Fix training hyperparameters (seq=2048, effective batch=8); add Lead AI Architect; document 57 agent capabilities trained in W2.7

README.md

@@ -56,7 +56,7 @@ This is a **from-scratch sovereign build**. It is not a fine-tune of any externa
 |---|---|
 | Training stage | W2.7 (mid-curriculum) |
 | Most recent chunk eval | 47/105 @ chunk 450 |
-| Scope | math, science, reasoning, chain-of-thought, UK knowledge, Celtic languages, MOTHER identity |
+| Scope | math, science, reasoning, chain-of-thought, UK knowledge, Celtic languages, MOTHER identity, agentic tool use, multi-step planning, RAG, memory, composition (see §3) |
 | Out of scope (separate future models) | code generation, creative writing, vision |
 
 This release is for **internal team testing**. It will fail on tasks outside its training scope.
@@ -74,7 +74,145 @@ W2.7 will continue to chunk 650, after which the W2.8 corpus addition (~330,000
 
 ---
 
-## 3. Locked Inference Rules
+## 3. Agent Capabilities Trained
+
+This checkpoint was trained on the **W2.7 agentic curriculum** in addition to the base reasoning corpus. The model has been exposed to 57 agent-related training categories spanning planning, tool-calling, chain composition, recovery, RAG, memory, and workflow execution.
+
+The per-category training loss values below are taken from chunk 484 (closest complete log to chunk 450); lower is better — values below 0.5 indicate the category is well-learned, 0.5-1.0 is partially learned, >1.0 needs more training.
+
+### 3.1 Agent reasoning & planning
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `agent_cot_planning` | 0.45 | Decompose a user goal into a stepped plan before acting |
+| `agent_cot_decomposition` | 0.42 | Break a multi-part task into independent sub-tasks |
+| `agent_cot_synthesis` | 0.39 | Combine multiple tool results into a single answer |
+| `agent_cot_verification` | 0.37 | Verify a tool result against the original acceptance criteria |
+| `agent_cot_replan` | 0.03 | Revise the plan mid-execution when an observation invalidates it |
+| `agent_args_validation` | 0.03 | Validate tool-call arguments before emitting the call |
+| `agent_args_hallucination_resist` | 0.08 | Refuse to invent arguments not present in the conversation |
+
+### 3.2 Tool calling
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `agent_call_documents` | 0.59 | Call doc tools (Drive, Notion, PDF/Word/Excel creation) |
+| `agent_call_microsoft` | 2.22 | Call Microsoft tools (Graph, Outlook, Teams) — *needs more training* |
+| `agent_call_google` | 1.03 | Call Google tools (Drive, Calendar, Gmail) |
+| `agent_call_code` | 1.42 | Call code-execution tools (Python, shell, sandbox) |
+| `agent_no_tool_needed` | 0.29 | Recognise when a question needs no tool and answer directly |
+| `tool_choice_routing` | 0.17 | Route a request to the correct tool of several plausible ones |
+
+### 3.3 Multi-step chains
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `agent_chain_3step` | 0.53 | Three-step sequential tool chains |
+| `agent_chain_5plus` | 0.59 | Five-or-more-step tool chains |
+| `agent_conditional_chain` | 0.43 | Branching chains where step N depends on step N-1's result |
+| `agent_parallel_calls` | 0.51 | Issue independent tool calls in parallel and merge results |
+
+### 3.4 Control flow & safety
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `agent_disambiguation` | 0.65 | Ask for clarification when the user's request is ambiguous |
+| `agent_error_recovery` | 0.58 | Recover gracefully from a failed tool call |
+| `agent_mid_chain_abort` | 0.32 | Abort a chain when a step reveals the original goal is unreachable |
+| `agent_loop_aggregation` | 1.13 | Aggregate results from a loop of tool calls |
+| `agent_oauth_required` | 1.12 | Recognise when a tool needs OAuth and surface that to the user |
+| `agent_unsafe_refusal` | 0.21 | Refuse unsafe, malicious, or out-of-policy requests |
+
+### 3.5 RAG (retrieval-augmented generation)
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `rag_single_call` | 1.34 | Single retrieval call before answering |
