savant-code 0.0.16 → 0.0.19

package/dev/fids/archive/FID-2026-0803-013-agent-roster-over-reporting.md

@@ -0,0 +1,146 @@
+# FID: Agent Roster Over-Reporting (13 spawnable agents reported as "the roster")
+
+**Filename:** `FID-2026-0803-013-agent-roster-over-reporting.md`
+**ID:** FID-2026-0803-013
+**Severity:** low (UX/consistency — no data loss, no correctness failure)
+**Status:** verified
+**Created:** 2026-08-03
+**Author:** Savant
+
+**Summary:**
+When a user asks the Savant orchestrator "what's the agent roster?", the model
+reports **13 agents** instead of the canonical **9 ECHO roles**. Root cause:
+the main-agent instructions prompt auto-appends a "You can spawn the following
+agents:" addendum built from the full `spawnableAgents` allowlist (13 entries
+including infrastructure helpers `basher`, `tmux-cli`, `browser-use`,
+`context-pruner` and the two Researcher tool libraries listed separately). The
+model treats that functional spawn list as the roster definition. The fix adds
+an explicit canonical **Agent Roster** section to the Savant system prompt that
+defines exactly the 9 ECHO roles and instructs the model to report only those
+when asked.
+
+---
+
+## RED — Evidence
+
+| # | Location | Issue |
+|---|----------|-------|
+| R-1 | `agents/savant/savant.ts:120-136` | `spawnableAgents` allowlist = 13 entries: `detective`, `scout`, `researcher-web`, `researcher-docs`, `basher`, `thinker`, `forge`, `verifier`, `tmux-cli`, `browser-use`, `context-pruner`, `recorder`, `scribe` |
+| R-2 | `packages/agent-runtime/src/templates/strings.ts:236-253` | For the main agent (non-inherited tools branch), the instructions prompt gets `addendum += "\n\nYou can spawn the following agents:\n\n" + agentDescriptions.join('\n')` where `agentDescriptions` maps **every** `spawnableAgents` entry to `- ${agentType}: ${template.spawnerPrompt}` |
+| R-3 | `ARCHITECTURE.md:22-33` | Canonical roster table = **9 roles**: Orchestrator, Detective, Forge, Verifier, Recorder, Thinker, Scout, Researcher, Scribe |
+| R-4 | `ARCHITECTURE.md:210-236` | "Helper Tool Libraries" section: `basher`/`tmux-cli`/`browser-use`/`context-pruner` are infrastructure helpers consumed BY the 9 roles — "do NOT constitute independent conversational agents". `researcher-web` + `researcher-docs` are the two tool libraries of the single Researcher role |
+| R-5 | Live behavior | User asked "the roster" → model listed multiple agents beyond the core 9 (parroting the 13-entry addendum) |
+
+### Finding details
+
+**R-1/R-2 (the mechanism):** The orchestrator legitimately needs all 13 in its
+spawn allowlist — `basher` runs typechecks, `browser-use` verifies UI,
+`tmux-cli` tests interactive CLIs, `context-pruner` auto-spawns via
+handleSteps. The bug is not the allowlist; it's that the runtime renders that
+functional list into the prompt verbatim as "You can spawn the following
+agents:", and the model conflates it with the roster definition when asked.
+The system prompt currently contains **no canonical roster section** — grep for
+`Agent Roster` / `The Agent Roster` across `agents/`, `common/src/`,
+`packages/agent-runtime/src/` returns zero hits in non-test code. The model
+only ever sees the 13-entry spawn list, so that's what it reports.
