Kashi
Quarterly Structural Integrity Audit
A structural assessment of team-level interaction patterns over a 90-day window. This audit reports on observable conversational structure — not content, not individual evaluation, not psychological state.
01 / Executive summaryWhat this report tells you
Over the past 90 days, Engineering Team Alpha's conversational structure stayed within stable patterns on two indicators and showed drift on two others. The team's overall Structural Integrity Score for this window is 0.82 (range 0.0–1.5 where 1.0 is the cross-customer reference; lower is better; see Section 04 for limitations).
Quarter Score · 2026-Q3
0.82
What this means: the team's structural patterns suggest fewer indicators of dialogue drift than the reference population of similar JP knowledge-work SMBs. What this does NOT mean: nothing about employee productivity, individual performance, sentiment, or compliance with any regulation. See Section 08.
Two patterns warrant manager attention:
- D02 Response-flow disruption shows a localized pattern — one channel of questions consistently does not receive same-meeting responses (see Section 02).
- D04 Voice presence drift shows a 30-day decline in floor share for two members. The pattern is structural; the reason is not inferred.
Two recommended manager actions are in Section 05. Renewal decision basis is in Section 06.
02 / Findings · D01–D04What each detector measured
Kashi measures four structural patterns. Each detector returns a categorical state (Stable / Watch / Attention) plus the structural evidence that triggered it. None of these detectors evaluate, score, or classify individuals.
| Detector | What it measures | What it does NOT measure |
|---|
| D01 · Floor-share asymmetry |
Distribution of speaking time across the team's meeting series; identifies trend changes vs the team's own prior window |
Speaking content, opinion quality, individual performance, who is "talking too much" |
| D02 · Response-flow disruption |
Whether questions / proposals consistently receive same-meeting responses or accumulate as unaddressed; structural only |
Whether the question is "good," whether the answer is "right," whether anyone is at fault |
| D03 · Interruption directionality |
Statistical asymmetry in who interrupts whom across the meeting series; reports the pattern, not the count |
Whether any interruption was rude, hostile, or warranted; never a "harassment" determination |
| D04 · Voice presence drift |
Trend changes in a member's speaking participation over 30/90/180-day windows relative to their own prior baseline |
Whether the member is disengaged, planning to leave, struggling, or "at risk"; the reason behind the trend |
D01 · Floor-share asymmetry
D01 · 2026-Q3
STABLE
Speaking-time distribution remained within the team's typical envelope across the 13 meetings analyzed. No member's share moved more than ±4 percentage points from their own 90-day prior baseline.
Confidence: Medium-high. 13 meetings is above the minimum-data threshold (10) but below the high-confidence threshold (24). Re-evaluate at end of Q4.
D02 · Response-flow disruption
D02 · 2026-Q3
WATCH
One conversational channel within the team — questions from one identified role-cluster (de-identified as Cluster B in this report) — received same-meeting response in 31% of instances, vs the team-wide rate of 68%. The pattern is consistent across 9 of 13 meetings.
Structural evidence excerpt · de-identified
Cluster B question (timestamp t+18:42) → no same-meeting response within 8-minute window → topic returned later by different cluster.
Cluster B question (timestamp t+34:11) → acknowledged ("noted, let's come back to it") → not addressed within the meeting.
Cluster B question (timestamp t+56:03) → no response → resolved offline in 1:1 (per next-meeting reference).
Excerpts are structural records of timing + response, not transcripts. Speaker identity is bound to a role-cluster, not a person.
Confidence: Medium. The pattern is consistent across 9 of 13 meetings. Plausible non-disruption explanations exist (e.g., the cluster's questions were appropriately deferred to async channels). The detector cannot disambiguate.
D03 · Interruption directionality
D03 · 2026-Q3
STABLE
Interruption pattern across the meeting series is approximately symmetric. No statistically significant directional asymmetry was identified in the analyzed window.
Confidence: Medium. Interruption detection requires conversational pacing within ±0.5s thresholds. 4 of 13 meetings included multi-party overlap segments where attribution was ambiguous; those segments were excluded from this calculation.
D04 · Voice presence drift
D04 · 2026-Q3
WATCH
Two members (de-identified as Member A and Member B) show declining floor share over the 30-day rolling window. Member A: 12% → 7% across the window. Member B: 9% → 5%. Both declines are above the 3-percentage-point trigger threshold and persisted across 4 consecutive 1-week sub-windows.
Structural evidence · de-identified time-series
Member A: Week 1 — 12% · Week 5 — 10% · Week 9 — 8% · Week 13 — 7% (monotonic decline)
Member B: Week 1 — 9% · Week 5 — 8% · Week 9 — 6% · Week 13 — 5% (monotonic decline)
These are observed structural patterns. The detector does NOT infer intent, engagement, or any individual-level state.
Confidence: Medium-high for the pattern. Zero confidence on causal interpretation — the drift could reflect work-load shifts, role changes, meeting-format changes, vacation, personal preference, or many other reasons the detector cannot see. Do not infer "at-risk employee" from this signal.
03 / Evidence handling policyHow we cite without identifying
Every "evidence excerpt" cited in Section 02 follows three rules:
- Role-cluster, not person. Members are referenced as "Member A / Cluster B" within this report. The mapping to actual identities is held by the customer's authorized administrator only — not by Kashi staff and not embedded in this report.
- Structural metadata, not transcript. Timestamps, response presence, and floor-share percentages are reported. Spoken content is never quoted. Kashi does not retain transcripts past the analysis-window TTL.
- k-anonymity threshold of 5. No team-level metric is reported when the analysis sub-population is below 5 members. Where the team is below this threshold (e.g., 4 members), only the team-wide aggregate is reported; no member-resolved evidence is shown.
What you receive: a report that lets a manager see the team's structural shape and decide what to do.
