@polycode-projects/the-mechanical-code-talker 0.2.0 → 0.4.0

package/README.md

@@ -36,6 +36,20 @@ statements it can store, retrieve, and answer from later. Text that doesn't
 fit the grammar still gets the tolerant strategies; nothing is rejected for
 being loose, fuzzy, or misspelled.
 
+**Negation and passive.** "Which modules do *not* import X?" computes a bounded
+**set complement** over the graph — and an honestly empty result stays a miss
+rather than a fabricated list (a non-enumerable type, like arbitrary *changes*,
+is refused outright). Reversible-passive questions traverse the right direction:
+"what is imported by Y" and "what does Y import" are understood as opposite
+edges, not the same one.
+
+**Response finishing.** Before an answer is printed it is segmented into typed
+spans — prose versus *protected* entities, paths, numbers, code, provenance, and
+receipts — and a small data-driven grammar pass runs on the prose spans only,
+under a guard that proves the protected spans came through byte-for-byte. Today
+that pass fixes the a/an article defect; broader voice and agreement rules are
+implemented but parked until they earn their place on the benchmark.
+
 ## How it remembers
 
 tmct's memory has two layers, both fed by every parsed request and response and
@@ -51,6 +65,29 @@ With no graph at all, tmct starts empty and remembers what you tell it — the
 vocabulary; a filtered **ConceptNet slice** (CC-BY-SA 4.0) is planned — see
 `ROADMAP.md` Phase 2.
 
+### Provenance and trust
+
+Every fact and text block records **where it came from and when**. Sources are
+first-class individuals — operator chat, a curated corpus, a provider graph, a
+web scrape, a rule-derived entailment — and a fact links back to *all* of them
+(`mgx:derivedFrom` / `mgx:statedBy` / `mgx:canonicalisedFrom`), timestamped with
+`mgx:createdAt`. From those links tmct computes a **deterministic, explainable
+trust score** — a source-type prior combined with corroboration (how many
+independent sources agree) and recency — never hand-set, always traceable to its
+inputs. Retrieval then ranks by **relevance × trust**, so a corroborated,
+operator-stated fact outranks a lone web scrape on the same question. When two
+trusted sources *disagree*, the `/memory` inspector shows **both sides with their
+provenance** rather than silently picking a winner.
+
+### Speculative inference (a maintenance job, not a chat cost)
+
+`tmct syllogise [--depth n] [--budget n]` is an offline, bounded, deterministic
+batch that forward-chains the memory's `rdfs:subClassOf` closure into new
+**entailed** facts — pre-deriving what the trusted sources already imply. It runs
+once automatically after seeding and on demand; the entailed facts are
+**low-trust and retractable** (never outranking a stated fact) and this never runs
+on the chat's hot path.
+
 ## What tmct deliberately is NOT
 
 - **It is not an indexer.** tmct keeps no codebase index of its own. It
@@ -58,7 +95,9 @@ vocabulary; a filtered **ConceptNet slice** (CC-BY-SA 4.0) is planned — see
   a code graph is out of scope. tmct's job is the *conversation*.
 - **It is not a reasoning model.** Where it "reasons", it does so by
   *calculation* surfaced as prose ("there are a lot of tests for a codebase of
-  that size") — deterministic, explainable, cheap. There is **no LLM anywhere
+  that size") — deterministic, explainable, cheap. Even its forward-chaining
+  entailment (`tmct syllogise`) is mechanical OWL rule materialization applied
+  offline, rule-by-rule and retractable — not an LLM. There is **no LLM anywhere
   in the product**. (An LLM-as-judge exists only in the offline eval harness
   that tunes tmct — see `SKILL_TUNING_CYCLE.md` — never in the product path.)
 - **It never guesses silently.** When it cannot resolve your question it says
@@ -70,10 +109,18 @@ vocabulary; a filtered **ConceptNet slice** (CC-BY-SA 4.0) is planned — see
 npm install -g @polycode-projects/the-mechanical-code-talker
 tmct                                  # bare = chat (the headline)
 tmct chat --repo /abs/path/to/repo    # chat over a specific repo's graph
+tmct init                             # scaffold .tmct/, tmct.toml, seed + provenance
+tmct syllogise                        # offline: pre-derive entailed facts (maintenance)
 ```
 
-Inside the chat: `/help` lists commands, `/exit` leaves. `TMCT_GRAPH_FILE`
-overrides the graph location.
+Inside the chat: `/help` lists commands, `/memory` inspects what tmct remembers
+(grouped by OWL class, with provenance and any contradictions), `/exit` leaves.
+`TMCT_GRAPH_FILE` overrides the graph location.
+
+`tmct init` is the onboarding surface for the repository interface below: it
+creates the `.tmct/` directory, writes the externalized `tmct.toml`
+configuration, seeds the tier-1 corpus, and records provenance — a host package
+or a bare user gets a working install in one command.
