Persona Builder · Adaptive voice bake-off
The Context on demand era includes GPT-6 Luna and Sol reasoning pairs and GPT-6 Astra UltraFast / GPT-6.1 Sol at low reasoning, plus thinking-disabled Wafer 31B FP8, Wafer 26B A4B FP8, Crusoe, and Inworld Gemma 4 baselines, and Muse Spark Open low/medium reasoning on Baseten. All pursue the adaptive docent-full goal through serial real voice CVI. Resource searches and guides load during the conversation. The September 30 WebSocket GPT arms, final-guard-audit Gemma arms, and October 1 A4B arm use the same carrier prompt and RQH #3142 preview; their RQH host is shown below. The earlier four HTTP adaptive-completion arms remain available as the transport baseline. This varies Charlie’s lead model; the shared RQH drafter route stays fixed. Earlier captures use different production carrier versions. The October 16 Inworld row runs the live production carrier 20dc098b (55,222 characters) with a GPT-6 Luna creator; the October 15 Wafer/Muse rows used carrier 4773c607 (54,797 characters) and a Gemini 2.5 Flash creator, so compare them as adjacent captures, not a controlled provider A/B. Historical GLM sequential and Gemma parallel fan-out eras retain 39 arms each. Cross-date and cross-era scores are historical context, not a controlled model or prompt A/B. Provider/configuration failures and unrun arms never receive model scores.
Guided closeout preview: Mercury 3 low and Wafer Gemma 4 FP8 use PR #1913, the real Portal client, the relevant museum PDF, and Gemini 3.6 Flash as creator. Shared test RQH; production carriers are unchanged. One completed capture per model; excluded earlier attempts include a Mercury setup failure, a premature Mercury Test handoff, and a Gemma idle shutdown. Compare the preview arms → Different runtime and creator from historical captures: this is not a controlled prompt-only improvement estimate. Capture findings, aborted attempts, and conversation-quality review →
October 19 Mercury 3: low reasoning directly through Inception, strict tool schemas, production carrier 20dc098b at capture, and GPT-6 Sol as the simulated creator. The older GPT/Muse/Wafer runs used a different carrier and creator. Compare Mercury 3 with Astra, Wafer Gemma 4, and Muse →
Tool calls: reviewed first-response execution now caps the Reliability bar and contributes to the overall score. Reliability is the lower of tool-call success and the existing dropped-turn/tool-leakage score. Missing calls count even when no guard intervened. Ambiguous turns are disclosed. Reviewed v4 and legacy scores have separate rankings; legacy reliability has no tool-execution review. Reliability now weighs 25%, autonomy 25%, build 35%, requirements 10%, and latency 5%.
Timing versus eventual success: Tool calls measures first-response dispatch on actionable creator turns, not the validity of emitted calls. Recovery and later batching are shown separately. One draft_pal call can dispatch several shared backend drafters; final completeness is not an independent prompt-quality score. Repeated requests can share the same later repair.
October 15 comparison: Muse Spark Open on Baseten (low and medium reasoning) and Wafer Gemma 4 31B FP8 (thinking disabled) use the same production carrier prompt and production RQH. Compare Muse Spark Open with Wafer Gemma 4 → Each setting has one adaptive capture; the conversations differ.
WebSocket experiment: this opt-in eval adapter reuses provider connections; production CVI transport is unchanged. Provider replay showed a median paired first-output saving of 338 ms for Astra and 361 ms for Sol, with uncertainty for Sol. The voice runs use different adaptive conversations, so their audible latency differences do not isolate transport. Read the measurements, grading corrections, and remaining CVI context race.
The original at-a-glance diagnostic chart, showing each reasoning arm independently. Its passed/total count is not the standings score: hard thresholds and overlapping checks would otherwise distort the ordering. A timeout or rejected provider credential is harness/configuration evidence; it is not presented as a model-quality failure.
Each required field and authored section for the models selected above. Initial means the snapshot at the first draft result; final means the completed capture. Face selection must precede drafting. A dash means the evidence was not captured and is excluded from the check percentage. Presence is mechanical completeness, not a judgement of content quality.
Every tool observed in either the model’s original turn or the turn-guard recovery. Counts are actual calls, not unique field updates.
tavus-glm-4.7 managed baseline alias
{
"model": "tavus-glm-4.7",
"speculative_inference": true
}
Evidence: the managed baseline alias is explicit; no reasoning control or trace is exposed
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-gemma-4-thinking managed alias (Cerebras)
{
"model": "tavus-gemma-4-thinking",
"speculative_inference": true
}
Evidence: 685 reasoning tokens in bake-off conversation c62f39421e7d8477
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-gemma-4 managed alias (Cerebras)
{
"model": "tavus-gemma-4",
"speculative_inference": true
}
Evidence: non-thinking managed alias selected explicitly
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-2.5-flash via Google AI Studio
{
"model": "gemini-2.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/"
}
Evidence: 285 hidden thought tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-2.5-flash via Google AI Studio
{
"model": "gemini-2.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: 0 hidden thought tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-3.5-flash via Google AI Studio
{
"model": "gemini-3.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/"
}
Evidence: 261 hidden thought tokens and a thought signature in the matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-3.5-flash via Google AI Studio
{
"model": "gemini-3.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
"extra_body": {
"reasoning_effort": "minimal"
}
}
Evidence: 0 billed hidden tokens on a simple probe, but a thought signature remained
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-luna via OpenAI, reasoning_effort=medium
{
"model": "gpt-5.6-luna",
"speculative_inference": true,
"base_url": "https://forecast-accessible-relationships-ebook.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "medium"
}
}
Evidence: 26 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-luna via OpenAI, reasoning_effort=none
{
"model": "gpt-5.6-luna",
"speculative_inference": true,
