QORUM INTELLIGENCE — CONTENT INVENTORY FOR REVIEW
Evidence snapshot: 20 October 2026. This is a coverage outline, not a new slide design.
Purpose: explain the work behind natural business conversation, what is improving,
and why Qorum's intelligence platform can become a valuable recurring service.
Read the numbered sections as the full proposed coverage, not twelve extra slides.
The intended format remains one animated page with a controlled progression.
1. THE EXPERIENCE WE ARE TRYING TO CREATE
Speak naturally; be understood; keep business context; find information; take the
right permitted action; hear a useful response; interrupt or correct it naturally.
Show the whole chain: microphone -> end-of-turn detection -> transcription ->
context and memory -> reasoning -> tools -> speech -> playback and interruption.
Success is reliable work completed as well as a quick conversational response.
2. WHY THIS IS HARD
Every stage adds delay or uncertainty. Silence detection can cut someone off or
wait too long. Transcripts can be wrong. Context can be expensive to assemble.
Tools may need several reasoning rounds. Generated text still needs synthesising,
delivery and playback. The next turn must know what the caller actually heard.
Distinguish first acknowledgement, first meaningful audio and completed action.
Do not imply that a filler cue makes the underlying task finish faster.
3. WHO DOES WHAT — CONFIGURED PROVIDERS AND MODELS
Source: schemas/qorum/qortex/manifest.yaml, read on 20 October.
This establishes configuration, not that every model was used in a live request.
Cerebras: fast text inference for the conversation slot; Qwen3.8-27B, low effort.
The DeepInfra equivalent is configured as fallback. Cerebras is not the audio layer.
DeepInfra: Qwen text models across instant, standard and advanced tiers;
Qwen3-Embedding-8B for retrieval; Qwen ASR 0.6B/1.7B and hosted Whisper turbo
for transcription options; MiMo-V2.5-tts for a speech option.
Deepgram: Nova-3 transcription; Flux live listening and turn detection; Aura-2
speech with Vesta as the configured default in the voice profile. A separate Flux
TTS family/Gemma profile exists in the local catalogue. Label listening and speech
separately, since the family name alone does not establish its modality.
Optional profiles: Google Gemini and Anthropic Claude for text tiers; OpenAI
text-embedding-3-small appears in those profiles. See configured-models.csv for
all 20 catalogue rows and exact configured remote IDs.
OCR/document intelligence: provider and model choice is not finalized. QRM-1448
is Parked. It records earlier Qwen-VL candidates hosted by DeepInfra, but no
confirmed fourth dedicated OCR provider. Document extraction, image description
and ordinary native-text extraction must be distinguished. Do not present the
pending OCR pipeline as a finished, operating capability.
4. WORK ALREADY TRACKED AS COMPLETED
Read-only Linear snapshot; Done is the issue's status, not a fresh demonstration
or end-to-end verification in this presentation session.
- Dictation, faster ASR and provider adapters: QRM-1372, 1387, 1389, 1390.
- Speech foundation, profiles and usage accounting: QRM-1374, 1375, 1391, 1412.
- Call session/API and continuing an existing chat: QRM-1384, 1408, 1409.
- Spoken rendering, read-aloud and tool results: QRM-1386, 1394, 1443.
- Call controls, holding behaviour and voice settings: QRM-1411, 1414, 1415,
1421, 1422.
- Latency measurement on server and client: QRM-1426, 1429.
- Conversation-specific model/effort and Cerebras routing: QRM-1427, 1445.
- Finding information and locating real UI targets: QRM-1406, 1437.
5. WORK STILL MOVING THROUGH REVIEW AND DELIVERY
- Monitoring: web call engine and pipelined playback/interruption, QRM-1381/1385.
- In Sprint: server short-first response release, QRM-1428; Safari/iPhone audio
capture fallbacks, QRM-1447.
- In Review: live Flux call client, QRM-1410; concurrent memory recall with bounded
waiting and assembly timing, QRM-1449.
- In Progress: honest, voice-auditioned acknowledgement corpus/API, QRM-1450.
- Backlog: stream released speech units to the client, QRM-1431; acknowledgement
timing/cadence, QRM-1451; interrupt handling, stale-reply suppression and echo
guard, QRM-1452/1453.
- Backlog: tool-correctness/effort evaluations, QRM-1438; spoken choices and
confirmations, QRM-1439/1440; natural end-call behaviour, QRM-1425/1430;
persistent call history, QRM-1432; voice administration and further bindings.
The full-duplex parent QRM-358 remains Blocked. Child progress does not mean the
whole conversational experience is complete. Statuses may change after this date.
6. WHERE THE TIME ACTUALLY GOES
Measure stages before choosing the next optimization. A historical five-turn
deployed-instance observation in QRM-1449 (20 October, 06:51-06:54 UTC) recorded:
- Memory query embedding: 0.94-3.06 seconds.
- Server first token: 2.29-5.22 seconds.
- Cerebras conversation steps: 0.41-1.25 seconds.
These measurements describe a small diagnostic sample, not an overall benchmark;
the stages overlap and should not simply be added. They motivated concurrent
memory work, still In Review in this snapshot. Fast model inference alone cannot
remove a serial context bottleneck.
QRM-1426's no-tool median <=2s and p90 <=3.5s are targets, not achieved results.
7. INTELLIGENCE, SPEED AND TRUST
Use the lowest reasoning effort that passes the real task evaluations. A fast
answer that skips the search/navigation tool is not a successful optimization.