 
 > Install-size note: tmct depends on wink-nlp's deterministic English language
 > model (~3.8 MB installed). That model is a lookup table, not an LLM.
@@ -88,6 +135,33 @@ The `exports` map and the chat primitives (`ask`, `resolveObject`,
 `relationKind`, `impactClosure`, `dispatchTool`, `fetchEntities`) are the
 extension surface.
 
+## The repository interface
+
+tmct is not an indexer, so it consumes a graph through a typed contract any
+producer can implement. That contract is now first-class: a **versioned
+(1.0.0), OWL-grounded, machine-readable service definition**
+(`docs/repository-interface.md` + a JSON schema) of every service, its
+arguments, result types, and error contract. A **miss is a value, not a throw** —
+the interface models "no answer" explicitly. tmct ships **reference providers**
+(a fixture graph and the empty/bootstrap graph) that implement every service,
+and a **runnable conformance suite**: tmct's own providers pass it in `npm test`,
+and any external graph producer (seonix first) runs the *same* suite against its
+native implementation to claim conformance. Conformance is the suite, not prose.
+The design deliberately inverts the original relationship — tmct was lifted out
+of seonix; now seonix reorients as a *user* that imports the tmct library and
+exposes its graph to tmct as a service, with the LLM agent staying outside tmct
+exactly as the no-LLM ethos requires.
+
+## Security and supply chain
+
+tmct is $0 to run and meant to be trusted offline, so the supply chain is
+hardened accordingly: CI runs **SAST and secret detection**, a **nightly
+`npm audit` + OSV-Scanner** job watches dependencies, releases are published with
+**npm provenance** (`--provenance`), and there is a coordinated-disclosure
+`SECURITY.md` policy. The content-address hash is single-sourced in
+`src/hash.mjs` so the cross-version-stable fact-id contract has exactly one
+definition.
+
 ## Provenance
 
 tmct began as a whole-package lift of the seonix chat surface (v0.1.0, then

package/ROADMAP.md

@@ -115,8 +115,11 @@ Keep the `exports` map and the primitives stable and documented as the
 internals are refactored.
 
 ### Shell work
-- **OpenTUI console shell** around `runTurn`; readline `runChat` stays as
-  `--plain` and as the test surface.
+- **Ink console shell** (`src/tui/app.mjs`, ink + react, no build step) around
+  the shared session sink; readline `runChat` stays as `--plain` and as the
+  test surface. *(Decision: OpenTUI ruled out — `@opentui/core` depends on Bun
+  FFI (`bun-ffi-structs`, native Zig renderer), not Node-clean; revisit when it
+  runs under plain Node.)*
 - Fold the surviving `bin/cli.mjs` arms into `bin/tmct.mjs`; delete `cli.mjs`.
 
 ---
@@ -170,6 +173,30 @@ in `SKILL_TUNING_CYCLE.md`:
   harness only** — the product stays no-LLM;
 - `CHATBENCH_0NN.md` artifacts and an autonomous cycle loop (no hard pause;
   each cycle logs its ranked decision menu and continues).
+- **The graded benchmark** (case-set v2, operator-specified 2026-07-04): a
+  scaled ladder fitted to HUMAN LANGUAGE STANDARDS, not AI-benchmark mechanics
+  (bAbI explicitly rejected — it tests expected-AI mechanics and overfits the
+  same way our own cases would). Every case carries a **CEFR band (A1–C2)** +
+  a **construction specialization** tag (TROG-2/CELF-5 style blocks, adapted:
+  naming/vocabulary, SVO queries, pronoun binding, reversible/passive,
+  relative/embedded clauses, quantifiers+counting, negation, temporal,
+  coordination/compositional, multi-turn discourse reference, declarative
+  assert+recall). Multiple questions per grade × specialization with a little
+  overlap — authored as a POOL ~10× the per-run need (deterministic generator;
+  each run samples ~10% stratified, ≥5 items per populated cell, recorded
+  seed), with COMBINATION cells alongside single-area cells so weakness is
+  attributable to an area alone or to a specific pairing.
+  **Ladder gating:** advanced grades are SKIPPED until every grade
+  below passes reliably; when a grade reaches unit-test-level reliability its
+  cases become ALWAYS-RUN deterministic tests (judge-free, promoted alongside
+  test/showcase.test.mjs). Winograd/WinoGrande items stay as permanent
+  ceiling markers; CHILDES as naturalistic easy-band input source. Licence
+  rule: TROG/CELF are commercial — borrow the construction taxonomy and
+  grading structure, author ORIGINAL items, never copy instrument content.
+  A case at 0% is a ceiling marker, not a failure.
+- **Retained showcase** (landed): the five most complex achieved sequences are
+  frozen as unit-timescale regression tests in `test/showcase.test.mjs`; a
+  showcase regression voids a cycle PASS regardless of the mean.