"base_url": "https://api.openai.com/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: 0 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-sol via OpenAI, reasoning_effort=high
{
"model": "gpt-5.6-sol",
"speculative_inference": true,
"base_url": "https://forecast-accessible-relationships-ebook.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "high"
}
}
Evidence: 3400 reasoning tokens on a complex matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-sol via OpenAI, reasoning_effort=none
{
"model": "gpt-5.6-sol",
"speculative_inference": true,
"base_url": "https://api.openai.com/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: 0 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-opus-4-6 via Anthropic, extended thinking
{
"model": "claude-opus-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/",
"extra_body": {
"thinking": {
"type": "enabled",
"budget_tokens": 1024
}
}
}
Evidence: signed thinking block and 41 thinking tokens in the native trace
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-opus-4-6 via Anthropic
{
"model": "claude-opus-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/"
}
Evidence: no thinking parameter was sent
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-sonnet-4-6 via Anthropic, extended thinking
{
"model": "claude-sonnet-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/",
"extra_body": {
"thinking": {
"type": "enabled",
"budget_tokens": 1024
}
}
}
Evidence: signed thinking block and 8 thinking tokens in the native trace
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-sonnet-4-6 via Anthropic
{
"model": "claude-sonnet-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/"
}
Evidence: no thinking parameter was sent
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
moonshotai/kimi-k3 via Cloudflare
{
"model": "moonshotai/kimi-k3",
"speculative_inference": true,
"base_url": "https://api.cloudflare.com/client/v4/accounts/19cbe0d2e934c6dc43052886b7cbd8f1/ai/v1"
}
Evidence: reasoning_content and non-zero reasoning tokens in matching probes
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
moonshotai/kimi-k3 via Cloudflare, reasoning_effort=none
{
"model": "moonshotai/kimi-k3",
"speculative_inference": true,
"base_url": "https://api.cloudflare.com/client/v4/accounts/19cbe0d2e934c6dc43052886b7cbd8f1/ai/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: no reasoning_content or reasoning tokens in the matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
accounts/fireworks/routers/kimi-k3-fast via Fireworks
{
"model": "accounts/fireworks/routers/kimi-k3-fast",
"speculative_inference": true,
"base_url": "https://api.fireworks.ai/inference/v1"
}
Evidence: reasoning_content and non-zero reasoning tokens in the matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
mercury-2 medium reasoning via Inception
{
"model": "mercury-2 medium reasoning via Inception",
"configuration_capture": "builder model only; other settings not captured",
"reasoning_profile": {
"mode": "on",
"setting": "reasoning_effort=medium",
"basis": "matching Inception provider probe",
"verified": true,
"evidence": "44 reasoning tokens in the matching direct probe"
}
}
Evidence: 44 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
mercury-2 instant reasoning via Inception
{
"model": "mercury-2 instant reasoning via Inception",
"configuration_capture": "builder model only; other settings not captured",
"reasoning_profile": {
"mode": "lowest",
"setting": "reasoning_effort=instant",
"basis": "Mercury Instant is the provider's lowest-latency reasoning mode",
"verified": true,
"evidence": "5 reasoning tokens remained in the matching direct probe"
}
}
Evidence: 5 reasoning tokens remained in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
grok-4.5 via xAI
{
"model": "grok-4.5",
"speculative_inference": true,
"base_url": "https://api.x.ai/v1"
}
Evidence: 192 reasoning tokens and reasoning_content in the matching high probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
grok-4.5 via xAI, reasoning_effort=low
{
"model": "grok-4.5",
"speculative_inference": true,
"base_url": "https://api.x.ai/v1",
"extra_body": {
"reasoning_effort": "low"
}
}
Evidence: 92 reasoning tokens and reasoning_content remained in the low probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-glm-4.7 managed baseline alias
{
"model": "tavus-glm-4.7",
"speculative_inference": true
}
Evidence: the managed baseline alias is explicit; no reasoning control or trace is exposed
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-gemma-4-thinking managed alias (Cerebras)
{
"model": "tavus-gemma-4-thinking",
"speculative_inference": true
}
Evidence: 685 reasoning tokens in bake-off conversation c62f39421e7d8477
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-gemma-4 managed alias (Cerebras)
{
"model": "tavus-gemma-4",
"speculative_inference": true
}
Evidence: non-thinking managed alias selected explicitly
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-2.5-flash via Google AI Studio
{
"model": "gemini-2.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/"
}
Evidence: 285 hidden thought tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-2.5-flash via Google AI Studio
{
"model": "gemini-2.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: 0 hidden thought tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-3.5-flash via Google AI Studio
{
"model": "gemini-3.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/"
}
Evidence: 261 hidden thought tokens and a thought signature in the matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemini-3.5-flash via Google AI Studio
{
"model": "gemini-3.5-flash",
"speculative_inference": true,
"base_url": "https://generativelanguage.googleapis.com/v1beta/openai/",
"extra_body": {
"reasoning_effort": "minimal"
}
}
Evidence: 0 billed hidden tokens on a simple probe, but a thought signature remained
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-luna via OpenAI, reasoning_effort=medium
{
"model": "gpt-5.6-luna",
"speculative_inference": true,
"base_url": "https://coaches-bones-railway-historical.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "medium"
}
}
Evidence: 26 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-luna via OpenAI, reasoning_effort=none
{
"model": "gpt-5.6-luna",
"speculative_inference": true,
"base_url": "https://api.openai.com/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: 0 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-sol via OpenAI, reasoning_effort=high
{
"model": "gpt-5.6-sol",
"speculative_inference": true,
"base_url": "https://coaches-bones-railway-historical.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "high"
}
}
Evidence: 3400 reasoning tokens on a complex matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-5.6-sol via OpenAI, reasoning_effort=none