+
+**R-5 (expected answer):** The 9 canonical ECHO roles with their
+responsibilities (as stated by the user and consistent with ARCHITECTURE.md):
+
+| Role | Responsibility |
+|------|----------------|
+| Savant (Orchestrator) | Routes work, enforces protocol, spawns agents |
+| Detective | Discovers bugs and issues with evidence before any code is written |
+| Forge | Implements code changes from a converged plan |
+| Verifier | Independent double-audit after implementation |
+| Thinker | Deep sequential reasoning for complex problems |
+| Scout | Explores codebases to gather context |
+| Researcher | Web search and documentation lookup |
+| Recorder | FID lifecycle management and tracking |
+| Scribe | Session summaries and knowledge capture |
+
+---
+
+## GREEN — Solution
+
+1. **R-1..R-5 fix:** Add an explicit `# Agent Roster` section to the Savant
+   default system prompt (`buildDefaultSystemPrompt` in
+   `agents/savant/savant.ts`, after the identity/date lines and before
+   `# General guidelines`). Content:
+
+   - A table of the **9 canonical ECHO roles** with one-line responsibilities
+     (matching ARCHITECTURE.md + the user's expected answer).
+   - A clarifying note: `researcher-web` and `researcher-docs` are the two tool
+     libraries of the single **Researcher** role; `basher`, `tmux-cli`,
+     `browser-use`, and `context-pruner` are **infrastructure helpers** — NOT
+     roster members.
+   - An explicit instruction: "When asked about the agent roster, report only
+     the 9 roles listed above."
+
+2. **No change to `spawnableAgents`** (R-1): the allowlist stays at 13 — it is
+   functionally correct. Removing infrastructure entries would break spawning.
+
+3. **No change to `strings.ts` addendum** (R-2): it remains the functional
+   "what can I spawn" list. The roster definition in the system prompt resolves
+   the conflation at the source.
+
+4. **Docs alignment:** `AGENTS.md` already carries the correct 9-role table —
+   no doc change needed. `ARCHITECTURE.md` already documents the
+   helper-libraries distinction — no change needed.
+
+---
+
+## AUDIT — Verification
+
+1. **Static (grep):** After the fix, `agents/savant/savant.ts` contains an
+   `# Agent Roster` heading; grep for `Agent Roster` in non-test TS now hits
+   exactly one source site.
+2. **Runtime (rendered prompt):** Typecheck the `agents` workspace, then run
+   the existing strings/template test suite in `packages/agent-runtime` — the
+   addendum tests (`strings.test.ts:158/200/277/445`) assert the *generic*
+   "You can spawn the following agents:" text, which we do NOT modify, so they
+   must stay green unchanged.
+3. **Model behavior:** A follow-up prompt asking "what's the roster?" should
+   now yield the 9 canonical roles (Savant signs this FID; verification of
+   model output is by the user at approval).
+4. **Call-graph:** `buildDefaultSystemPrompt` is invoked from
+   `buildSystemPrompt` (single caller) — template string change only, no
+   exported symbol changes, no callers to update.
+
+---
+
+## Resolution — COMPLETE (operator-approved)
+
+Operator approved the GREEN solution ("re-read the ECHO spec, run the
+perfection loop on the FID, then code"). AUDIT on the FID document passed: insertion point
+verified (`buildDefaultSystemPrompt` at `agents/savant/savant.ts:504`;
+identity and date lines 512-514; `# General guidelines` at 516); single
+caller (`buildSystemPrompt:478`) — zero callers to update (Law 4); zero
+existing roster text in non-test TS; zero tests reference the system prompt
+content.
+
+Implemented: added `# Agent Roster` section (9 canonical ECHO roles table +
+Researcher tool-libraries + infrastructure-helpers clarification + "report
+only the 9" instruction) between the date line and `# General guidelines` in
+`buildDefaultSystemPrompt`.
+
+**Verification evidence (Double Audit):**
+
+- Method 1 (static): grep confirms `# Agent Roster` at
+  `agents/savant/savant.ts:516` with the 9-role table and helper-libraries
+  note; `spawnableAgents` unchanged at 13 entries (lines 120-136);
+  `buildDefaultSystemPrompt` has a single caller (`buildSystemPrompt:478`).
+- Method 2 (runtime): `agents` workspace typecheck → exit 0; CLI typecheck
+  (bundle regenerated via `prebuild:agents`) → exit 0; eslint
+  `agents/savant/savant.ts --max-warnings 0` → clean; `lint:md` → clean;
+  `packages/agent-runtime` strings template suite 11/11 pass (addendum
+  untouched). Independent review: clean.