 
 Inside this loop, two earlier ambitions become **tuning levers** rather than
 standalone features:
@@ -183,7 +210,7 @@ calculations, not inference: deterministic, explainable, cheap.
 Let tmct run linters or tests to observe whether something actually worked,
 reporting the observation — measurement, not reasoning.
 
-### Item 11 — Formal logical reasoning over the ontology (Prolog / Progol) — exploratory, gated
+### Item 11 — Formal logical reasoning over the ontology (Prolog / Progol) — exploratory, gated; matured into Phase LATER tier-5 "entailment-on-miss"
 Apply real rules of inference (modus tollens, etc.) to formulas extracted from
 parsed prose, checked against the axioms in the ontology (item 3) — a step
 beyond item 5's arithmetic. The sketch: map OWL constructs into templates over
@@ -199,6 +226,392 @@ then use Progol to theorem-prove against parsed prose".)*
 
 ---
 
+## Phase 4 — The wiring wave (operator-directed 2026-07-04)
+
+Five subsystems are built, tested, and consumed by NOTHING in the answer path — they measured
+zero on case-set v1 because no case could see them. This phase wires each into answering, with
+**unit tests at the seam AND graded-benchmark cells that measure it** (the graded pool creates
+the cases that make each lever visible). Wired as one operator-directed wave; cycle-level
+attribution resumes per-lever afterwards.
+
+| # | Wiring | Seam | Unit test | Bench coverage |
+|---|---|---|---|---|
+| W1 | **Templates → render path** | answer rendering consumes `data/templates/responses.jsonl` via `src/corpus/templates.mjs` instead of hardcoded strings (same output first — byte-stable swap — then variation) | render parity + slot lint | every existing case re-measures the swap; `via:"template"` provenance |
+| W2 | **retrieveBlocks → miss path** | a bare-question miss consults the memory block index before the honest miss; a hit answers with the recalled block + provenance ("you asked this on …") | recall hit/miss seam | memory-recall cells (mr-asked-before flips) |
+| W3 | **seedMemory → bootstrap** | first run in a graph-less repo seeds a capped corpus slice (limit ~500) into `.tmct/memory/`; banner says so honestly | seeded-bootstrap test | bootstrap-empty + vocabulary cells ("what is a cache?") |
+| W4 | **Asserted Facts → answers** | "what is a module?" / "is a module a component?" consults remembered `rdfs:subClassOf`/`rdf:type` facts alongside the code graph, cited with provenance | fact-lookup seam | assert-recall cells |
+| W5 | **Corpus on-demand** | unknown-term misses may consult the corpus slice (local first; network tier only behind an explicit flag) | on-demand seam, offline-degrades test | naming-vocabulary cells at higher grades |
+
+Answer-path **provenance** lands with W1: every turn record carries `via`
+(composed | template | count | recall | fact | corpus) — the field the dual-banding
+benchmark (Phase 5) and the memory inspector read.
+
+### Corpus tiering policy (the committed/seeded/on-demand cutoff)
+
+- **Tier 1 — committed & shipped in the npm tarball**: small, load-bearing, licence-clean,
+  diffable — the lexicon, response templates, phrasebook, the relation→OWL map, and the CORE
+  ConceptNet slice. Budget: **~2 MB total tarball**; rule: what the product needs to be useful
+  offline out of the box.
+- **Tier 2 — fetched at seed time into install-local folders** (`.tmct/corpus/` per repo, or a
+  user-level cache): growable corpora — extended ConceptNet neighbourhoods, acquired template
+  libraries (Phase 5), any corpus > ~2 MB. Fetched once by `tmct seed` (or first bootstrap with
+  consent), checksummed, provenance-recorded, never committed.
+- **Tier 3 — on-demand at question time**: unbounded/live sources (ConceptNet API for unknown
+  terms, paper phrase-mining), consulted ONLY behind an explicit opt-in flag, cached down into
+  tier 2 after use. **Network failure degrades to the honest miss** — the $0-offline default is
+  inviolable; tiers 2-3 are additive, never required.
+- **Tier 4 — learn-on-miss (Phase LATER, not now)**: the term IS in the lexicon and the query
+  built cleanly, but the graph+memory+corpus all return nothing → web search, clean the fetched
+  text into tmct's own dialect (the ACE-ish controlled grammar), store on disk (tier 2), ingest,
+  THEN answer — the full acquisition loop at question time. See Phase LATER.
+
+### Memory inspection (seeing into the memory)
+
+Graph-vis exploration hasn't earned its keep; the in-ethos answer is TEXT. A `/memory` chat
+command + `tmct memory` CLI: the memory graph grouped by **OWL superclass** (Utterance, Fact,
+Session; code classes when present), counts per class with **balanced samples scaled to class
+size** (log-scaled so a 10,000-fact class shows ~8 exemplars and a 3-session class shows all 3),
+top facts ranked by provenance breadth (corpus+chat-agreed facts first), recent utterance pairs,
+and the block-index summary (blocks, tokens, top PageRank blocks). Same renderer serves
+`/stats`-style terse and `why`-style verbose.