{
"model": "gpt-5.6-sol",
"speculative_inference": true,
"base_url": "https://api.openai.com/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: 0 reasoning tokens in the matching direct probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-opus-4-6 via Anthropic, extended thinking
{
"model": "claude-opus-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/",
"extra_body": {
"thinking": {
"type": "enabled",
"budget_tokens": 1024
}
}
}
Evidence: signed thinking block and 41 thinking tokens in the native trace
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-opus-4-6 via Anthropic
{
"model": "claude-opus-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/"
}
Evidence: no thinking parameter was sent
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-sonnet-4-6 via Anthropic, extended thinking
{
"model": "claude-sonnet-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/",
"extra_body": {
"thinking": {
"type": "enabled",
"budget_tokens": 1024
}
}
}
Evidence: signed thinking block and 8 thinking tokens in the native trace
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
claude-sonnet-4-6 via Anthropic
{
"model": "claude-sonnet-4-6",
"speculative_inference": true,
"base_url": "https://api.anthropic.com/v1/"
}
Evidence: no thinking parameter was sent
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
moonshotai/kimi-k3 via Cloudflare
{
"model": "moonshotai/kimi-k3",
"speculative_inference": true,
"base_url": "https://api.cloudflare.com/client/v4/accounts/19cbe0d2e934c6dc43052886b7cbd8f1/ai/v1"
}
Evidence: reasoning_content and non-zero reasoning tokens in matching probes
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
moonshotai/kimi-k3 via Cloudflare, reasoning_effort=none
{
"model": "moonshotai/kimi-k3",
"speculative_inference": true,
"base_url": "https://api.cloudflare.com/client/v4/accounts/19cbe0d2e934c6dc43052886b7cbd8f1/ai/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
Evidence: no reasoning_content or reasoning tokens in the matching probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
grok-4.5 via xAI
{
"model": "grok-4.5",
"speculative_inference": true,
"base_url": "https://api.x.ai/v1"
}
Evidence: 192 reasoning tokens and reasoning_content in the matching high probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
grok-4.5 via xAI, reasoning_effort=low
{
"model": "grok-4.5",
"speculative_inference": true,
"base_url": "https://api.x.ai/v1",
"extra_body": {
"reasoning_effort": "low"
}
}
Evidence: 92 reasoning tokens and reasoning_content remained in the low probe
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
google/gemma-4-31b-it via Crusoe serverless
{
"model": "google/gemma-4-31b-it",
"speculative_inference": true,
"base_url": "https://api.inference.crusoecloud.com/v1/"
}
Evidence: tool-calling probe returned finish_reason=tool_calls with no reasoning_content and reasoning_effort was not sent
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
google/gemma-4-31b-it via Crusoe serverless
{
"model": "google/gemma-4-31b-it",
"speculative_inference": true,
"base_url": "https://api.inference.crusoecloud.com/v1/"
}
Evidence: tool-calling probe returned finish_reason=tool_calls with no reasoning_content and reasoning_effort was not sent
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
google/gemma-4-31b-it via Crusoe serverless, throttled by the provider
{
"model": "google/gemma-4-31b-it",
"speculative_inference": true,
"base_url": "https://api.inference.crusoecloud.com/v1/"
}
Evidence: tool-calling probe returned finish_reason=tool_calls with no reasoning_content and reasoning_effort was not sent; the provider returned HTTP 429 max_requests_limit during this run; 12 turns produced no output
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B via Wafer dedicated B200 endpoint, prod RQH
{
"model": "Gemma-4-31B",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/v1",
"extra_body": {
"thinking": {
"type": "disabled"
}
}
}
Evidence: default and thinking-disabled probes returned no reasoning_content with completion_tokens_details.reasoning_tokens=0; a thinking-disabled tool-calling probe returned finish_reason=tool_calls with valid arguments
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B via Wafer dedicated B200 endpoint, prod RQH
{
"model": "Gemma-4-31B",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/v1",
"extra_body": {
"thinking": {
"type": "disabled"
}
}
}
Evidence: same-day probe returned reasoning_tokens=0 by default and with thinking disabled, and a thinking-disabled tool probe returned finish_reason=tool_calls with valid arguments
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-gemma-4 managed alias, served by Crusoe gemma-4-responsive, prod RQH
{
"model": "tavus-gemma-4",
"speculative_inference": true
}
Evidence: every logged request for conversation cb0ccde83c6a84fc was served by [email protected] on attempt 1 with no fallback and no reasoning parameter
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemma-4-31b via the Cerebras API directly, prod RQH
{
"model": "gemma-4-31b",
"speculative_inference": true,
"base_url": "https://api.cerebras.ai/v1"
}
Evidence: probe returned completion_tokens_details.reasoning_tokens=0 with no reasoning parameter sent; the run sends none
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
GLM-5.2 via Wafer dedicated B200 endpoint, prod RQH
{
"model": "GLM-5.2",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/v1",
"extra_body": {
"thinking": {
"type": "disabled"
}
}
}
Evidence: default, thinking-disabled, and reasoning_effort=none probes all returned reasoning_tokens=0 with clean content; a thinking-enabled probe cleanly emitted reasoning_content (169 reasoning tokens), so off-by-default is configuration, not a capability gap; a thinking-disabled streamed two-tool probe returned valid per-call JSON
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
GLM-5.2 via Wafer dedicated B200 endpoint, prod RQH
{
"model": "GLM-5.2",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/v1",
"extra_body": {
"thinking": {
"type": "disabled"
}
}
}
Evidence: same-day probe returned reasoning_tokens=0 by default and with thinking disabled, and a thinking-disabled tool probe returned finish_reason=tool_calls with valid arguments
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
tavus-gemma-4 managed alias, served by Crusoe gemma-4-responsive, prod RQH, 2026-09-15
{
"model": "tavus-gemma-4",
"speculative_inference": true
}
Evidence: 24 of 24 requests for conversation cb417c4233d5e4d6 were served by [email protected] on attempt 1 with reasoning_settings chat_template_kwargs.enable_thinking=false, no fallback hops, no [LLM error] or ttft_timeout lines; the harness sent no reasoning_effort or extra_body
