation form, Live SSE view,
+  trace detail (6 tabs: Overview / Cast / Agent detail / Peer claims /
+  Questions / Tool calls), Compare two runs (ΔCIQ + per-question delta),
+  Docs (`docs/spec/*.md` + `docs/design/*.md` rendered with sidebar TOC,
+  on-page anchors, syntax highlighting via markdown-it + highlight.js).
+
+### Added — public API
+- `BlockReason` dataclass with `rule_id`, `spec_ref`, `docs_url`,
+  `hint`, `telemetry`; 12 v0.1 rule IDs in `hypermind.rule_ids`.
+- `HyperMindAgent.open(mode="sandbox" | "libp2p")` — canonical factory
+  entry point for v1.0.
+- `hypermind.testing` module — re-exports `testbed_cohort` and other
+  test-only helpers.
+- `hypermind.crypto.hlc.ClockSource` Protocol + `SystemClock` default
+  with skew-detection warning; `hmctl doctor --check-clock`.
+- `PortableInclusionProof` (RFC 6962 §2.1.1 audit-path + leaf-index +
+  tree-size) — independent verifiers no longer need pymerkle.
+- `Recorder.tail(namespace_id=…)` — async generator for SSE; backbone
+  of the live mode.
+
+### Changed
+- **`HyperMindAgent.sandbox` (instance attribute) renamed to
+  `is_sandbox`** to stop shadowing the `HyperMindAgent.sandbox()`
+  classmethod factory. JSON `introspect()` payload unchanged
+  (still keyed `"sandbox"`).
+- **Persona-panel scenario moved into the package** at
+  `src/hypermind/simlab/`; `examples/scenario_persona_panel.py` is now
+  a thin shim that calls `simlab.run_persona_sim(SimConfig)`.
+
+### Removed
+- Static viewers replaced by SimLab: `docs/sim_debugger.html`,
+  `docs/_sim_debugger_template.html`, `docs/eval_dashboard.html`,
+  `docs/crisis_behaviours.html`, `docs/_crisis_viewer_template.html`,
+  `docs/msft_outlook.html`, all stray traces under `docs/`,
+  `examples/serve_persona_sim.py`, `examples/scenario_msft_outlook.py`.
+- Old marketing landing `docs/index.html` — Vue Home (`/`) is now the
+  single canonical landing, served by `hmctl simlab`.
+
+### Fixed
+- **`pyproject.toml` version was `0.1.0a0`** while `__version__` was
+  `0.5.0a0` — `pip show hypermind` and `/v1/healthz` disagreed. Both
+  now report `0.9.0a0`.
+- **CI: `ENVELOPE_SCHEMA_HASH` is now CRLF-invariant.** Both `wire.py`
+  and `reference_verifier/verifier.py` normalise `\r\n` → `\n`.
+- **CI: DCO action SHA pin corrected** (`tisonkun/actions-dco@f1024cd`).
+- **`mypy --strict` clean across the public API surface** (0 errors,
+  down from 25).
+- **Tests: deterministic clock for HLC merge-branch coverage**.
+- **Tests: bump IBLT cells/δ ratio from 6× to 10×** to remove flakiness.
+- **Examples: ruff lint cleanup** + `two_node_tcp.py` uses a poll loop
+  instead of a fixed sleep.
+
+### Security
+- **FROST hedged-random nonces (RFC 9591 §4.2)** replace pure-deterministic
+  §4.1 — closes Cycle-1 finding C1 (Critical). Process restart no longer
+  re-opens Drijvers replay window.
+- **LAMPS draft-16 §6 domain prefix** on Composite Ed25519+ML-DSA-65 —
+  closes stripping/splicing class.
+- **`ENVELOPE_SCHEMA_HASH` bound to actual CDDL bytes** via
+  `importlib.resources` — restores two-implementation interop guarantee.
+- **CapToken canonical-encoding gate** — `from_bytes()` now rejects
+  non-canonical CBOR (mirrors wire.py A-ALG-BOUND).
+- **Secret material redacted** in `__repr__`/`__str__`; pickle refused
+  with `TypeError` on `Keypair`/`FrostShare`.
