Integration log · v1 → v2
What the research library changed 可視
Kashi began as a working prototype. Then two research libraries
landed — 24 business PDFs + 18 technical-dev memos, 42 documents total.
Below is an honest table of what changed, why, and which memo drove it.
2026-04-21 · Companion to /business.html · Source docs available to invited partners at the partner review page (password required)
42
total research documents integrated
11
risks now in register (was 6)
6
success metric layers (was 4)
01Positioning & language
strikethrough = what we used to say ·
regular text = what we say now
| Area | Before | After | Driver memo |
|---|
| Hero positioning |
"Governance infrastructure for workplace power dynamics" |
"Privacy-bounded meeting governance that surfaces repeated interaction asymmetries under uncertainty" |
category-entry · measurement-science |
| Thesis line |
"The pattern is the harm" |
"The pattern may constitute evidence consistent with uneven conversational treatment over time" |
measurement-science · legal-procedural-fairness |
| Detection claim |
"Detect harmful team dynamics earlier" |
"Surface repeated interaction asymmetries earlier, as contestable structural signals for human review" |
measurement-science · product-wedge |
| Buyer framing |
"The CEO's instrument for seeing the bill before it arrives" (front-door) |
"A bounded governance instrument" (front-door) · "CEO's instrument" demoted to closed-sponsor context only |
category-entry · product-wedge · trust · founder-narrative |
| Anti-surveillance line |
"Not an employee-monitoring tool" (overclaim — we DO create person-linked telemetry) |
"Not a general-purpose employee-surveillance product. A tightly restricted governance system with structural metadata under explicit anti-misuse limits" |
labor-relations · trust |
| Detector purity claim |
"All deterministic, metadata-only, no content reading" (contradicted by embedding-based detectors) |
Three lanes: 3 structural-only (default employer-facing) · 4 hybrid text-informed (tenant opt-in) · refused. Machine-readable at src/lib/pipeline/detector-registry.ts |
detector-boundary · model-boundary · measurement-science · legal-max-hardening |
02Market & ROI
| Area | Before | After | Driver |
|---|
| Hero cost metric |
"¥7.9M per leave case × N cases = big savings" |
Hidden-loss portfolio (presenteeism + regrettable attrition + formal-escalation premium + leave tail). ¥7.3T of Japan's ¥7.6T loss is presenteeism, not absence |
rebuilt-ROI · research-synthesis |
| ROI math |
8× / 3.3× / 1.3× by company size (overstated; assumed full leave-cost capture) |
A × C × I × E discount framework: Strict 0.54% = ¥0.69M · Mid 1.75% = ¥2.22M · Upper 4.80% = ¥6.10M captured per 30-case year |
rebuilt-ROI |
| Pricing defense |
Enterprise ¥10M base justified by leave-cost savings |
Enterprise base = governance-service fee (consultation + CAIQ pack + SOC 2 + CSM + compliance artifacts); leave-cost capture alone requires 94.4% which is indefensible |
rebuilt-ROI · procurement-security |
| SAM sizing |
"JP mid-market, ~45k companies" |
Same ~45k SAM but segmented by knowledge-work filter, blended ACV ¥4.2M, revenue ceiling ~¥190B/yr |
research-synthesis · funding-pathways |
03Go-to-market & expansion
| Area | Before | After | Driver |
|---|
| Expansion order |
EU-first (DE → NL → UK), US second |
Japan-first doctrine · Singapore second (PSG/EDG grants) · NL/UK year 4 · DE only after works-council playbook matures · US year 5+ after SOC 2 |
cross-cultural · funding-pathways · labor-politics |
| Buyer model |
One buyer: "the CEO" |
Three buyers with distinct rhetorical registers: Economic sponsor (CEO/CFO) · Operational gatekeeper (HR/Legal/Compliance) · Trust counterparty (worker rep / union / works council) |
founder-narrative · labor-politics · trust |
| Sales framing |
"Intro to the CEO at a 50-500-person JP company" |
"Requires executive sponsor + legal review + worker-representation alignment before pilot launch. All three, or none." |
labor-relations · labor-politics · rollout |
| Channel stack |
5 channels (CEO outbound, labor-law, academic, JP VC, content) |
8 channels (+ worker-rep/union engagement, + occupational-health channel, + ombuds/neutral-review) with priority weighting |
labor-politics · rollout · ally-path |
| Financing |
Pure equity (pre-seed ¥50M → seed ¥500M → Series A ¥3B) |
Equity + non-dilutive track in parallel: JP AI Adoption Subsidy 2026 (P1), Workstyle Reform Subsidy (P1), JOHAS routes, Singapore PSG/EDG. Customer-side subsidy is the enterprise pricing unlock |
funding-pathways |
04Risk register
Expanded from 6 risks to 11. The 5 new risks are not edge cases — they are
category-killers that v1 silently dismissed.