+
+**Amendment (2026-08-03):** Operator provided the exact canonical roster
+content — the section now uses the ARCHITECTURE.md table format with # / Agent /
+Phase / Responsibility columns (numbered 1-9) and a dedicated "Important
+distinction" subsection listing the infrastructure helpers
+(`researcher-web`/`researcher-docs`, `basher`, `tmux-cli`, `browser-use`,
+`context-pruner`) as spawnable-but-not-roster-members. Re-verified after the
+content replacement: `prebuild:agents` bundle regenerated (12 refs),
+`agents` typecheck exit 0, eslint clean.

package/dev/fids/archive/FID-2026-0803-014-freebuff-to-savant-rebrand-sweep.md

@@ -0,0 +1,185 @@
+# FID: FreeBuff → Savant Rebrand Sweep (195+ instances across 34 tracked files)
+
+**Filename:** `FID-2026-0803-014-freebuff-to-savant-rebrand-sweep.md`
+**ID:** FID-2026-0803-014
+**Severity:** medium (branding consistency + contract hygiene — no runtime data loss,
+but stale FreeBuff identifiers/contract keys leak into a Savant-only product)
+**Status:** complete
+**Created:** 2026-08-03
+**Author:** Savant
+
+**Summary:**
+Repo-wide audit of 195+ `freebuff` / `FreeBuff` / `FREEBUFF` /
+`Freebuff` instances across 34 tracked files. Categories: (A) **live
+contract code** — the `freebuff` YAML namespace + `FreeBuffProtocolConfig`
+parser in `protocol-config.ts`, the `cli.update_freebuff_failed` telemetry
+event in the released savant-free wrapper, and the stale `FREEBUFF_MODE=true`
+dev-script env var; (B) **kept protocol documents** —
+`ECHO-freebuff.md` and `dev/nova/specs/echo-v0.1.2-freebuff.md` remain named
+and intact because they are the authoritative FreeBuff ECHO adaptation; (C)
+**safe docs/config prose** — current Savant-facing documentation and selected
+historical records were updated, while legal attributions, protocol-routing
+directives, explicit historical records, and `LEARNINGS.md` were preserved.
+
+---
+
+## RED — Evidence
+
+### Category A — live contract code (4 files, 27 instances plus compatibility normalization)
+
+- **RB-A1 — `common/src/util/protocol-config.ts`:** Replaced the
+  `FreeBuffProtocolConfig` and `.freebuff` runtime shape with
+  `SavantProtocolConfig` and `.savant`. The parser prefers `savant.protocol`
+  and normalizes the documented legacy `freebuff.protocol` alias. The sole
+  consumer reads `.maxIterations`; no `.freebuff` runtime field is exposed.
+- **RB-A2 — `protocol.config.yaml` and parser tests:** The active
+  `savant.protocol` contract and legacy `freebuff.protocol` compatibility alias
+  are both covered by focused tests. This keeps the retained protocol docs
+  operational without making FreeBuff the active runtime naming.
+- **RB-A3 — release wrapper telemetry:** Renamed
+  `cli.update_freebuff_failed` to `cli.update_savant_free_failed` in both the
+  released wrapper and wrapper-safety test.
+- **RB-A4 — `package.json`:** Renamed the development script environment
+  variable from `FREEBUFF_MODE` to `SAVANT_FREE_MODE`, matching the actual
+  runtime reader and binary build injection.
+
+### Category B — freebuff-named protocol files (kept by operator decision)
+
+| # | File | Decision | Proper usage |
+|---|------|----------|--------------|
+| RB-B1 | `ECHO-freebuff.md` | Keep filename and protocol content | Protocol marker for the FreeBuff ECHO adaptation |
+| RB-B2 | `dev/nova/specs/echo-v0.1.2-freebuff.md` | Keep filename and protocol content | Governing FreeBuff single-agent protocol referenced by `FREEREADME.md` |
+
+These are not stale Savant runtime identifiers. They intentionally document a
+separate protocol lineage and remain the canonical references for FreeBuff
+sessions. Their names, version tag `0.1.2-freebuff`, and routing references are
+therefore excluded from the branding sweep.