+
+## Phase 5 — The cycle-4+ tuning arc (near-term: make the floor reliable)
+
+> **STATUS: in progress — the two headline B1 levers shipped.** Negation as a bounded SET
+> COMPLEMENT ("which X do not <verb> Y") and reversible-passive traversal ("X is imported by Y")
+> are live; the harness meta-fixes and the rest of the ranked lever board continue.
+
+> Detailed plans: **PLAN_CYCLE_4.md** (the tuning arc) and **PLAN_DEPENDENCY_STRATEGY.md**
+> (the pre-build dependency audit — verdict: no dep changes now; a standing adoption
+> register + avoid-list for phases 6-9; two near-term actions).
+
+The immediate work: drive the graded benchmark up the CEFR ladder, one lever per cycle, per
+`SKILL_TUNING_CYCLE.md`. Cycle 3 (post-wiring-wave, CHATBENCH_003) gave the first full-spectrum
+reading — A1 1.72 / A2 1.70 / **B1 0.77 (the cliff)** / B2 0.97 / C1 1.07 / C2 0.69 — and two
+META-fixes gate everything:
+
+- **Meta-1 — fix the harness artifact BEFORE trusting groundedness.** The judge systematically
+  scores TRUTHFUL product output (/describe attributes, recall frames, session ids) as
+  fabrication because `FIXTURE_CONTEXT` omits the detail the product legitimately emits.
+  Measurement integrity: lands in its own cycle, re-measured, before any product lever — else
+  every groundedness delta is confounded. A harness correction, logged as such (like cycle-2's
+  H1a/H1b), not a product change.
+- **Meta-2 — the ladder rule: get B1 reliable before judging C-grades.** Don't pay to judge a
+  ceiling while the floor leaks. A/B grades carry the judged spend; C1/C2 stay tier-1-only
+  ceiling markers, judged only occasionally to confirm they're still ceilings (the existing
+  `--ladder`/`--grade` flags).
+
+Then the product levers, one per cycle, in ranked order: **(1) B1 negation operator** — the
+deepest gap, the engine tokenizes "not"/"don't" as an entity ("no module matching 'not' found");
+add set-complement to traversal; predicted B1 0.77→~1.05, ~10 hard fails cleared. **(2)
+reversible-passive** ("is imported by" reverses edge direction). **(3) under-covered pool
+growth** (B1 pronoun/temporal, C1 temporal — instrument fix, parallelizable). **(4)
+assert-recall read-back**, **(5) quantifier+temporal composition**, **(6) the help-text honesty
+leak** (hardcoded examples naming non-fixture entities — a real product fabrication, distinct
+from meta-1's harness artifact), **C2 ceiling LAST**. Operator decision 2026-07-05: **do all of
+them**. Exit criterion in the plan (roughly: B1 grade mean ≥ ~1.5 with all cells dual-draw
+agreeing unlocks C-grade judging).
+
+## Near-term actions (from the dependency audit, PLAN_DEPENDENCY_STRATEGY.md)
+
+> **STATUS: both shipped.** The wink browser-loader seam is added (shared model loader with a
+> browser registration path) and `fnv1a` is single-sourced into `src/hash.mjs`.
+
+Two concrete, low-risk actions the audit surfaced — not features, not dep changes:
+
+1. **Fix wink's `createRequire` browser-loader gap** — the wink model IS the browser build, but
+   our adapters load it via `createRequire(import.meta.url)` (`ask-nlp.mjs:29`, `prose-nlp.mjs:31`),
+   which is Node-only. Browser mode needs a bundler `import` path. **A Phase 8 (browser-mode)
+   blocker** — budget it into `PLAN_REPOSITORY_INTERFACE.md`; it is a wiring fix, not a dependency
+   change (the model is already browser-capable).
+2. **Single-source `fnv1a`** — extract the content-address hash to one `src/hash.mjs` so the
+   cross-version-stable fact-id contract has a single definition. Trivial refactor, do any time;
+   no dependency (the audit confirmed home-grown FNV-1a is the correct choice — sync + browser +
+   version-stable, which every library candidate fails).
+
+## Provenance & trust — the unified source-link primitive (cross-cutting)
+
+> **STATUS: shipped.** `mgx:createdAt` universal; `Source` first-class individuals linked by
+> `mgx:derivedFrom` / `mgx:statedBy` / `mgx:canonicalisedFrom`; a deterministic `computeTrust`
+> (source-type prior × corroboration × recency); retrieval weighted by relevance × trust; the
+> `/memory` inspector surfaces contradictions with provenance. Legacy `mgx:factProvenance` kept
+> as a compat shim.
+
+> Detailed plan: **PLAN_PROVENANCE_TRUST.md**.