+| `rag_synthesis` | 1.90 | Synthesise across multiple retrieved chunks — *needs more training* |
+| `rag_with_citation` | 1.25 | Include source citations in the synthesised answer |
+| `rag_empty_fallback` | 0.25 | Handle "no relevant results" gracefully |
+| `rag_not_needed` | 1.34 | Decline to retrieve when the question doesn't warrant it |
+
+### 3.6 Working memory
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `memory_store` | 0.05 | Persist a fact across turns in a session |
+| `memory_recall` | 0.33 | Retrieve a previously-stored fact when relevant |
+| `memory_multi_turn` | 0.19 | Carry intermediate state through a multi-turn session |
+| `memory_empty` | 0.12 | Handle the cold-start case where memory is empty |
+
+### 3.7 Composition (multi-modal chains)
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `compose_calc_chain` | 0.53 | Compose calculator + downstream tool |
+| `compose_memory_calc` | 0.61 | Combine memory recall with calculation |
+| `compose_rag_calc` | 0.34 | Retrieve facts, then compute with them |
+| `compose_rag_multi` | 0.28 | Multi-step retrieval-and-reason chains |
+| `compose_rag_web` | 0.10 | Combine internal retrieval with web search |
+| `compose_web_calc` | 0.13 | Web search + calculation |
+| `compose_web_memory` | 0.65 | Web search + memory storage |
+
+### 3.8 Error recovery
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `recovery_admit_failure` | 0.19 | Honestly admit when a tool failed rather than fabricate |
+| `recovery_alternate` | 0.09 | Try an alternative tool or strategy after failure |
+| `recovery_malformed` | 0.17 | Detect and repair malformed tool output |
+| `recovery_rewrite` | 0.10 | Rewrite a failing query in a way more likely to succeed |
+
+### 3.9 Web search primitives
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `web_search_single` | 0.11 | One-shot web search |
+| `web_search_reading` | 0.19 | Fetch and read a specific URL |
+| `web_search_fallback` | 0.26 | Use web search when internal sources fail |
+
+### 3.10 Chat behaviour
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `chat_greeting` | 0.37 | Conversational greetings |
+| `chat_acknowledgement` | 0.37 | Acknowledge received instructions |
+| `chat_identity` | 0.51 | Maintain MOTHER/MSAI identity in conversation |
+| `chat_helpful_refusal` | 0.54 | Decline politely and offer alternatives |
+| `chat_length_match` | 1.25 | Match response length to question complexity |
+| `chat_multi_turn` | 0.19 | Maintain coherence across conversation turns |
+
+### 3.11 Pre-built workflows
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `workflow_invoice_send` | 0.68 | End-to-end invoice creation + send workflow |
+| `workflow_meeting_prep` | 0.86 | Meeting preparation (calendar + brief generation) |
+| `workflow_msai_specific` | 1.26 | MSAI-internal workflows (deal flow, tenant comms) |
+| `workflow_proposal_pipeline` | 1.06 | Proposal authoring pipeline |
+| `workflow_report_generation` | 0.93 | Reporting workflows (status, financial, ops) |
+
+### 3.12 AUTM integration
+
+| Category | Loss @ chunk 484 | Purpose |
+|---|---|---|
+| `autm_agent` | 0.19 | Generic AUTM agent calling convention |
+| `autm_vertical` | 0.19 | AUTM vertical-specific dispatching |
+
+### Loss interpretation guide
+
+- **< 0.30** — well-learned; production-trustable
+- **0.30 – 0.60** — partially learned; usable but supervise outputs
+- **0.60 – 1.00** — emerging; expect inconsistent behaviour
+- **> 1.00** — still training; treat outputs as unreliable
+
+### W2.8 will strengthen the weak categories
+
+The W2.8 corpus (~330,000 new records, currently in build) targets the categories above 1.0 — particularly `agent_call_microsoft` (2.22), `rag_synthesis` (1.90), `agent_call_code` (1.42), `rag_single_call` / `rag_not_needed` (1.34), `chat_length_match` (1.25), and `workflow_msai_specific` (1.26). W2.8 also adds new categories: `doc_format_subtyping`, `verifier_loop`, `args_validation_adversarial`, `execution_graph_dag`, `cot_replan_observation`, `tool_failure_recovery`, `rag_synthesis_grounded`, `retrieval_arbiter`, `multi_agent_orchestration`, and `memory_synthesis`.
+
+---
+
+## 4. Locked Inference Rules
 