+
+### Category C — safe docs/config prose and history (~28 files, 150 instances)
+
+| File | Count | Notes |
+|------|-------|-------|
+| `savant-free/SPEC.md` | 40 | `FREEBUFF_MODE`/`IS_FREEBUFF`/`FREEBUFF` block letters → `SAVANT_FREE_MODE`/`IS_SAVANT_FREE`/`SAVANT_FREE` |
+| `CHANGELOG.md` | 24 | Historical entries — FreeBuff → Savant (preserving prose structure) |
+| `dev/test-prompts/archive/release-az-test-fid-2026-0728-008.md` | 17 | T3.x protocol tests reference renamed spec files |
+| `dev/session-summaries/2026-07-31-freebuff-echo-compliance-remediation.md` | 12 | Filename itself is dated (`-freebuff-`) — content renamed; filename keeps historical date prefix (documented) |
+| `dev/session-summaries/2026-07-19-fid-026-debugging-and-rename.md` | 12 | Prose |
+| `FREEREADME.md` | 12 | Kept as the FreeBuff session directive; its protocol-routing references remain intentional |
+| `dev/session-summaries/2026-07-19-fid-026-phase-b-rebrand.md` | 5 | Prose |
+| `dev/fids/archive/FID-2026-0803-001` (3), `...-002` (3), `...-003` (1) | 7 | Archived FID prose (FreeBuff references) |
+| `dev/session-summaries/2026-07-17-1000.md` (3), others (1 each ×3) | 6 | Prose |
+| `docs/reports/codebuff-discord-feedback.md` (3), `docs/gravity-integration-starter.md` (3) | 6 | Prose |
+| `dev/test-prompts/archive/*` (5 more files, 1-3 each) | 8 | Prose |
+| `LEARNINGS.md` (2) | 2 | Preserved verbatim per operator instruction; `dev/releases/v0.0.2.md` and safe current-facing docs were updated |
+| `savant-free/README.md` | 2 | `FREEBUFF_MODE` prose |
+
+### Out of scope (documented)
+
+- `research/` (vendored `servers-main` reference copy, 144 tracked files) and
+  `resources/` (untracked scans) — vendored third-party reference material, not
+  product code. FreeBuff hits there (if any) are upstream/reference prose.
+- `.git/lost-found/` — git garbage, not tracked working tree.
+- `bun.lock`, `database.db`, `cli/bin/`, `.env.local` — generated/ignored.
+
+---
+
+## GREEN — Solution
+
+1. **RB-A1/A2 — protocol config rename (lockstep):**
+   - `common/src/util/protocol-config.ts`: `FreeBuffProtocolConfig` →
+     `SavantProtocolConfig`; active runtime field `freebuff` → `savant`;
+     parser locals were renamed to Savant terminology; `savant.protocol` is
+     preferred while `freebuff.protocol` remains an explicit compatibility
+     alias normalized into `.savant`.
+   - `protocol.config.yaml`: the active contract is `savant:` with
+     `'0.1.2-savant'`; the legacy `freebuff.protocol` alias remains explicitly
+     documented for FreeBuff-session compatibility and is normalized by the
+     parser.
+   - `protocol-config.test.ts`: Savant fixture, legacy FreeBuff alias fixture,
+     and expectations cover both accepted namespaces. Law 4:
+     `readProtocolConfig` consumers remain unchanged (`.maxIterations` only);
+     both namespaces normalize into the `.savant` runtime property, so no
+     `.freebuff` runtime field is exposed.
+
+2. **RB-A3 — telemetry event (both sides):** `cli.update_freebuff_failed` →
+   `cli.update_savant_free_failed` in BOTH `savant-free/cli/release/index.js`
+   and `cli/src/__tests__/release/wrapper-safety.test.ts`. The wrapper-safety
+   test asserts the released wrapper's exact config, so both must change in one
+   commit.
+
+3. **RB-A4 — dev script env var:** `FREEBUFF_MODE=true` →
+   `SAVANT_FREE_MODE=true` in `package.json` (`dev:savant-free` script) —
+   aligns with the actual code (`build-binary.ts:176` reads `SAVANT_FREE_MODE`).