+
+*(Operator-specified 2026-07-05, from the observation that Phase-6 canonicalise-and-link,
+tier-4 learn-on-miss, and the ConceptNet slice all share one shape: raw source preserved,
+derived form linked back.)* Promote that shape to a FIRST-CLASS primitive used everywhere a
+fact enters memory:
+
+- **Every fact/block carries a `Source` and a link to it** — one predicate family
+  (`mgx:derivedFrom` / `mgx:canonicalisedFrom` / `mgx:statedBy`) instead of the current
+  per-writer `mgx:factProvenance` string. Sources are first-class individuals (class `Source`:
+  operator-chat, corpus:conceptnet, learned:web:<url>, entailed:<rule>, provider:seonix), so a
+  fact can cite MANY sources (the existing "|"-union becomes real edges).
+- **Everything created is TIMESTAMPED** (`mgx:createdAt`), universally — Facts don't carry one
+  today (only Utterances do), a Phase-6-trust gap to close: recency is a trust input and the
+  novelty signal (below) needs it. Backfill on write; the timestamp is itself provenance.
+- **Calculable trust scores per source**, deterministic and explainable: a source-type prior
+  (operator > provider graph > curated corpus > web > unverified entailment) combined with
+  corroboration (how many independent sources assert the same fact — the union already tells us)
+  and recency/agreement signals. Trust is a computed attribute, never hand-set, always
+  traceable to its inputs.
+- **Trust as RETRIEVAL WEIGHTING**: `retrieveBlocks` / fact lookup / the memory inspector rank
+  by relevance × trust, not relevance alone — a corroborated operator-stated fact outranks a
+  lone web scrape on the same query. Contradiction becomes visible (two high-trust sources
+  disagree → surface both with their provenance, never silently pick).
+- **Feeds tier-5**: the Syllogist's entailed facts get a derived trust (min/product of premise
+  trusts × rule confidence) — a conclusion is only as trustworthy as its weakest premise, and
+  that number is computed, not asserted.
+
+## Phase 6 — Formulaic competence: the template-acquisition learning loop
+
+> **STATUS: shipped.** A technical (C1) register of templates and productive/performance
+> dual-banding (computed from the `via` provenance) are live in the benchmark.
+
+> Detailed plan: **`PLAN_FORMULAIC_COMPETENCE.md`**.
+
+The operator's insight upgraded to the strategy: a consistently-failed C1/C2 graded cell whose
+answer EXISTS as a stable phrasing in technical prose is not a ceiling — it is a
+**template-acquisition lever**. tmct learns the way human learners do: formulaic chunks first
+(Wray's formulaic sequences), productive competence later.
+
+- **Dual banding**: every graded score splits into a **productive band** (composed answers only)
+  and a **performance band** (templates allowed), computed from the `via` provenance (W1). The
+  band GAP is a first-class metric: how much fluency is memorized vs generated.
+- **Template-lane benchmarking**: cases that target templated capability are TAGGED as such —
+  a template-carried C1 pass counts in the performance band and never inflates the productive
+  band; template-lane cells get their own agreement/reliability treatment (they are additional
+  benchmarking, not replacements — a level we would otherwise expect to fail at is being
+  deliberately faked, and the bench must say so).
+- **The shopping list**: each cycle, the write-up extracts consistently-failed C1/C2 cells and
+  ranks them by template-acquirability (does a stable technical-prose phrasing exist? is the
+  slot structure mechanical — counts, comparisons, provenance we already compute?). Acquiring
+  the template IS the lever; the graded bench measures the flip in the performance band.
+- **Mechanical conclusions at paragraph grade**: counting + comparison + superlatives (item 5)
+  composed through acquired C1-register templates — "X has 340 tests across 12 suites, unusually
+  dense for a codebase this size" — tech-domain answers can be genuinely advanced while the CEFR
+  banding tells us honestly how good the conversation AROUND them is.
+- **Generalization path**: fixed tech domain first (templates hand-picked from technical-paper
+  register); then template acquisition generalizes — mining candidate templates from corpus
+  blocks (tier-2), scored by slot-fillability, promoted into `data/templates/` with provenance.
+
+## Phase 7 — Response finishing: the grammar pass (tone of voice dropped for now)
+
+> **STATUS: shipped.** Answers segment into typed spans (prose vs protected); the grammar-rule
+> pass runs on prose spans only under a protected-span invariance guard. The a/an article fix is
+> active; broader voice/agreement rules are implemented-but-parked.
+
+> Detailed plan: **`PLAN_RESPONSE_FINISHING.md`**.
+
+*(Refined 2026-07-05; decisions settled with the operator. Fact invariance is achieved by
+CONSTRUCTION, not by hope — finishing operates over a SEGMENTED answer, never a raw string.