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemma-4-31b via Hopper (api.withhopper.com), prod RQH, 2026-09-15
{
"model": "gemma-4-31b",
"speculative_inference": true,
"base_url": "https://api.withhopper.com/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
Evidence: same-day probes of gemma-4-31b returned no reasoning_content and no reasoning token usage by default, with the knobs pinned, and even with enable_thinking=true; tool probes returned finish_reason=tool_calls with valid JSON arguments non-streamed, as clean streamed deltas, and for parallel calls, and the tool-result round trip returned plain content
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-NVFP4 via Wafer dedicated B300 endpoint, prod RQH, 2026-09-15
{
"model": "Gemma-4-31B-IT-NVFP4",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
Evidence: post-fix probes of Gemma-4-31B-IT-NVFP4 returned usage.reasoning_tokens=0 with and without the knobs, tool-result round trips returned plain content with no channel markup streamed or not, and streamed tool calls (a long draft_pal brief, parallel calls, a follow-on call after a result) accumulated to valid JSON arguments with finish_reason=tool_calls
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 via Wafer dedicated B300 FP8 endpoint, prod RQH, 2026-09-16
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
Evidence: same-day probes of Gemma-4-31B-IT-FP8 returned usage.reasoning_tokens=0 with and without the knobs, tool-result round trips returned plain content with no channel markup streamed or not, and streamed tool calls (a long draft_pal brief, parallel calls, a follow-on call after a result) accumulated to valid JSON arguments with finish_reason=tool_calls
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 via Wafer dedicated B300 FP8 endpoint, prod RQH on CVI canary 3.339.0 with the HTTP/2 client, 2026-09-17
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
Evidence: FP8 probes returned usage.reasoning_tokens=0 with and without the knobs and clean streamed tool arguments; the CVI's FEATURE_FLAGS_LOADED snapshot for conversation c0a3d394f140c4bc on pod phoenix4-canary (service.version 3.339.0) shows cvi_hopper_client True, and all 22 requests carried reasoning_effort none plus enable_thinking=false to tavus.wafer.ai/fp8 on attempt 1
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 via Wafer dedicated B300 FP8 endpoint, prod RQH on CVI canary with the HTTP/2 client, 2026-09-17
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
Evidence: FP8 probes returned usage.reasoning_tokens=0 with and without the knobs and clean streamed tool arguments; the run's CVI FEATURE_FLAGS_LOADED snapshot, pod and service.version are recorded in findings/wafer-gemma4-dedicated.md; every request carried reasoning_effort none plus enable_thinking=false
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-26B-A4B-IT-FP8 via Wafer dedicated 8x B300 endpoint, prod RQH on CVI canary, 2026-09-21
{
"model": "Gemma-4-26B-A4B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/gemma26/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
Evidence: same-day probes of Gemma-4-26B-A4B-IT-FP8 at tavus.wafer.ai/gemma26/v1 with the knobs returned usage.reasoning_tokens=0 and reasoning_content null, two parallel streamed update_persona_field calls accumulated to valid JSON arguments with finish_reason=tool_calls, and the turn after the tool results returned plain speech with no channel markup streamed or not; the run's CVI audit for conversation cf8b96158a126492 on pod phoenix4-canary (service.version 3.343.0, cvi_hopper_client True) logged 20 LLM requests, 0 LLM errors, 0 Gemma tool-parse events and 0 spoken tool calls
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-luna via OpenAI Responses, reasoning.effort=none
{
"model": "gpt-6-luna",
"speculative_inference": true,
"base_url": "https://lesson-define-insulation-attempts.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
{
"prompt": {
"portal_commit": "1cce76b7c27a399830ccefbe588d801e02f161ad",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "7f94ec4d283f6911ed0bbd99c60b9b4b56ff48e9fe2783b8f35e20cab1b9d228",
"carrier_version": "0e41db90",
"carrier_characters": 39463,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-luna",
"effort": "none",
"requests": 37,
"input_tokens": 1141468,
"cached_tokens": 1021404,
"reasoning_tokens": 0,
"output_tokens": 3382,
"first_request_tokens": 18429,
"max_request_tokens": 40315
}
}
Evidence: Completed docent-full capture ca3d0f46db62c4d9: OpenAI Responses reported 0 reasoning tokens over 37 completed requests at reasoning.effort=none. CVI 3.345.0 audit recorded 0 LLM errors and 0 invalid tool batches. Per-request usage and the exact carrier hash are retained.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-luna via OpenAI Responses, reasoning.effort=low
{
"model": "gpt-6-luna",
"speculative_inference": true,
"base_url": "https://lesson-define-insulation-attempts.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low"
}
}
{
"prompt": {
"portal_commit": "1cce76b7c27a399830ccefbe588d801e02f161ad",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "7f94ec4d283f6911ed0bbd99c60b9b4b56ff48e9fe2783b8f35e20cab1b9d228",
"carrier_version": "0e41db90",
"carrier_characters": 39463,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-luna",
"effort": "low",
"requests": 32,
"input_tokens": 929353,
"cached_tokens": 789150,
"reasoning_tokens": 3574,
"output_tokens": 7157,
"first_request_tokens": 18427,
"max_request_tokens": 36376
}
}
Evidence: Completed docent-full capture c5e33b442a11d465: OpenAI Responses reported 3,574 reasoning tokens over 32 completed requests at reasoning.effort=low. CVI 3.345.0 audit recorded 0 LLM errors and 0 invalid tool batches. Per-request usage and the exact carrier hash are retained.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-sol via OpenAI Responses, reasoning.effort=none
{
"model": "gpt-6-sol",
"speculative_inference": true,
"base_url": "https://lesson-define-insulation-attempts.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "none"
}
}
{
"prompt": {
"portal_commit": "1cce76b7c27a399830ccefbe588d801e02f161ad",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "7f94ec4d283f6911ed0bbd99c60b9b4b56ff48e9fe2783b8f35e20cab1b9d228",
"carrier_version": "0e41db90",
"carrier_characters": 39463,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-sol",
"effort": "none",
"requests": 32,
"input_tokens": 844871,
"cached_tokens": 619395,
"reasoning_tokens": 0,
"output_tokens": 2647,
"first_request_tokens": 18429,
"max_request_tokens": 32612
}
}