| Risk # | Name | What kills us if we ignore it | Driver |
|---|
| R1-R3 | Pilot lawsuit · MS/Zoom copies · Works-council consent | Present in v1. Mitigations refined. | all |
| R4 · NEW | Adversarial adaptation / metric gaming | Managers reroute dominance into channels Kashi doesn't see (agenda control, 1:1s, async, hierarchical laundering). Metric improves, harm doesn't. | adversarial · adversarial-for-devs |
| R5 · NEW | Institutional capture / selective deployment | Org buys Kashi, carves leadership out of scope, surveils ICs only. "Kashi surveils workers, protects leaders" becomes the public story. | anti-capture |
| R6 · NEW | Metadata leakage / inference reconstruction | Employer infers worker opened pattern page / created vault / marked confounds — directly actionable retaliation risk under MHLW guidelines. No content needed. | retaliation · privacy-retention |
| R7 · NEW | False-negative laundering | Clean dashboard cited as exoneration ("we ran Kashi, no signal, therefore no harm"). EEOC base rate: 3 of 4 targets never speak up. | false-negative |
| R8 · UPGRADED | ASR + diarization input contamination | 46.8% tcpWER single-channel, 54.9% overlap-heavy, 2× racial WER disparity. Demographic bias could become the scandal. | scientific-attack-surface · speaker-identity |
| R9-R10 | Regulator walkback · Scandal | Present in v1. Mitigations refined. | all |
| R11 · NEW | Goodhart within Manager Mirror | Even private self-view is gameable. Keigo/interruption improve on the measurable surface while dominance expresses in untouched channels. | measurement-science |
05Architecture & moat
| Area | Before | After | Driver |
|---|
| Moat layers |
4 layers: product (refusals) · process (consultation) · data (baselines) · regulatory (compliance) |
5 layers (+ Layer 5: anti-inference architecture — retaliation defense via telemetry partitioning, min-group-size suppression, batching/redaction) · Layer 1 upgraded to procedural-fairness spine (contestability state machine, bounded-context rule, retention-under-challenge) |
retaliation · legal-max-hardening · legal-procedural-fairness |
| Detector output |
Scalar intensity enum (calm/watch/concern) |
ConfidenceBundle with 5 sub-components + reason codes + abstention state + evidence grade. Types declared in src/lib/types.ts |
confidence-abstention · measurement-science |
| Detector taxonomy |
All 7 detectors treated as equivalent "structural" |
Three lanes declared at src/lib/pipeline/detector-registry.ts. Employer-facing default = structural-only. Tenant feature flag required for hybrid lane. |
detector-boundary · model-boundary |
| Meeting-type handling |
Universal baseline, cross-type pooling |
MeetingType + scoringMode as first-class fields. MVP scoreable = weekly_sync, standup, design_critique. Rest = observation-only. Cross-type pooling blocked. |
meeting-type-normalization |
| Governance "will not do" |
10 items |
12 items (+ no decryption of evidence server-side, + no company-wide health score) + 3-lane accountability model (self-correction / governed remediation / formal review) + anti-laundering rule + responsibility-boundary CAN/CANNOT table |
manager-adoption · labor-relations · measurement-science |
06Measurement & success
| Area | Before | After | Driver |
|---|
| Success model |
4 layers: adoption · behavior · responsiveness · strategic |
6 layers (+ remediation quality — was the handling actually fair? · + human recovery — did the affected person recover?) · resists Goodhart by requiring worker-side fairness rating |
remediation-outcomes |
| North-star metric |
Detection pass rate (3/3 on seed, 0 FP on control) |
Paying JP companies with signed 就業規則 amendment AND completed Lane-B remediation cycle with worker-rated fairness ≥ 4/5 |
remediation-outcomes · labor-relations |
| Evidence quality |
Binary: signal detected / not detected |
4-state dashboard: signal detected / no qualifying signal in observed data / insufficient observation / out-of-scope. Absence of signal ≠ absence of problem. |
false-negative |
07What was REMOVED
Worth calling out explicitly. These five items were in v1 as strengths;
the research made them liabilities:
- "The pattern is the harm" — philosophically satisfying but legally indefensible. Replaced with "evidence consistent with uneven conversational treatment."