+Tone-of-voice synonym substitution is DROPPED: once every term with technical significance is
+protected — entities, paths, vocabulary, receipts, provenance — the substitutable surface is
+mostly connectives: high accuracy risk, thin reward. "Keen on the trickery to make a helpful
+product, but not at the cost of accuracy." Moved to Phase LATER should a provably-safe subset
+ever emerge.)*
+
+- **The segmentation IR (the foundation, lever 1)**: every answer becomes a list of typed spans
+  before it becomes text — `prose` vs PROTECTED (`entity`, `path`, `number`, `code`,
+  `provenance`, `receipt`). Protected spans are byte-copied through finishing; only prose spans
+  are ever touched. The W1 template renderer is already slot-aware (slots ARE the protected
+  spans); composed renders adopt segmentation progressively via a conservative masker. Phase 5's
+  dual banding reads the same spans.
+- **Grammar pass (lever 2)**: a data-driven rule table (TOML, item-7 formats) over prose spans —
+  article selection ("a artifact" → "an artifact", an observed defect class), subject–verb
+  agreement against slot plurality, capitalization, list/terminal punctuation. Grammar
+  corrections IMPROVE accuracy (they fix our own generated defects); that is why they survive
+  the tone cut. Neutral behavior is byte-stable except where a rule fixes a genuine defect —
+  each rule lands as a bench-measured lever.
+- **Memory decision (settled)**: memory stores BOTH — the **as-spoken** turns live as larger
+  prose blocks on the graph (the honest record), and the **canonical** form is derived and
+  LINKED to its source prose blocks (canonise + link, never replace). Recall and folding read
+  canonical; provenance walks back to as-spoken.
+- **Verification**: unit invariance checker (protected-span multiset identical pre/post) +
+  golden files per rule + the graded bench measuring each grammar rule as a lever.
+
+## Phase 8 — The Repository Interface (seonix inverts to a tmct user)
+
+> **STATUS: shipped.** A versioned (1.0.0), OWL-grounded service contract
+> (`docs/repository-interface.md` + `.schema.json`); a typed graph-service with a first-class
+> miss contract (a miss is a value, not a throw); fixture + bootstrap reference providers; a
+> runnable conformance/compatibility suite; and `tmct init` (scaffold `.tmct/`, `tmct.toml`,
+> tier-1 seed, provenance).
+
+> Detailed plan: **`PLAN_REPOSITORY_INTERFACE.md`**.
+
+*(Operator-specified 2026-07-05; upgraded from research item to a build phase. tmct was spun OUT
+of seonix; this inverts the relationship: seonix reorients as a USER that imports the tmct
+library and exposes its graph to tmct as a typed service. Grows item 14's provider adapter from
+a passive payload loader into the product's primary integration surface.)*
+
+**Phase deliverables — define, reference-implement, and test the interface:**
+1. **The interface DEFINITION**: the typed, OWL-grounded service contract as a versioned
+   document + machine-readable shape (docs/repository-interface.md + a JSON-schema/typedef of
+   every service, its arguments, result types, and error contract) — tmct owns and versions it.
+2. **A REFERENCE IMPLEMENTATION tmct ships itself**: the in-repo provider (fixture graph +
+   bootstrap/empty graph) implementing EVERY service of the interface — the executable
+   specification any external producer reads first.
+3. **The contract test suite (the compatibility kit)**: a runnable suite any implementation is
+   tested against — tmct's reference implementation passes it in `npm test`; seonix runs the
+   SAME suite against its native implementation to claim conformance. Conformance = the suite,
+   not prose.
+4. **The session-handle lifecycle**, implemented: create/dispose context handles (focus, last,
+   memory dir, lexicon), provider-owned caching, documented re-entrancy — proven by the
+   contract suite's concurrent-session cases.
+5. **`tmct init`** shipped as part of this phase (it is the interface's onboarding surface).
+
+- **tmct defines the adapter shape** — not the producer. Rationale: tmct is the brittle side
+  (query interpretation), so it must own and optimize around a STABLE interface; because the
+  vocabulary is OWL-grounded, the human/code world is already quantized into types both sides
+  understand — the interface is built from those shared types, not ad-hoc JSON.
+- **A rich instruction set, translated from what seonix already exposes**: survey seonix's
+  native tool surface (describe / members / subclasses / impact / callers / callees / tests-for /
+  untested / history / exports / architecture / search / context / snippet / locate / digest —
+  the dispatchTool catalog tmct carried at the lift) and translate it into tmct's language as
+  the REPOSITORY INTERFACE: a consistent set of typed services any graph producer implements
+  natively (seonix first; the empty/bootstrap and fixture providers are degenerate
+  implementations tmct ships itself).
+- **The flow** (LLM-agent front door): Claude Code et al. is briefed to use seonix → when the
+  agent judges it useful, seonix's "ask" tools pass NATURAL LANGUAGE to tmct → seonix calls the
+  tmct library in-process with the query PLUS a callbacks object (functions implementing the
+  repository interface over its native graph) → tmct resolves the query mechanically, calling
+  back into seonix's services for graph truth → results return through seonix to the LLM agent.