Evidence: Completed docent-full capture c54ce809dbf13436: OpenAI Responses reported 0 reasoning tokens over 32 completed requests at reasoning.effort=none. CVI 3.345.0 audit recorded 0 LLM errors and 0 invalid tool batches. Per-request usage and the exact carrier hash are retained.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-sol via OpenAI Responses, reasoning.effort=low
{
"model": "gpt-6-sol",
"speculative_inference": true,
"base_url": "https://lesson-define-insulation-attempts.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low"
}
}
{
"prompt": {
"portal_commit": "1cce76b7c27a399830ccefbe588d801e02f161ad",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "7f94ec4d283f6911ed0bbd99c60b9b4b56ff48e9fe2783b8f35e20cab1b9d228",
"carrier_version": "0e41db90",
"carrier_characters": 39463,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-sol",
"effort": "low",
"requests": 29,
"input_tokens": 804802,
"cached_tokens": 694708,
"reasoning_tokens": 779,
"output_tokens": 4230,
"first_request_tokens": 18429,
"max_request_tokens": 34613
}
}
Evidence: Completed docent-full capture c51b17c106a7c435: OpenAI Responses reported 779 reasoning tokens over 29 completed requests at reasoning.effort=low. CVI 3.345.0 audit recorded 0 LLM errors and 0 invalid tool batches. Per-request usage and the exact carrier hash are retained.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-astra via OpenAI Responses, service_tier=ultrafast, reasoning.effort=low
{
"model": "gpt-6-astra",
"speculative_inference": true,
"base_url": "https://issn-laura-beth-notification.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "ultrafast"
}
}
{
"prompt": {
"portal_commit": "08e83ddce53d6fc3f7b7326ee086d49f3b1fcd93",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "962938c7acf8cd070a992a2cb5a71802e87037b5cb880c9a886224137767e2f6",
"carrier_version": "bf68ba0d",
"carrier_characters": 42849,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-astra",
"effort": "low",
"requests": 30,
"requested_service_tier": "ultrafast",
"service_tiers": [
"ultrafast"
],
"served_models": [
"gpt-6-astra"
],
"input_tokens": 789665,
"cached_tokens": 652718,
"reasoning_tokens": 158,
"output_tokens": 2753,
"first_request_tokens": 19422,
"max_request_tokens": 31751
}
}
Evidence: 30 matching Responses requests served as gpt-6-astra at ultrafast/low; 158 reasoning tokens in conversation c918295b844ab4c2. Low is the lowest supported mode, not off.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6.1-sol via OpenAI Responses, service_tier=default, reasoning.effort=low
{
"model": "gpt-6.1-sol",
"speculative_inference": true,
"base_url": "https://issn-laura-beth-notification.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "default"
}
}
{
"prompt": {
"portal_commit": "08e83ddce53d6fc3f7b7326ee086d49f3b1fcd93",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "962938c7acf8cd070a992a2cb5a71802e87037b5cb880c9a886224137767e2f6",
"carrier_version": "bf68ba0d",
"carrier_characters": 42849,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6.1-sol",
"effort": "low",
"requests": 31,
"requested_service_tier": "default",
"service_tiers": [
"default"
],
"served_models": [
"gpt-6.1-sol"
],
"input_tokens": 739281,
"cached_tokens": 661307,
"reasoning_tokens": 208,
"output_tokens": 2614,
"first_request_tokens": 19425,
"max_request_tokens": 29044
}
}
Evidence: 31 matching Responses requests served as gpt-6.1-sol at default/low; 208 reasoning tokens in conversation c0822b66ef33f4df. Low is the lowest supported mode, not off.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 directly via Wafer, thinking disabled, prod RQH
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "08e83ddce53d6fc3f7b7326ee086d49f3b1fcd93",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "962938c7acf8cd070a992a2cb5a71802e87037b5cb880c9a886224137767e2f6",
"carrier_version": "bf68ba0d",
"carrier_characters": 42849,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed context-on-demand capture c3ef321b297604f5: all 25 CVI requests used Gemma-4-31B-IT-FP8 directly at tavus.wafer.ai/fp8/v1 with reasoning_effort=none and enable_thinking=false. CVI 3.353.1 audit found zero LLM errors, invalid batches or spoken tool syntax. No provider reasoning-token total was captured.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemma-4-responsive directly via Crusoe, thinking disabled, prod RQH
{
"model": "gemma-4-responsive",
"speculative_inference": true,
"base_url": "https://api.inference.crusoecloud.com/v1",
"extra_body": {
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "08e83ddce53d6fc3f7b7326ee086d49f3b1fcd93",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "962938c7acf8cd070a992a2cb5a71802e87037b5cb880c9a886224137767e2f6",
"carrier_version": "bf68ba0d",
"carrier_characters": 42849,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed context-on-demand capture c9f0605e0770f45c: direct gemma-4-responsive at api.inference.crusoecloud.com/v1 with enable_thinking=false verified in CVI request logs; no managed-alias fallback. Provider reasoning-token totals were not captured.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-astra via OpenAI Responses HTTP, service_tier=ultrafast, reasoning.effort=low; adaptive discovery through completion
{
"model": "gpt-6-astra",
"speculative_inference": true,
"base_url": "https://route-distinction-actors-rebel.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "ultrafast"
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-astra",
"effort": "low",
"requests": 28,
"requested_service_tier": "ultrafast",
"service_tiers": [
"ultrafast"
],
"served_models": [
"gpt-6-astra"
],
"input_tokens": 909979,
"cached_tokens": 721266,
"reasoning_tokens": 274,
"output_tokens": 4155,
"first_request_tokens": 19476,
"max_request_tokens": 41452
}
}
Evidence: Completed adaptive capture c9d9be2c50092483: all 28 matching Responses requests served gpt-6-astra at ultrafast/low, with 274 reasoning tokens. CVI 3.356.0 recorded no LLM errors, invalid tool batches, or tool syntax sent to speech. Low is the lowest supported mode, not off.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6.1-sol via OpenAI Responses HTTP, service_tier=default, reasoning.effort=low; adaptive discovery through completion
{
"model": "gpt-6.1-sol",
"speculative_inference": true,
"base_url": "https://route-distinction-actors-rebel.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "default"
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6.1-sol",
"effort": "low",
"requests": 22,
"requested_service_tier": "default",
"service_tiers": [
"default"
],
"served_models": [
"gpt-6.1-sol"
],
"input_tokens": 612176,
"cached_tokens": 503767,
"reasoning_tokens": 212,