What you do not receive: the ability to look up who said what.
04 / Structural Integrity ScoreWhat 0.82 means and what it does not
The Score is a bounded indicator aggregating the four detector outputs (D01–D04), each weighted by its confidence in the analyzed window. The reference scale is calibrated against a cross-customer benchmark population of similar JP knowledge-work SMBs (50–500 employees).
| Score range | Categorical label | Operating implication |
|---|
| 0.0 – 0.75 | Stable | No detector triggered Attention; no operational change recommended this window |
| 0.75 – 1.05 | Routine review | One or more Watch findings; manager review recommended within 14 business days |
| 1.05 – 1.30 | Active review | Multiple Watch findings or one Attention finding; manager review + documented action within 14 days |
| > 1.30 | Escalated review | Multiple Attention findings; review with HR sponsor recommended; never an automated HR or discipline trigger |
What the Score is NOT: not a measure of team performance, productivity, engagement, harassment, psychological safety, manager quality, or any individual's value or risk. It is a single aggregate number summarizing observable structural patterns over a 90-day window. It is designed to be one input among many for a manager's judgment — not a verdict.
Score versioning: the aggregation formula and reference population are versioned. This report uses Score v0.9 (synthetic-data calibration). Future versions may produce different numerical values for the same underlying structural patterns; categorical labels will remain stable.
05 / Recommended manager actionsTwo structural responses
The findings in Section 02 suggest two recommended manager actions over the next 14 business days. Both are framed as structural questions for the manager to investigate, not directives.
Action 1 — Response-flow review (D02)
Structural question: what work pattern explains Cluster B's questions being deferred more often than the team-wide rate? Possible reasons include async-by-design (the cluster's questions are appropriately deferred to written channels), agenda-driven (the cluster's topics fall later in the meeting and run out of time), or interpersonal (the questions are not surfacing as priority for the response-givers).
Manager's options: walk the next 2 meetings with the Cluster B participant to identify the operational pattern; introduce an "open questions" agenda item; surface the structural finding to the cluster directly and ask if their preferred resolution channel is being honored.
SLA-aligned action: document the chosen path within 14 business days. Kashi tracks whether a documented action was logged (not what action was chosen).
Action 2 — Voice presence review (D04)
Structural question: the floor-share decline for two members is observable; the cause is not. The manager has context Kashi does not (work assignments, recent role changes, meeting-format changes, personal communication preferences, vacation).
Manager's options: use the team's existing 1-on-1 cadence to surface the structural observation in a non-evaluative way ("I noticed your participation pattern has shifted — I want to make sure we're hearing from you; what would help?"); review meeting-format changes in the analyzed window; check role-load changes.
Important: the Kashi finding is structural-only. Do NOT use this finding as input to any individual performance evaluation, role-change decision, or HR action. See Section 08.
The 14-business-day Signal-to-Action SLA is structural too — Kashi measures whether a documented action was logged. The chosen action is the manager's call, in consultation with HR if needed. Kashi never prescribes the action.
06 / Quarter-over-quarter comparisonPlaceholder · populated from Q4 forward
Quarter-over-quarter comparison shows the directional change in the Structural Integrity Score and the categorical state of each detector. For your first quarterly audit, only the current quarter is populated; historical comparison begins with the second audit.
2026-Q3 (current)
0.82
first audit · baseline
2026-Q4
—
populated next audit
From Q4 onward, this section will display: Score delta vs prior quarter, detectors that changed categorical state, and a "what changed" summary referencing the structural pattern (never an individual). A multi-quarter trend chart is added after 4 audits.
07 / Data scope · confidence notesWhat was analyzed
| Dimension | This audit | Minimum / Recommended |
|---|
| Meetings analyzed in window | 13 | min 10 · recommended 24+ |
| Window length | 90 days | 90 days (standard) |
| Team members in scope | 12 | min 5 (k-anonymity) |
| Total turns analyzed | ~1,840 | — |
| Total speaking time | ~22 hours | — |
| Excluded segments | 4 multi-party overlap segments (D03 ambiguity) | standard exclusion |
| Confidence band | Medium-high | 10+ meetings = medium; 24+ = high |
Confidence caveat: 13 meetings is above the floor for analysis but below the recommended threshold of 24 for high-confidence pattern claims. Two-quarter trends become more interpretable from the next audit onward.
08 / What this audit does NOT measureThe boundary, in plain language
Explicit non-measures
- Individual performance or productivity. Kashi cannot and does not measure whether an employee is performing their job well. Speaking-time decline is structural — not a performance signal.
- Sentiment, emotion, mood, psychological state. Kashi has zero emotion-recognition or sentiment-analysis components. The detectors compute timing and direction; they do not infer feelings.
- "At-risk employee" or attrition prediction. Kashi makes no claim that the structural patterns predict who will quit. The patterns reflect what happened in meetings, not what will happen in HR records.
- Harassment, hostility, intimidation. The detectors never determine whether any interaction was inappropriate. Kashi reports the structural shape; investigation of any specific instance is HR + manager judgment using channels other than this audit.
- Manager quality. A team's Score reflects the team's structural patterns, not the manager's evaluation. Lower Scores do not imply "better manager."
- Compliance certification. This report is NOT a regulatory certification. It is a structural diagnostic. Compliance posture is a separate process involving qualified counsel.
- Cause-and-effect for any specific incident. If a team incident occurred, this audit does not establish what caused it.
- Anything that requires content reading. Kashi reads structural metadata (timing, who-spoke-when, response-presence). Spoken content is not retained past TTL and is not the basis for any reported finding.
If a customer or partner asks Kashi to measure something on this list: the answer is no — by design. The Kashi product is built such that these capabilities are physically excluded at the data layer, not just turned off by policy.