+  The mechanical interpreter becomes the NL front-end for any agent-facing graph tool; the LLM
+  stays outside tmct, exactly as the no-LLM ethos requires.
+- **In-process lifecycle research (the hard part)**: seonix calls tmct directly, and the
+  interface is wider than the in-house chat — so define explicitly what is HELD IN MEMORY
+  between function calls: an explicit session/context handle (focus, last-answer, memory dir,
+  loaded lexicon) created and disposed by the caller instead of process-global state; graph
+  caching delegated to the provider (tmct never caches provider truth — the known source.mjs
+  process-cache staleness in long-lived servers becomes the provider's concern, by contract);
+  re-entrancy and concurrent-session guarantees documented per service.
+- **seonix chat becomes tmct chat + a pointer**: seonix's chat surface loads tmct's chat with
+  the repository-interface handle — one chat implementation, N graph backends.
+- **Browser mode**: the same inversion works in seonix's browser/code-browser surface — seonix
+  finds its own graph (it already ships one to the page) and embeds an OFF-THE-SHELF tmct: the
+  engine core (interpret / ask / render, lexicon, templates) is already pure JS with no
+  node-only dependency — wink's eng-lite-web-model is literally the browser build — so the
+  repository interface + a browser storage seam for memory (or provider-supplied persistence)
+  is all that separates the npm package from running in the page. The fs/readline/child_process
+  seams stay node-side; the browser gets the library surface, not the shell.
+- **Distribution: `tmct init`** — a CLI command that initializes a local directory for tmct:
+  seeds/links the text corpuses (tier-1/2 policy applies), writes the externalized configuration
+  (tmct.toml — the seonix.toml pattern), creates `.tmct/`, and records provenance — so a host
+  package (seonix) or a bare user gets a working install with one command.
+
+## Phase 9 — Speculative inference: a step toward the Syllogist
+
+> **STATUS: shipped.** `tmct syllogise [--depth n] [--budget n]` — an offline, bounded,
+> deterministic maintenance job that forward-chains the `rdfs:subClassOf` closure into low-trust,
+> retractable entailed facts; runs once after seeding, never on the chat hot path.
+
+*(Operator-specified 2026-07-05. Tier-5 entailment answers a MISS on demand; this is the step
+before it — PROACTIVELY extending memory with inferences that will be useful later, forward and
+backward chaining over the OWL base during idle/fold time rather than at query time.)*
+
+> Detailed plan: **`PLAN_SPECULATIVE_INFERENCE.md`**.
+
+**A maintenance job, not a query-time cost.** Speculative inference runs as an explicit
+batch — `npx tmct syllogise --depth <N>` (default depth bounded, e.g. 32) — and **once
+automatically after seeding** (the W3 bootstrap seed is the natural trigger: a fresh corpus is
+exactly when pre-deriving the useful closure pays off). Never on the chat's hot path.
+
+**The selection criterion, sharpened by the operator (2026-07-05):** the guiding question is
+*"what do the assertions of the sources I TRUST allow me to infer about this topic that is of
+RELEVANCE"* — so **novelty × trust is the primary driver**, not an afterthought: the pass walks
+outward from high-trust premises (the provenance primitive) toward novel, relevant conclusions,
+timestamping each so recency and novelty stay computable. The mechanics are the easy half
+(bounded forward chaining materializes entailments; backward chaining from frequent query shapes
+pre-derives likely answers). **The residual hard half is still the FRAME PROBLEM / relevance
+realization — unsolved in the general case and not pretended otherwise;** trust+novelty+relevance
+are the tractable approximation, not a solution. The plan's job is
+to make it TRACTABLE in tmct's narrow, closed world, not to solve it: usefulness is approximated
+from what the system actually gets asked (query-shape frequency), what connects to recent focus,
+what a cheap forward step yields that isn't already stored, and — the honest guardrail — a hard
+budget (inference is bounded, its output trust-scored via the provenance primitive, and anything
+speculative is retractable and never outranks a stated fact). Everything else is deferred to the
+plan's open questions, where the relevance problem is named as the open research risk it is.
+
+### Open-source the ACE-OWL parser as a standalone library
+> **STATUS: deferred follow-up** — not yet started; still gated on the Phase 8 library-surface
+> work settling the extraction boundary. See `PLAN_OSS_ACE_PARSER.md`.
+
+*(Operator-specified 2026-07-05, from the dependency audit's publish-not-replace finding.)* The
+pure-JS, ESM, dependency-free ACE-OWL controlled-grammar parser (`src/grammar/ace.mjs` +
+`lexicon.mjs`) that turns controlled-English sentences into OWL-labelled triples is a RARE thing:
+the reference implementation (APE) is GPL + SWI-Prolog (native), so there is no permissive,
+browser-capable, npm-installable ACE→OWL parser in the JS ecosystem. tmct's is exactly that.