"output_tokens": 2966,
"first_request_tokens": 19476,
"max_request_tokens": 33034
}
}
Evidence: Completed adaptive capture cc57d2af27993463: all 22 matching Responses requests served gpt-6.1-sol at default/low, with 212 reasoning tokens. CVI 3.356.0 recorded no LLM errors, invalid tool batches, or tool syntax sent to speech. Low is the lowest supported mode, not off.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 directly via Wafer, thinking disabled, RQH #3142 preview; adaptive discovery through completion
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed adaptive capture ce993455063e0428 used Gemma-4-31B-IT-FP8 at tavus.wafer.ai/fp8/v1 with reasoning_effort=none and enable_thinking=false. CVI 3.356.0 recorded 21 requests, no invalid tool batches or tool syntax sent to speech. A native draft tool timeout and late successful result are retained as a runtime finding.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemma-4-responsive directly via Crusoe, thinking disabled, RQH #3142 preview; adaptive discovery through completion
{
"model": "gemma-4-responsive",
"speculative_inference": true,
"base_url": "https://api.inference.crusoecloud.com/v1",
"extra_body": {
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed adaptive capture cdfea25af66cd463 used gemma-4-responsive at api.inference.crusoecloud.com/v1 with enable_thinking=false in the CVI request logs. CVI 3.356.0 recorded 24 requests, no LLM errors, invalid tool batches, spoken tool syntax, or native tool timeouts.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-astra via Responses WebSocket, service_tier=ultrafast, reasoning.effort=low; adaptive discovery through completion
{
"model": "gpt-6-astra",
"speculative_inference": true,
"base_url": "https://route-distinction-actors-rebel.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "ultrafast"
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-astra",
"effort": "low",
"requests": 30,
"transports": [
"websocket"
],
"websocket_requests": 30,
"websocket_reused_requests": 29,
"requested_service_tier": "ultrafast",
"service_tiers": [
"ultrafast"
],
"served_models": [
"gpt-6-astra"
],
"input_tokens": 1121216,
"cached_tokens": 835167,
"reasoning_tokens": 113,
"output_tokens": 4168,
"first_request_tokens": 19473,
"max_request_tokens": 49090
}
}
Evidence: Completed adaptive capture c1efb469aafbb429: all 30 provider requests used WebSockets at ultrafast/low; 29 reused the open connection. Served models: gpt-6-astra. 113 reasoning tokens. CVI 3.356.0 recorded 0 LLM errors and 0 invalid tool batches. Low is the lowest supported mode, not off.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6.1-sol via Responses WebSocket, service_tier=default, reasoning.effort=low; adaptive discovery through completion
{
"model": "gpt-6.1-sol",
"speculative_inference": true,
"base_url": "https://route-distinction-actors-rebel.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "default"
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6.1-sol",
"effort": "low",
"requests": 27,
"transports": [
"websocket"
],
"websocket_requests": 27,
"websocket_reused_requests": 26,
"requested_service_tier": "default",
"service_tiers": [
"default"
],
"served_models": [
"gpt-6.1-sol"
],
"input_tokens": 777023,
"cached_tokens": 673888,
"reasoning_tokens": 138,
"output_tokens": 3605,
"first_request_tokens": 19473,
"max_request_tokens": 35596
}
}
Evidence: Completed adaptive capture c9401bc2a869c4eb: all 27 provider requests used WebSockets at default/low; 26 reused the open connection. Served models: gpt-6.1-sol. 138 reasoning tokens. CVI 3.356.0 recorded 0 LLM errors and 0 invalid tool batches. Low is the lowest supported mode, not off.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 directly via Wafer, thinking disabled; shared guard race fix; adaptive discovery through completion
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed shared-guard-fix capture cda9f56e7f545480 used Gemma-4-31B-IT-FP8 directly at tavus.wafer.ai with thinking disabled, verified in CVI request logs. CVI 3.356.0 recorded 20 requests, 0 LLM errors, 0 invalid tool batches, 0 spoken tool-syntax turns, 0 native tool timeouts, and 0 discarded results.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gemma-4-responsive directly via Crusoe, thinking disabled; shared guard race fix; adaptive discovery through completion
{
"model": "gemma-4-responsive",
"speculative_inference": true,
"base_url": "https://api.inference.crusoecloud.com/v1",
"extra_body": {
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pbf29388c3a4",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed shared-guard-fix capture cac14ea5cbf7a481 used gemma-4-responsive directly at api.inference.crusoecloud.com with thinking disabled, verified in CVI request logs. CVI 3.356.0 recorded 28 requests, 0 LLM errors, 1 invalid tool batches, 0 spoken tool-syntax turns, 0 native tool timeouts, and 0 discarded results.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-26B-A4B-IT-FP8 directly via Wafer, thinking disabled; adaptive discovery through completion
{
"model": "Gemma-4-26B-A4B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/gemma26/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "62774194b8c2ab1bc5405477b5b8e8126ada0e19",
"portal_dirty": false,
"carrier_persona_id": "pd7c6eea24bb",
"carrier_sha256": "a44087f118d69097a284153d9113e5bd0d0e17b90372da370398161cc34cf118",
"carrier_version": "95776bf7",
"carrier_characters": 42874,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: Completed capture ca1c12fa92eb2422: all 34 CVI requests used Gemma-4-26B-A4B-IT-FP8 directly at tavus.wafer.ai/gemma26/v1 with reasoning_effort=none and enable_thinking=false. CVI 3.356.0: 0 LLM errors, 2 invalid-JSON tool batches, 8 responses with tool syntax sent to speech, 0 tool timeouts or discarded results. Supplemental provider probes returned reasoning_tokens=0; no conversation-wide provider token total was captured.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 directly via Wafer with thinking disabled; same carrier as Muse Spark Open
{
"model": "Gemma-4-31B-IT-FP8",
"speculative_inference": true,
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "1f2483b3dd236e251ab1479d9b57d19ed30dfc97",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "abee42e04b71909af259c6249a3c6adfd3678333f3837570881a03cb7c6c61dd",
"carrier_version": "4773c607",