+- **GitHub Actions pinned to commit SHAs** with Renovate config.
+- **Hash-pinned `requirements.lock`** generated via
+  `pip-compile --generate-hashes`; `pip-audit` CI gate added.
+- **Per-sim `Recorder.scoped()`** isolates SSE event streams between
+  concurrent sims; no cross-sim leakage on `/v1/sims/{id}/events/stream`.
+
+## [0.6.1] - 2026-04-27
+
+Outcome of an internal expert review cycle (10 reviewers, 2 cycles). This release
+lands the Horizon-0 janitorial fixes (~1.5 engineer-weeks of work) plus partial
+Horizon-1 workstreams (WS-A, WS-B, WS-E, WS-F, WS-H, A2-1).
+
+### Security
+- **FROST hedged-random nonces (RFC 9591 §4.2)** replace pure-deterministic §4.1 — closes Cycle-1 finding C1 (Critical). Process restart no longer re-opens Drijvers replay window. (A2-1)
+- **LAMPS draft-16 §6 domain prefix** on Composite Ed25519+ML-DSA-65 signatures — closes stripping/splicing class. (WS-E)
+- **`ENVELOPE_SCHEMA_HASH` bound to actual CDDL bytes** via `importlib.resources` instead of placeholder string literal — restores two-implementation interop guarantee. (H0-4)
+- **CapToken canonical-encoding gate** — `from_bytes()` now rejects non-canonical CBOR (mirrors wire.py A-ALG-BOUND). (H0-8)
+- **Secret material redacted** in `__repr__` and `__str__`; pickle refused with `TypeError` on `Keypair`/`FrostShare`. (H0-3)
+- **GitHub Actions pinned to commit SHAs** with Renovate config; reduces supply-chain attack surface. (H0-1)
+- **Hash-pinned `requirements.lock`** generated via `pip-compile --generate-hashes`; `pip-audit` CI gate added. (WS-A)
+- **Dual-vendor PQ adapter scaffold** (`hypermind.crypto.pq.MlDsaBackend` Protocol + QuantcryptBackend + OqsBackend); FIPS 204 cross-check test for namespaces requiring PQ. (WS-A)
+
+### Added
+- `BlockReason` dataclass with `rule_id`, `spec_ref`, `docs_url`, `hint`, `telemetry`; 12 v0.1 rule IDs in `hypermind.rule_ids`. (H0-6)
+- `HyperMindAgent.open(mode="sandbox" | "libp2p")` — canonical factory entry point for v1.0. (WS-F)
+- `hypermind.testing` module — re-exports `testbed_cohort` and other test-only helpers. (WS-F)
+- `hypermind.crypto.hlc.ClockSource` Protocol + `SystemClock` default with skew-detection warning; `hmctl doctor --check-clock` integration. (WS-H)
+- `PortableInclusionProof` (RFC 6962 §2.1.1 audit-path + leaf-index + tree-size) — independent verifiers no longer need pymerkle. (WS-E)
+- `tests/test_examples_runnable.py` — CI-enforced: every `examples/0N_*.py` must run with exit 0. (H0-5)
+- Stub examples for v0.2 / v0.7 features that were referenced but unimplemented. (H0-5)
+- `fuzz/` directory with `atheris` harnesses for `wire.from_bytes`, `CapToken.from_bytes`, `BloomFilter.from_cbor`, Merkle inclusion proof; nightly `fuzz.yml` workflow. (WS-B)
+- CycloneDX SBOM auto-generation on release tags via `sbom.yml` workflow. (WS-A)
+- `py.typed` marker + strict mypy on public API surface (`__init__.py`, `agent.py`, `wire.py`, `capability.py`, `errors.py`); `typecheck.yml` workflow. (H0-12)
+- Bounded `Recorder` ring buffer (default `maxlen=10_000`) with `records_dropped_total` counter; per-namespace `Recorder.scoped()`. (Reliability prep)
+
+### Changed
+- `pyproject.toml` classifiers now include Python 3.13, `Topic :: Scientific/Engineering :: Artificial Intelligence`, `Topic :: Security :: Cryptography`; SPDX expression `license = "Apache-2.0"`. (H0-2)
+- `HyperMindAgent.from_env()` now emits `PendingDeprecationWarning` — use `.open(mode=...)` instead. (WS-F)
+- `CHARTER.md` adds §2.1.1 Bootstrap Period clause: founder + 1 advisor interim quorum until N≥3 maintainers from ≥2 organisations are seated. Bootstrap Period sunsets at v0.7 release. (H0-10)
+
+### Documentation
+- Internal expert review synthesis (30 findings, 3 horizons, 90-day operational plan, foundation evaluation matrix, v1.0 audit-readiness statement, 8-row production-profile gates) drove this release.