Explain bounded tool loops, correct arguments, confirmations, permissions,
business scope, auditability, fallbacks and truthful error handling. Each extra
model round can reprocess context and add cost as well as time.
Short speech and streaming can improve responsiveness; long reasoning and complex
tools still need time. A neutral acknowledgement must never claim success early.
8. MAKING THE CONVERSATION FEEL NATURAL
Cover endpointing, partial/live words, short-first speech, pronunciation/rendering,
cadence, cues, interruption, echo prevention and suppressing obsolete replies.
Keep distinct what is completed, under review and planned. The proposed 300ms cue
is acknowledgement timing (QRM-1451), not a promise of a completed answer in 300ms.
Explain why mobile/Safari microphones and playback behaviour need separate work.
9. FUTURE OPTIONS — LABEL THEM AS FUTURE
Talker/reasoner separation (QRM-1442) is Parked: a fast conversational layer could
hand off to a stronger worker. It is not the current architecture being claimed.
OCR/document extraction choices (QRM-1448) are also Parked. Evaluate real documents,
quality, latency, cost and data handling before choosing the provider/profile.
Broader autopilot work must be separated from the current voice delivery tasks.
10. WHY QORUM RATHER THAN A STANDALONE ASSISTANT?
Claude and ChatGPT already support tools and integrations. Qorum's proposed value
is connected business information, governed actions, workflows, UI context,
organizational permissions, service/support and the complete user experience.
Qorum can still use Claude when it is the right engine for a task.
Describe 'our intelligence' as Qorum's orchestration, context, routing, business
tools and experience. This is not a claim that Qorum trained or owns the foundation
model weights. Externally hosted inference remains a purchased input.
11. THE COMMERCIAL MODEL
Two recurring sources: platform subscriptions and a metered/packaged business
intelligence service. Model selection, caching, context control and specialized
models can improve supplier costs, subject to equivalent quality and reliability.
Include speech, embeddings, OCR, retries, hosting and support in the real unit
economics. A text-token estimate is only one part of the delivered service cost.
Claude caching and cheaper Claude models can also lower the comparison baseline.
12. THE AGREED ILLUSTRATIVE 50-CLIENT EXAMPLE
Revenue: 50 clients x R10,000/month = R500,000/month.
Baseline assumption Optimized target assumption
AI supplier cost/client R5,000 R500
Total AI supplier costs R250,000 R25,000
Revenue less those costs R250,000 R475,000
This is R225,000 more monthly contribution before other costs, or 1.9 times the
baseline contribution. Supplier costs are ten times lower. It is not ten times
the profit, and neither column is net profit. R500 has not been proved as the cost
of serving the same workload at the same quality. A 100% markup is a 50% margin.
Use R5,000 as an explicit scenario, not a measured typical Claude bill.
CALIBRATING THE R5,000 SCENARIO
At an illustrative R18/USD, 5,000 Sonnet 5.5 model calls/month, each with 20,000
uncached input tokens and 1,000 billed output tokens, calculate to R4,500 in text
tokens. This is per client organization; users and automations share that volume.
One user request can require several model calls. Model choice, output/reasoning
length, tool rounds, caching and monthly use can move the bill substantially.
See pricing-notes.txt and claude-cost-scenarios.csv for the assumptions and sources.
PACKAGED EVIDENCE
linear-snapshot.csv: 50 relevant issue records, statuses and source links.
configured-models.csv: 20 configured catalogue entries, not usage telemetry.
claude-cost-scenarios.csv: nine calculated token-only workload scenarios.
economics-assumptions.json: the agreed revenue and illustrative cost comparison.
This inventory is represented in the six builds of the intelligence slide.
CLAUDE COST ASSUMPTIONS — 20 OCTOBER 2026
Official source checked:
https://platform.claude.com/docs/en/about-claude/pricing
Standard uncached API prices in USD per million input/output tokens:
Sonnet 5.5: $2 input / $10 output.
Opus 5.5: $4 input / $20 output.
Haiku 5.5: $0.10 input / $0.50 output, for each prompt up to 100,000 tokens.
The Haiku rate is higher above that prompt-length threshold.
Scenario assumptions, not measured customer usage:
- Exchange rate R18/USD is a planning assumption, not a current FX quote.
- Per model call: 20,000 uncached input tokens and 1,000 billed output tokens.
- Input includes system instructions, context, history, tool schemas/results.
- Billed output includes reasoning tokens where applicable, not only spoken text.
- A user request or conversational turn may involve multiple model calls.
- Monthly volume is per client organization, across its users and automations.
Formula:
USD = calls x ((20,000 x input_rate + 1,000 x output_rate) / 1,000,000)
ZAR = USD x 18
At 5,000 model calls/month:
Sonnet 5.5 = $250 = R4,500.
Opus 5.5 = $500 = R9,000.
Haiku 5.5 = $12.50 = R225.
These calculations exclude speech recognition, speech generation, separately
billed tool services, OCR from another provider, embeddings, hosting, support,
tax and payment/FX fees. They also exclude cache charges/discounts and batch
discounts. Prompt caching and task-appropriate smaller models can substantially
reduce a Claude baseline; those options must be included in a fair comparison.
R5,000/client/month is plausible for a specified business workload, but is not
an established average. R500/client/month for an optimized Qorum service remains
a hypothetical target until equivalent-workload measurements establish it.
Official Anthropic API pricing ↗
Download token scenarios · Download economics assumptions
orumALFHARON / 21 OCT 2026
ORU / ALONGSIDE YOUR TEAM