"carrier_characters": 54797,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: 2026-10-15 capture ce427e8e9385143e uses Gemma-4-31B-IT-FP8 directly at tavus.wafer.ai/fp8/v1 with reasoning_effort=none and enable_thinking=false. Its carrier hash matches both Muse Spark Open arms. Provider reasoning-token totals were not captured. CVI 3.370.0 recorded 24 requests, 0 LLM errors, 0 invalid batches, 1 spoken tool-syntax turn, 1 tool timeout, and 1 dropped late result.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6.1-sol via OpenAI Responses HTTP, service_tier=ultrafast, reasoning.effort=low; same carrier as Muse/Wafer comparison
{
"model": "gpt-6.1-sol",
"speculative_inference": true,
"base_url": "https://broad-lived-cio-preferred.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "ultrafast"
}
}
{
"prompt": {
"portal_commit": "1f2483b3dd236e251ab1479d9b57d19ed30dfc97",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "abee42e04b71909af259c6249a3c6adfd3678333f3837570881a03cb7c6c61dd",
"carrier_version": "4773c607",
"carrier_characters": 54797,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6.1-sol",
"effort": "low",
"requests": 31,
"transports": [
"http"
],
"websocket_requests": 0,
"websocket_reused_requests": 0,
"requested_service_tier": "ultrafast",
"service_tiers": [
"ultrafast"
],
"served_models": [
"gpt-6.1-sol"
],
"input_tokens": 1153572,
"cached_tokens": 944151,
"reasoning_tokens": 138,
"output_tokens": 3950,
"first_request_tokens": 24755,
"max_request_tokens": 47832
}
}
Evidence: 31 Responses HTTP requests in c982cf2c84a6b41e returned gpt-6.1-sol and ultrafast at low effort; 138 reasoning tokens.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
gpt-6-astra via OpenAI Responses HTTP, service_tier=ultrafast, reasoning.effort=low; same carrier as Muse/Wafer comparison
{
"model": "gpt-6-astra",
"speculative_inference": true,
"base_url": "https://broad-lived-cio-preferred.trycloudflare.com/v1",
"extra_body": {
"reasoning_effort": "low",
"service_tier": "ultrafast"
}
}
{
"prompt": {
"portal_commit": "1f2483b3dd236e251ab1479d9b57d19ed30dfc97",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "abee42e04b71909af259c6249a3c6adfd3678333f3837570881a03cb7c6c61dd",
"carrier_version": "4773c607",
"carrier_characters": 54797,
"context_placement": "trailing"
},
"provider_usage": {
"model": "gpt-6-astra",
"effort": "low",
"requests": 37,
"transports": [
"http"
],
"websocket_requests": 0,
"websocket_reused_requests": 0,
"requested_service_tier": "ultrafast",
"service_tiers": [
"ultrafast"
],
"served_models": [
"gpt-6-astra"
],
"input_tokens": 1557732,
"cached_tokens": 1244804,
"reasoning_tokens": 270,
"output_tokens": 4109,
"first_request_tokens": 24752,
"max_request_tokens": 51647
}
}
Evidence: 37 Responses HTTP requests in cdd4852dd453046c returned gpt-6-astra and ultrafast at low effort; 270 reasoning tokens. Earlier read_pal_section harness failure is unranked.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
meta-models/muse-spark-open directly via Baseten, reasoning_effort=low; production RQH
{
"model": "meta-models/muse-spark-open",
"speculative_inference": true,
"base_url": "https://inference.baseten.co/v1",
"extra_body": {
"reasoning_effort": "low",
"temperature": 1.0,
"max_tokens": 16384
}
}
{
"prompt": {
"portal_commit": "1f2483b3dd236e251ab1479d9b57d19ed30dfc97",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "abee42e04b71909af259c6249a3c6adfd3678333f3837570881a03cb7c6c61dd",
"carrier_version": "4773c607",
"carrier_characters": 54797,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: 2026-10-15 direct Baseten low streamed probe returned valid ping arguments. Meta documents low/medium/high/xhigh and no reasoning-off mode. No per-CVI reasoning token count inferred. https://dev.meta.ai/docs/muse-spark-open/settings
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
meta-models/muse-spark-open directly via Baseten, reasoning_effort=medium; production RQH
{
"model": "meta-models/muse-spark-open",
"speculative_inference": true,
"base_url": "https://inference.baseten.co/v1",
"extra_body": {
"reasoning_effort": "medium",
"temperature": 1.0,
"max_tokens": 16384
}
}
{
"prompt": {
"portal_commit": "1f2483b3dd236e251ab1479d9b57d19ed30dfc97",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "abee42e04b71909af259c6249a3c6adfd3678333f3837570881a03cb7c6c61dd",
"carrier_version": "4773c607",
"carrier_characters": 54797,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: 2026-10-15 direct Baseten medium streamed probe returned valid ping arguments. Meta documents low/medium/high/xhigh and no reasoning-off mode. No per-CVI reasoning token count inferred. https://dev.meta.ai/docs/muse-spark-open/settings
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
inworld/models/gemma-4-31b-it via Inworld Inference with thinking disabled; production carrier 20dc098b, GPT-6 Luna creator
{
"model": "inworld/models/gemma-4-31b-it",
"speculative_inference": true,
"base_url": "https://api.inworld.ai/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal_commit": "bf3f8569f07182d1bb5b631b2540d863afb38514",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "3551d5944e8b0dbe3d788079a276aee4e90d015ccc3bc75b37f26a8efd27b28f",
"carrier_version": "20dc098b",
"carrier_characters": 55222,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: 2026-10-16 capture c1fc06863f69044f on production carrier 20dc098b: all 23 CVI [LLM request] rows used inworld/models/gemma-4-31b-it at api.inworld.ai/v1, is_tavus_managed=false, with reasoning_effort=none and enable_thinking=false (CVI 3.376.0, SigNoz). 0 LLM errors and invalid tool batches; 1 native tool-call turn spoken aloud; 1 draft-tool timeout with its late result dropped. Inworld's own default is non-thinking; only an explicit reasoning effort produced reasoning text in probes.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
mercury-3 directly via Inception, reasoning_effort=low, strict tool schemas; October 19 production carrier and GPT-6 Sol creator
{
"model": "mercury-3",
"speculative_inference": true,
"base_url": "https://api.inceptionlabs.ai/v1#groq",
"extra_body": {
"reasoning_effort": "low",
"max_tokens": 16384
}
}
{
"prompt": {
"portal_commit": "0a4aff7c39fc681977ccc70f06939c3b7b62d614",
"portal_dirty": false,
"carrier_persona_id": "p1abc173c043",
"carrier_sha256": "3551d5944e8b0dbe3d788079a276aee4e90d015ccc3bc75b37f26a8efd27b28f",
"carrier_version": "20dc098b",
"carrier_characters": 55222,
"context_placement": "trailing"
},
"provider_usage": {}
}
Evidence: 2026-10-19 direct Inception probe returned mercury-3 with valid streamed ping arguments at low effort. Inception documents instant/low/medium/high; low is not the lowest mode. No conversation-wide reasoning token count inferred. https://docs.inceptionlabs.ai/capabilities/reasoning-efforts Completed c098b52f4162d458: all 30 CVI request logs use mercury-3 directly at api.inceptionlabs.ai/v1 with reasoning_effort=low and is_tavus_managed=false (CVI 3.376.0).
