Lab dashboard
start hereThis page replays one conversation through the production API, one contribution at a time, and shows what each analyser makes of it. Pick a conversation below, then open one section at a time — the cards are in pipeline order, from what arrives to what is derived from it. The replay controls at the top stay live while a section is open.
Open a conversation
…
No generated conversations yet — the built-in one above is always available.
Use your own AI interaction
Nothing here is limited to the samples. There are three ways to put a real conversation through the same measures, in increasing order of effort.
Open a section
one at a time
Ingestion topology
Spec §2, §6, §13.3, §15 — one image, two Cloud Run services, GCS authoritative
Live packet flow animated on each ingested turn
synchronous (returns in the 201)
asynchronous (Cloud Tasks, OIDC)
event fan-out (Pub/Sub → SSE)
Response contract
Layer 1 completes inside the request. Everything semantic is queued, so the client gets deterministic numbers immediately and semantic curves arrive over the event stream.
Cloud Run may terminate the container the moment a response is written, so no unawaited promise is allowed to carry Layer 2 work (§15).
Thread laboratory
Spec §18.1 — the interface calls the same API a third-party app would
Thread configuration
Participants
LCC reference model v2.1.0
Transcript 0 ingested
API inspector
Spec §13, §18.2 — request, response envelope, and the SSE stream
Request
Press Replay to ingest the first turn.
Response
—
Event stream text/event-stream
Endpoint surface
Asynchronous job board Cloud Tasks
| Task | Handler | Turn | State | Attempts |
|---|
Handlers are idempotent and keyed on threadId + turnId + algorithmVersion, so a retry after a 5xx cannot double-write a curve point.
Layer 1 — content-agnostic interaction analytics
Spec §8 — deterministic, runs for every valid thread, never judges correctness
Token production by participant derived
| Ratio | Formula | Value | Provenance |
|---|
Timing observed
| Turn | TTFT | Gen. dur. | TPOT | E2E |
|---|
Human adjacent-turn latency HL
Completion length vs. adjacent latency OLS on log(1+HL)
Linguistic measures per language adapter
| Measure | Value | Method |
|---|
Lexical density here is a stop-list approximation, not a POS tagger; it is reported with its method string and is not ICCC Communication Density.
Speech acts and structural derivatives
Spec §8.8–§8.9 — schema-constrained LLM classification over functional segments
Act distribution LLM inferred
Act → act transitions
Rows: preceding act. Columns: following act. Counts only, no valuation.
Structural measures
LCC — Learner's Characteristic Curves
Spec §10 — requires a tutorial context and a versioned reference model
—
Quadrant curves R and N are model-inferred
RN relevant + new
RO relevant + old
IN irrelevant + new
IO irrelevant + old
Soft membership per contribution RN+RO+IN+IO = 1
Current contribution
Expectation coverage non-duplicative
Misconception evidence
Instructional state descriptive
ICCC — Individual Conversation Characteristics Curves
Spec §11 — six measures per participant, Social Impact revisable
Six aligned curves
Dashed segments are provisional Social Impact: the forward window is not yet complete, so those points will be revised when later contributions arrive.
Contribution inspector
Social Impact revision ledger versioned
Per-participant snapshot at the playhead accessible table alternative
| Participant | Participation | Responsivity | Cohesion | Social impact | Newness | Comm. density | Status |
|---|
Group tutoring — two models, no merged score
Spec §12 — cross-model patterns are descriptive only
LCC × ICCC pattern map
Each mark is one learner contribution: horizontal = LCC RN, vertical = ICCC Social Impact. There is no arithmetic combining the two axes.
Observed patterns in this thread
Authorization laboratory
Spec §17, §18.6 — deny by default, scoped grants, immediate revocation
Grant active
GET /api/v1/threads/:threadId/turns …
—
Audit trail
Object layout
Spec §14 — opaque owner key, versioned analyses, no prefix scans on read paths
object layout 0 objects
Write semantics
| Idempotency key | threadId + turnId + clientRequestId |
| Concurrency | ifGenerationMatch |
| Replay of a stored turn | 200 + DUPLICATE_TURN |
| Analysis overwrite | never — new versions/{n}.json + pointer swap |
| Owner key | derived HMAC, never an email or JWT subject |
Object keys are built from validated opaque identifiers only. No component is ever derived from an email address, JWT subject, or client-supplied string.
Metric dictionary
Spec §23 — every metric carries its prohibited interpretation
| Construct | Layer | Operational definition | Range | Prohibited interpretation | Help |
|---|
Research-integrity notice · spec §27
These measures describe observable and computationally inferred interaction patterns. They do not, by themselves, establish learning, mastery, cognitive effort, engagement, trust, factual quality, intelligence, personality, or causal influence. Such interpretations require external measures and empirical validation.
Provenance: this interface contains no analytical implementation of its own. It creates a real thread, ingests each contribution through POST /api/v1/threads/:threadId/turns, and every number shown is read back from an API response. Scrubbing backwards replays the snapshot the API returned at that point rather than recomputing anything locally, which is what makes a provisional Social Impact value visibly become final. When the banner reports mock mode, speech acts are rule-based and embeddings come from a hashing vectoriser rather than a learned semantic model — the values are real API output, but the instruments behind them are stand-ins.