+Extract it to its own MPL-2.0 package (tmct depends on it back), so the wider RDF/OWL/semantic-web
+JS community gains a controlled-natural-language front-end that runs in the browser. Gated on the
+Repository Interface library-surface work (Phase 8) settling the extraction boundary; see
+`PLAN_OSS_ACE_PARSER.md`. Sibling publish-candidates (the bounded-Damerau fuzzy matcher, the
+PageRank+IDF block ranker) may follow the same path if there is demand.
+
+## Phase LATER — recognized, deferred, not now
+
+Features we have deliberately shaped seams for but will not build until the phases above have
+earned them:
+
+### Tone-of-voice adaptation (dropped from Phase 6, 2026-07-05)
+Per-voice synonym/phrase substitution over prose spans. Dropped because tmct's protected-span
+analysis leaves too little safely-substitutable text: any term with technical significance is
+untouchable, and accuracy outranks helpfulness trickery. Revisit only if a provably-safe
+substitutable subset emerges (e.g. connective-only voice profiles, or per-voice template
+overrides authored as whole alternatives rather than substitutions). The grammar-preference
+half of the idea survives inside Phase 6's rule table.
+
+### Tier-4 corpus: learn-on-miss acquisition
+The strongest miss signal tmct can emit is: *lexicon term recognized, query built cleanly,
+zero matches anywhere* — the question was well-formed and the knowledge is simply absent. The
+tier-4 loop answers it by learning: web search on the resolved term → clean the fetched text
+into tmct's own dialect (normalize into the ACE-OWL controlled grammar; whatever survives the
+grammar becomes Facts, whatever doesn't becomes tier-2 text blocks under the PageRank index) →
+store on disk with source provenance → ingest → answer the original question from the newly
+learned material, citing what was just learned and from where. Strictly opt-in, network tier
+rules apply (offline default inviolable; failure degrades to the honest miss). Prerequisites:
+W1-W5 wired and measured, the Phase-5 template/dialect cleaning machinery (the "clean dialect"
+IS the acquisition format), and a provenance-trust policy for web-sourced facts (never blended
+silently with graph/operator facts — the `via`/provenance discipline extends to "learned:web").
+
+### Tier-5: entailment-on-miss — "the Syllogist" (deductive inference over the OWL base)
+*(Item 11 matured from exploratory sketch to a designed tier; the "theorem-prove against
+parsed prose" thread of the original code-talker ideas.)*
+
+**The concept, classically:** answering from the **deductive closure** of a knowledge base —
+KB ⊨ φ ("the knowledge base *entails* φ") — content that is nowhere ASSERTED in the graph,
+memory, or corpus, but is a logical CONSEQUENCE of what is. Deductive inference (modus ponens,
+modus tollens, syllogistic chains) predates ELIZA by ~2,300 years (Aristotle's syllogisms →
+Frege's predicate logic → Robinson's resolution principle 1965 → Kowalski's "logic as a
+programming language" → Prolog's SLD resolution; on the rules side, forward-chaining production
+systems and the Rete algorithm; on the OWL side, description-logic reasoners and the RDFS/OWL
+entailment regimes). tmct's version: a well-formed query misses everywhere → run the inference
+layer over the OWL-encoded facts + axioms → if the answer is ENTAILED, materialize it as a Fact
+with `via:"entailed"` and a **proof-chain provenance** (the applied rules + premise facts,
+renderable as a chain of thought in words: "every cache is a store; every store is a component;
+so a cache is a component") → the same query now yields an answer that shows its derivation.
+
+**Worked shape (modus tollens over the code graph):** axiom "every tested module is covered by
+a suite"; fact "m.mjs is covered by no suite" ⊨ "m.mjs is not tested" — never asserted,
+honestly derived, provenance = the two premises + the rule name.
+
+**Engine choice (the Prolog / graph-query question):** the classical candidates are embedded
+Prolog (SLD, backward-chaining, item 11's original sketch), a graph query syntax (SPARQL under
+entailment regimes / datalog / openCypher), or a description-logic tableau reasoner. The
+recommended target is **OWL 2 RL** — the profile DESIGNED to be implemented as forward-chaining
+rules (datalog-style semi-naive materialization, polynomial, decidable): pure-JS implementable,
+mechanical, explainable rule-by-rule — exactly in ethos. Prolog-style backward chaining stays
+the fallback for query-time-only derivation if materialization proves too eager. Progol/ILP
+(learning NEW rules from examples) remains a separate, further-out spike.
+
+**Gates:** the full-domain lexicon + OWL encoding in a queriable structure (Phases 2+4+5 and
+tier-4's acquisition feed it), the provenance-trust policy (entailed facts must never silently
+mix with asserted ones — a wrong axiom poisons the closure, so entailments are retractable by
+provenance), and bench cells that measure inference specifically (premises in, conclusion
+asked, derivation shown).
+
 ## Explicitly out of scope (for now)
 
 - No AWS, no benchmark rig — tmct is a published npm library + CLI with a