+
+4. **RB-B — protocol-file names kept:**
+   - No `git mv` was performed. `ECHO-freebuff.md` and
+     `dev/nova/specs/echo-v0.1.2-freebuff.md` remain the canonical FreeBuff
+     protocol references by operator decision.
+   - `FREEREADME.md` continues to route FreeBuff sessions to those files and
+     `freebuff.protocol`; the parser accepts that namespace as a compatibility
+     alias while preferring the active `savant.protocol` contract.
+
+5. **RB-C — safe docs/archives prose sweep:** Case-preserving replacement
+   was applied only where the reference was current Savant-facing prose or a
+   safe normalization target. Protocol-routing directives, legal attributions,
+   explicit historical records, `LEARNINGS.md`, and the kept protocol docs were
+   not rewritten. Strategic docs now say Savant-Free where they discuss the
+   future free product.
+
+6. **No unintended runtime behavior change:** the active protocol contract,
+   legacy FreeBuff alias, development mode variable, and telemetry wrapper are
+   synchronized. Legal notices, historical records, protocol routing, and
+   `LEARNINGS.md` remain intentionally preserved.
+
+---
+
+## AUDIT — Verification
+
+1. **Static sweep (Method 1):** `git grep` confirms zero deprecated
+   `FREEBUFF_*` identifiers and zero `.freebuff` runtime consumers in active
+   source (`cli/src`, `common/src`, `sdk/src`, `packages`, `agents`, and
+   `savant-free/cli`). The sole active-source `freebuff` token is the explicit
+   compatibility parser/test for the documented `freebuff.protocol` alias.
+   Remaining matches are classified as retained protocol docs, legal
+   attribution, explicit history, `LEARNINGS.md`, or the legacy `.freebuff/`
+   ignore rule.
+2. **Runtime (Method 2):** Typecheck ×4 (sdk/common/agent-runtime/cli) exit 0;
+   `common` test suite pass (protocol-config tests updated); `cli` wrapper-safety
+   test pass (telemetry key both sides); `agents` typecheck exit 0; ESLint
+   `--max-warnings 0`; `lint:md` exit 0.
+3. **Law 4 call-graph:** no file rename was performed by approved scope.
+   `FreeBuffProtocolConfig`, `update_freebuff_failed`, and deprecated runtime
+   `FREEBUFF_MODE` readers are absent from active code; the kept protocol names
+   are intentionally present only in their documented protocol/history boundary.
+4. **Spec consistency:** the active `savant.protocol` contract and documented
+   legacy `freebuff.protocol` alias in `protocol.config.yaml` match
+   `common/src/util/protocol-config.ts`; the kept FreeBuff protocol paths remain
+   valid in `FREEREADME.md` and the protocol test prompts.
+
+---
+
+## Resolution — IMPLEMENTED (operator-approved scope: full sweep minus .md renames)
+
+Operator approved: "Full sweep minus .md renames." This means code, config,
+telemetry, and stale runtime environment identifiers were renamed to Savant;
+selected current-facing `.md` content was updated; `.md` filenames were not
+renamed; and the FreeBuff protocol documents plus `LEARNINGS.md`, legal notices,
+explicit historical records, and legacy compatibility rules were preserved.
+
+Implemented: RB-A1..A4 (protocol parser/config compatibility, telemetry, env
+script, and active `IS_SAVANT_FREE` wiring); RB-C safe current-facing docs and
+strategic wording; RB-B intentionally kept. Active-source verification found
+no deprecated `FREEBUFF_*` identifiers or `.freebuff` runtime consumers. The
+remaining name matches are documented protocol, legal, historical, learning,
+or compatibility boundaries.
+
+### Closeout evidence
+
+- `common` protocol-config focused test: 3 passed, 0 failed (including the
+  legacy `freebuff.protocol` normalization regression test).
+- Active source grep: no `FREEBUFF_*` or `.freebuff` runtime identifiers.
+- `SAVANT_FREE_MODE` readers and `savant` protocol config are synchronized;
+  the FreeBuff protocol alias is covered by a focused parser regression test.
+- Strategic docs no longer describe a planned FreeBuff hosting dependency.
+- `LEARNINGS.md` restored verbatim per operator instruction.