 **Deviation from these rules produces incorrect or degenerate output.** They are not suggestions — they are the inference recipe the model was trained against.
 
@@ -97,7 +235,7 @@ A working reference is included as `inference.py` in this repo. The canonical im
 
 ---
 
-## 4. Architecture Detail
+## 5. Architecture Detail
 
 ```
 MotherCoreModel
@@ -139,7 +277,7 @@ Forward return:
 
 ---
 
-## 5. Training
+## 6. Training
 
 ### Corpus (W2.7)
 
@@ -173,7 +311,7 @@ Training was performed at sequence length **2048** using physical microbatches o
 
 ---
 
-## 6. Sovereign Build Posture
+## 7. Sovereign Build Posture
 
 MOTHER CORE is part of MSAI's sovereign AI stack — built end-to-end in the UK on UK-resident infrastructure. The training, weights, tokeniser, and corpus are owned by MSAI. The training datacentres are MSAI-operated (Wright Avenue, Dundee; with additional sites in Durham and Manchester). No US cloud provider is in the inference or training path.
 
@@ -181,7 +319,7 @@ This positioning matters for UK government, defence, and regulated-enterprise cu
 
 ---
 
-## 7. Intended Use & Out-of-Scope Use
+## 8. Intended Use & Out-of-Scope Use
 
 **In scope (this checkpoint):**
 - Reasoning and chain-of-thought tasks at modest difficulty
@@ -200,7 +338,7 @@ This positioning matters for UK government, defence, and regulated-enterprise cu
 
 ---
 
-## 8. Evaluation
+## 9. Evaluation
 
 The internal eval suite at chunk 450 scores **47/105 (44.8%)** across:
 
@@ -221,7 +359,7 @@ Eval suite and methodology are MSAI-internal. Comparable public benchmarks (MMLU
 
 ---
 
-## 9. Limitations & Known Failure Modes
+## 10. Limitations & Known Failure Modes
 
 1. **Single-turn only** — no chat-style multi-turn coherence
 2. **Format-brittle** — the `Question:\n\n...\n\nAnswer:` template is required; other formats produce OOD output
@@ -234,7 +372,7 @@ Eval suite and methodology are MSAI-internal. Comparable public benchmarks (MMLU
 
 ---
 
-## 10. Usage
+## 11. Usage
 
 ### Quick test from a clean Python environment
 
@@ -279,7 +417,7 @@ python inference.py "What is the capital of Scotland?"
 
 ---
 
-## 11. License
+## 12. License
 
 **MSAI Sovereign License — Internal & Partner Use Only.**
 
@@ -289,7 +427,7 @@ For licensing enquiries: contact MediaStream AI Limited via the company website.
 
 ---
 
-## 12. Citation
+## 13. Citation
 
 ```
 @misc{msai-mother-core-2026,
@@ -303,9 +441,10 @@ For licensing enquiries: contact MediaStream AI Limited via the company website.
 
 ---
 
-## 13. Contact
+## 14. Contact
 
 - Organisation: MediaStream AI Limited (MSAI)
 - Founder & CEO: Christopher Kenna
+- Lead AI Architect: Christopher Kenna
 - Web: https://mediastreamai.com
 - Infrastructure: UK sovereign (Dundee, Durham, Manchester)