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
mercury-3 low reasoning; real Portal PR 1913 on test RQH; museum PDF; Gemini 3.6 Flash creator
{
"model": "mercury-3",
"base_url": "https://api.inceptionlabs.ai/v1#groq",
"extra_body": {
"reasoning_effort": "low",
"max_tokens": 16384
}
}
{
"prompt": {
"portal": {
"url": "https://charlie-guided-closeout.tavus-preview.io",
"mode": "remote",
"branch": "sam/charlie-guided-closeout",
"commit": "663ed67e02c2784d0929899cd1149c34b0842946",
"checkout_branch": null,
"dirty": null,
"commit_source": "GitHub preview deployment 39322604624 and deployed-head sticky comment"
},
"rqh": {
"url": "https://test.rqh.tavusapi.com",
"mode": "remote",
"container": null,
"branch": null,
"observed_origin": "https://test.rqh.tavusapi.com",
"configured_matches_observed": true,
"preview_pr": null,
"commit": null,
"commit_source": null
},
"carrier": {
"persona_id": "p75e99695c86",
"prompt_version": "992c1fbb",
"prompt_version_source": "browser",
"stale_warning": null,
"status": "observed"
},
"cvi": {
"mode": "hosted",
"target": "p-509ff669-phoenix4-canary-deployment-6c9f844458-496cf",
"service_version": "3.379.0",
"source": "CVI audit logs"
},
"builder_llm_model": "mercury-3",
"built_persona_model": "tavus-gemma-4"
},
"provider_usage": {}
}
Evidence: Conversation cdc866f4fd519482 logs identify mercury-3, direct Inception endpoint and reasoning_effort=low; launcher output budget 16384. Carrier overlay 992c1fbb, Portal 663ed67e0.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
Gemma-4-31B-IT-FP8 thinking disabled; real Portal PR 1913 on test RQH; museum PDF; Gemini 3.6 Flash creator
{
"model": "Gemma-4-31B-IT-FP8",
"base_url": "https://tavus.wafer.ai/fp8/v1",
"extra_body": {
"reasoning_effort": "none",
"chat_template_kwargs": {
"enable_thinking": false
}
}
}
{
"prompt": {
"portal": {
"url": "https://charlie-guided-closeout.tavus-preview.io",
"mode": "remote",
"branch": "sam/charlie-guided-closeout",
"commit": "663ed67e02c2784d0929899cd1149c34b0842946",
"checkout_branch": null,
"dirty": null,
"commit_source": "GitHub preview deployment 39322604624 and deployed-head sticky comment"
},
"rqh": {
"url": "https://test.rqh.tavusapi.com",
"mode": "remote",
"container": null,
"branch": null,
"observed_origin": "https://test.rqh.tavusapi.com",
"configured_matches_observed": true,
"preview_pr": null,
"commit": null,
"commit_source": null
},
"carrier": {
"persona_id": "p75e99695c86",
"prompt_version": "992c1fbb",
"prompt_version_source": "browser",
"stale_warning": null,
"status": "observed"
},
"cvi": {
"mode": "hosted",
"target": "p-509ff669-phoenix4-canary-deployment-6c9f844458-k5nwx",
"service_version": "3.379.0",
"source": "CVI audit logs"
},
"builder_llm_model": "Gemma-4-31B-IT-FP8",
"built_persona_model": "tavus-gemma-4"
},
"provider_usage": {}
}
Evidence: Conversation cb8c3480aafe3451 logs identify Gemma-4-31B-IT-FP8, Wafer endpoint, reasoning_effort=none and enable_thinking=false. Carrier overlay 992c1fbb, Portal 663ed67e0.
A blank turn with a small max_tokens or reasoning budget is configuration evidence, not automatically a model-quality failure.
docent-full. The adaptive creator pursues the same museum-docent requirements across every arm while responding to what each Charlie actually said. The PR #1913 arms preflight the same relevant museum PDF; older captures retain their original source fixtures. Each current run preflights a live readable/presentable PDF and disables the generic per-turn acknowledgement; explicit long-running draft narration remains. Reviewed standings = 25% autonomy + 35% final build completeness + 10% creator requirement coverage + 25% reliability + 5% continuous first-audio latency. Legacy: Standings score = 45% autonomy + 35% build + 10% requirements + 5% reliability + 5% latency. Reviewed v4 reliability is capped by first-response tool-call success; legacy scores are ranked separately. Autonomy itself is 70% self-handoff coverage and 30% guard-free turns. Binary checks are diagnostic gates, not additive weights. Provider/runtime failures are excluded instead of receiving a zero. Timeline and transcript inspectors are rendered by the same core/timing_viz.py component used by the Timeline family.