+- No `.md` filenames were renamed; FreeBuff protocol files remain usable.

package/dev/nova/reports/2026-08-02-top-25-ai-inference-providers-competitive-intel.md

@@ -0,0 +1,213 @@
+# Top 25 AI Inference Providers — Competitive Intelligence Report
+
+**Date:** August 2, 2026
+**Purpose:** Competitive landscape analysis for Savant Inference Gateway
+**Scope:** 25 providers across 5 categories
+
+---
+
+## Executive Summary
+
+The AI inference market in 2026 is fragmented across four distinct tiers: unified gateways, custom silicon, managed API providers, and enterprise cloud platforms. **OpenRouter dominates gateways** (5.5% markup, 400+ models, 10M+ users). **Groq leads custom silicon** (LPU ASIC, 1,000 TPS, $0.05/M input). **DeepInfra leads managed APIs** (300+ models, $0.10/M for 70B). **AWS/Azure/GCP dominate enterprise** with SOC2/HIPAA/FedRAMP compliance.
+
+**Key gaps Savant Inference can exploit:**
+1. **No self-hosted binary gateway exists** — all are cloud-only SaaS
+2. **No gateway offers both OpenAI-compat + provider-native dual mode** — Cloudflare deprecated OpenAI compat, OpenRouter is OpenAI-only
+3. **No gateway does multi-modal routing** (text + image + audio + video) natively
+4. **Zero-markup BYOK is validated** — flo2's entire pitch is "zero markup OpenRouter alternative"
+5. **ECHO Protocol governance is unique** — no competitor offers cryptographic quality gates
+
+---
+
+## Category 1: Unified Gateways & Aggregators
+
+| Provider | Status | Models | Pricing | Free Tier | Lock-in |
+|----------|--------|--------|---------|-----------|---------|
+| **OpenRouter** | ✅ Live | 400+ | 5.5% platform fee | 25+ models, 50 req/day | Low |
+| **OneInfer** | ❌ Defunct | — | — | — | — |
+| **Cloudflare AI GW** | ✅ Live | 25+ providers | Free (you pay providers) | Free on all CF plans | HIGH (CF ecosystem) |
+| **TokensMind** | ✅ Live (China) | Unknown | Enterprise sales | No | High |
+| **flo2** | ✅ Beta | All (BYOK) | $0 markup | Free during beta | None |
+
+### Key Takeaways:
+- OpenRouter's 5.5% markup is the market price — every BYOK competitor positions against it
+- Cloudflare AI Gateway is free but forces provider-native API (OpenAI compat deprecated)
+- flo2 validates zero-markup BYOK routing + racing/A/B testing features
+- **Gap:** No self-hosted gateway option exists
+
+---
+
+## Category 2: Ultra-Low Latency & Custom Silicon
+
+| Provider | Architecture | Cheapest Model | Speed | Free Tier | On-Prem |
+|----------|-------------|----------------|-------|-----------|---------|
+| **Groq** | LPU ASIC | $0.05/M (Llama 8B) | 1,000 TPS | ✅ Rate-limited | ✅ GroqRack |
+| **Cerebras** | Wafer-Scale Engine | $0.35/M (GPT OSS 120B) | ~3,000 TPS | ✅ $5 credits | ❓ |
+| **SambaNova** | RDU Dataflow | $0.22/M (GPT OSS 120B) | — | ✅ | ✅ SambaRack |
+| **GMI Cloud** | NVIDIA GPUs | GPU-hour based | — | ❌ (credits) | ❌ |
+
+### Key Takeaways:
+- Groq is cheapest overall (Llama 8B at $0.05/$0.08/M)
+- Cerebras has highest raw throughput (~3K TPS) but 2-3x pricier than Groq
+- SambaNova beats Cerebras on GPT OSS 120B ($0.22 vs $0.35 input)
+- All support OpenAI-compatible APIs — straightforward Gateway integration
+
+---
+
+## Category 3: Fast Managed APIs for Open Weights
+
+| Provider | Models | Cheapest 70B | Free Tier | Unique Feature |
+|----------|--------|-------------|-----------|----------------|
+| **Fireworks AI** | 200+ | — | $1 credits | Fine-tuning at base price |
+| **Together AI** | 200+ | $1.04/M | $1 credits | PTU for predictable cost |
+| **DeepInfra** | 300+ | $0.10/M | $1 credits | FP4 quantization, largest catalog |
+| **SiliconFlow** | 50+ | — | Permanently free models | Best free tier, China-optimized |
+| **DeepSeek API** | 2 | $0.14/M | $5 credits | Cheapest first-party, 1M context |