- "CEO's instrument for seeing the bill" as top-of-document identity — reads as surveillance-for-executives. Demoted to closed-sponsor context.
- "Metadata only, no content" blanket claim — falsified by our own embedding-based detectors. Replaced with explicit lane taxonomy.
- EU-first expansion — DE requires works-council co-determination on monitoring tech; we don't have the playbook yet. Japan is actually the easier first market.
- Single-CEO buyer — reads as top-down imposition to worker reps, who will block deployment. Three-buyer model required.
08What the research did NOT change
Worth also naming. Several architectural choices held up across all
42 documents and required no revision:
- Structural-only detector core — 3 detectors (intrusive-interruption, chilling-delta, Gini) remain the default employer-facing lane. Endorsed by every relevant memo.
- Per-speaker 90-day baseline — defended explicitly by baseline-calibration + measurement-science memos.
- Refusal of company-wide health bar — multiple memos independently endorsed the refusal.
- Role-based visibility with audit trail — endorsed by trust + legal-procedural-fairness.
- Victim-owned E2E evidence vault — endorsed by retaliation + legal-max-hardening as one of the strongest architectural defenses.
- "Review-worthy event" language (not "harassment detected") — endorsed throughout.
- EU AI Act Art. 5 no-emotion-inference alignment — endorsed as architecturally required.
The one-sentence summary: Kashi's product architecture
was mostly right from v1. Its language, ROI math, buyer model, and
risk register were not. The research didn't require a rewrite of
the product; it required a rewrite of how the product presents itself
and defends its claims.
09v2 → v3 delta (2026-04-21 post-GitHub-push)
Small but important shipped-delta after the wave-3 docs pass. This
is the bridge commit: UX commitment from the Ideas_wave3 research
lands in the live product, not just in types.
| Area | v2 (declared) | v3 (shipped) | Driver |
|---|
| ConfidenceBundle emission |
Types declared in types.ts; not yet emitted. Detectors still return legacy scalar outputs. |
Backward-compat wrapper at confidence-bundle.ts. Every ManagerMirror + PatternSummary now carries overallGrade + overallAbstention + overallReasonCodes[]. UI shows grade badge with hover rationale on /demo/mirror + /demo/ceo. |
confidence-abstention · measurement-science |
| Weakest-link grade logic |
No aggregation logic specified |
Headline grade = MIN of (inputQuality, exposureSupport, contextSupport, detectorConfidence, aggregationSupport). "A signal is only as strong as its weakest component." |
confidence-abstention |
| Tech-doc §7.5 |
§7 listed 23 improvements, ranked but not sequenced |
New §7.5 "Next 3 builds" critical review — picks Japan data residency + meeting_type gating + RLS CI tests as the first post-funding Sprint 1. Also flags 3 items the original §7 missed (webhook ingest, feature-flag admin UI, detector cost telemetry). |
internal critical review |
| Code shipped |
0 lines of ConfidenceBundle consumption |
~250 lines: wrapper + type extensions + 2 seed updates + 2 UI components + 3 doc updates |
this pass |