+- `docs/migration/v0.7.md` — migration guide for `from_env` → `open` and namespace re-exports. NEW.
+- `ROADMAP.md` updated through v1.0 with WS-A/B/C/D/E/F/G/H workstream items, A2-1/A2-2/A2-3/A2-4/A2-5 architectural shifts, 8-row v0.8 production-profile gates, formal-methods workstream for v0.9.
+- `CLAUDE.md` updated: active milestone notes Horizon-0 landed; canonical entry point now `HyperMindAgent.open()`.
+
+### Roadmap pulls-left
+- **SCITT IETF submission moved from v0.4 → v0.5** per panel WS-D (CRITICAL): WG charter is solidifying now; submitting at v0.4 is too late to shape it. Target: `draft-zysec-scitt-contested-claims-00` to datatracker within 30 days; IANA early-allocation request within 60 days; SCITT WG shepherd within 90 days.
+- **HLC clock-source contract moved from v0.9 → v0.5** per panel finding C5 — wire-version prerequisite, not freeze-milestone item.
+
+### Known issues
+- Bootstrap Period in effect: only one named maintainer. Critical-path milestone for v0.7 is closing the W6 maintainer #2 hire (panel 90-day plan).
+- libp2p adoption (A2-3) deferred to v0.2 release — `mode="libp2p"` raises `NotImplementedError` until then.
+- Dispute close() witness k-of-n quorum (A2-2) deferred to v0.8 — current single-witness path emits soft warning.
+- Tamarin / ProVerif / hax-rust formal-methods workstream queued for v0.9 (NLnet-fundable, ~18 engineer-months).
+
+### Removed
+None. All changes are additive or deprecation-shim'd to preserve v0.6 callers.
+
+## [0.1.0a0] - 2026-04-27 — Substrate complete (alpha)
+
+### Added
+
+- Full normative specification (18 sections) covering wire format,
+  cryptography, dispute resolution, capabilities, transport, agent
+  lifecycle, observability, MCP integration, CLI, and governance.
+- Reference Python SDK implementing the entire specification.
+- `hmctl` command-line tool for keygen, signing, verification,
+  capability minting, dispute filing, and transparency-log inspection.
+- Hybrid post-quantum signing (classical + PQ in a single envelope).
+- FROST `k`-of-`n` threshold signatures for multi-party authorisation.
+- RFC 6962 Merkle transparency log for the knowledge plane.
+- Bonded dispute FSM with stake, challenge, response, and slashing states.
+- Capability chain-of-trust with delegation, attenuation, and revocation.
+- Per-`(agent, topic)` reputation scoring.
+- `A-ALG-BOUND` wire-level defence binding algorithm identifiers into
+  the signed envelope to defeat algorithm-substitution attacks.
+- 361 tests at 97% coverage.
+- `reference_verifier/` interoperability suite for cross-implementation
+  conformance testing.
+- Six end-to-end example scenarios under `examples/`.
+
+### Security
+
+- Two cycles of self-heal review — 14 fixes pinned in
+  `tests/test_self_heal_cycle1.py` and `tests/test_self_heal_cycle2.py`.
+
+[Unreleased]: https://github.com/ZySec-AI/hypermind/compare/v0.1.0a0...HEAD
+[0.1.0a0]: https://github.com/ZySec-AI/hypermind/releases/tag/v0.1.0a0
+
+---
+
+_Older versions: pre-history; not tagged._