+| **Novita AI** | 100+ | — | $1 credits | Agent Sandbox, batch 50% off |
+
+### Key Takeaways:
+- DeepInfra has largest catalog (300+) and cheapest 70B ($0.10/M)
+- DeepSeek API is cheapest for first-party models ($0.14/$0.28/M)
+- SiliconFlow wins free tier (permanently free models, no credit required)
+- All support OpenAI-compatible APIs
+- DeepSeek additionally supports Anthropic API format
+
+---
+
+## Category 4: Developer-Centric GPU & Serverless
+
+| Provider | Model | Deployment | Unique Feature |
+|----------|-------|------------|----------------|
+| **Baseten** | Truss framework | Managed | Open-source deployment framework |
+| **Modal** | Python-native | Serverless | Sub-second cold starts, no YAML |
+| **Replicate** | Community models | Serverless | Easiest model deployment |
+| **RunPod** | GPU pods | On-demand | Fastest spin-up times |
+| **Fal.ai** | Multimedia | Serverless | Optimized for image/audio/video |
+| **HF Endpoints** | Hub-integrated | Managed | Deepest model ecosystem |
+
+### Key Takeaways:
+- These are deployment platforms, not API gateways — different market segment
+- Modal and Fal.ai are most relevant for custom model serving
+- None offer unified API routing across providers
+
+---
+
+## Category 5: Enterprise & Full-Stack Cloud
+
+| Provider | Pricing Model | Compliance | Unique Capability |
+|----------|--------------|------------|-------------------|
+| **Google Vertex AI** | Per-token + per-hour | SOC2, HIPAA, FedRAMP | TPU acceleration, Gemini models |
+| **AWS SageMaker** | Per-instance-hour + Bedrock per-token | SOC2, HIPAA, FedRAMP | Broadest model marketplace (20+) |
+| **Azure AI Foundry** | Per-token + PTU reservations | SOC2, HIPAA, FedRAMP, IL4/IL5 | Exclusive OpenAI access |
+| **Nebius AI** | Per-GPU-hour + per-token | SOC2, ISO 27001, HIPAA | Best GPU price/performance (bare-metal) |
+
+### Key Takeaways:
+- Azure has exclusive OpenAI model access (GPT-5 series)
+- AWS has broadest model marketplace via Bedrock
+- Nebius offers best GPU price/performance (bare-metal, non-virtualized)
+- All have SOC2/HIPAA/FedRAMP compliance
+
+---
+
+## Pricing Comparison: Cost per 1M Tokens (70B-class model)
+
+| Provider | Input | Output | Notes |
+|----------|-------|--------|-------|
+| **DeepInfra** | $0.10 | $0.32 | Llama 3.3 70B — cheapest |
+| **DeepSeek API** | $0.14 | $0.28 | V4 Flash — cheapest first-party |
+| **Groq** | $0.59 | $0.79 | Llama 3.3 70B — fastest |
+| **SambaNova** | $0.60 | $1.20 | Llama 3.3 70B |
+| **Together AI** | $1.04 | $1.04 | Llama 3.3 70B — symmetric |
+| **Cerebras** | $0.35 | $0.75 | GPT OSS 120B |
+| **OpenRouter** | Cost + 5.5% | Cost + 5.5% | Pass-through + markup |
+| **flo2** | $0 markup | $0 markup | BYOK, you pay providers |
+| **Cloudflare** | $0 markup | $0 markup | Free gateway, you pay providers |
+
+---
+
+## Free Tier Comparison
+
+| Provider | Free Offer |
+|----------|-----------|
+| **SiliconFlow** | Multiple permanently free models (GLM-Z1, embeddings, ASR, image) |
+| **DeepSeek API** | $5 free credits |
+| **OpenRouter** | 25+ models, 50 req/day |
+| **Groq** | Rate-limited free tier |
+| **Cerebras** | $5 free credits |
+| **SambaNova** | Free tier (details on signup) |
+| **Fireworks AI** | $1 free credits |
+| **Together AI** | $1 free credits |
+| **DeepInfra** | $1 free credits |
+| **Novita AI** | $1 free credits |
+| **Cloudflare AI GW** | Free on all plans |
+| **flo2** | Free during beta |
+
+---
+
+## Competitive Gaps Savant Inference Can Exploit
+
+### 1. Self-Hosted Binary Gateway
+**No provider offers a self-hosted binary.** All are cloud SaaS. Enterprise teams wanting data locality + no vendor dependency have zero options. Savant's Rust-based proxy compiled to a single binary is a unique differentiator.
+
+### 2. Dual API Mode (OpenAI + Provider-Native)
+Cloudflare deprecated OpenAI compat. OpenRouter is OpenAI-only. A gateway that does BOTH would be unique. Savant's universal schema translation (OpenAI → Anthropic/DeepSeek/Google) is already designed for this.
+
+### 3. ECHO Protocol Governance
+No competitor offers cryptographic quality gates, FID-bound execution, or the Perfection Loop FSM. This is a moat for enterprise compliance.
+
+### 4. Multi-Modal Routing
+Most gateways focus on text LLMs. Cloudflare supports voice/video (WebSockets) but it's Beta. First-class multi-modal routing could be a differentiator.
+
+### 5. Zero-Overhead Proxy
+LiteLLM (Python) fails at 500 RPS. Kong adds 25-40ms overhead. Savant's Pingora-based proxy achieves 11μs overhead at 5,000 RPS. This is a 100-1000x latency advantage.
+
+### 6. Transparent Cost-Plus Pricing
+OpenRouter's 5.5% markup is opaque. CommandCode locks users to CLI. Savant's cost-plus model (resell at provider cost, earn from subscriptions) builds developer trust.
+
+---
+
+## Integration Priority for Gateway
+
+### Tier 1 (MVP — must have):
+1. **DeepSeek API** — Cheapest first-party, 1M context, dual API format
+2. **Groq** — Fastest inference, LPU ASIC, OpenAI-compat
+3. **OpenRouter** — Largest model catalog, pass-through pricing
+
+### Tier 2 (v1.1 — high value):
+4. **DeepInfra** — Largest catalog (300+), cheapest 70B
+5. **Together AI** — Most complete platform, PTU for predictable cost
+6. **Fireworks AI** — Best fine-tuning value, serverless training
+
+### Tier 3 (v1.2 — enterprise):
+7. **Cerebras** — Highest throughput (~3K TPS)
+8. **SambaNova** — Competitive pricing, on-prem option
+9. **Cloudflare AI Gateway** — Free, but deep ecosystem lock-in
+
+### Tier 4 (enterprise — deferred):
+10. **Azure AI Foundry** — Exclusive OpenAI access
+11. **AWS SageMaker/Bedrock** — Broadest enterprise ecosystem
+12. **Google Vertex AI** — TPU acceleration, Gemini models
+
+---
+
+## Sources
+
+- OpenRouter: openrouter.ai
+- Cloudflare AI Gateway: cloudflare.com/ai
+- flo2: flo2.com
+- Groq: groq.com
+- Cerebras: cerebras.ai
+- SambaNova: sambanova.ai
+- GMI Cloud: gmicloud.ai
+- Fireworks AI: fireworks.ai
+- Together AI: together.ai
+- DeepInfra: deepinfra.com
+- SiliconFlow: siliconflow.cn
+- DeepSeek API: deepseek.com
+- Novita AI: novita.ai
+- Google Vertex AI: cloud.google.com/vertex-ai
+- AWS SageMaker: aws.amazon.com/sagemaker
+- Azure AI Foundry: azure.microsoft.com/en-us/products/ai-studio
+- Nebius AI: nebius.com
+
+*All pricing data as of August 2, 2026*