docs(specs): pseudo-anonymization-readlayer-v0 prebuild (#177) #338

Merged

pdurlej merged 1 commit from claude/pseudo-anonymization-readlayer-v0-prebuild into main

2026-05-23 10:31:04 +02:00

claude commented

2026-05-17 18:22:40 +02:00

Collaborator

Summary

Pre-implementation Spec Kit for local pseudo-anonymization filter between cloud cousins (Hermes, claude-cloud, codex-cloud) and operator's vault content (Tier 0). Uses openai/privacy-filter (Apache 2.0, 1.5B sparse-MoE, 8 PII categories). Defense-in-depth on top of physical separation (vault-tier-3-architecture-v0).

BATCH G prebuild 4/4. Final prebuild in BATCH G.

What's in

docs/specs/pseudo-anonymization-readlayer-v0/ — 4-file Spec Kit
prompts/codex-pseudo-anonymization-readlayer.md — execution prompt

Slices

A (Full, FIRST): ADR committing to openai/privacy-filter + arch
B (Full, canary 3+3): privacy_filter.py with redact() + CI perf gate (<500ms median)
C (Full, canary 3+3): vault_read_anonymized.py gatekeeper
D (Full, canary 3+3): Hermes integration as first cloud cousin through filter
E (Lite): platformctl privacy-filter test CLI for operator verification

Constitution principles (P1-P8)

Defense-in-depth NOT primary / default-deny on filter fail / <500ms perf budget / audit log per access / complementary to physical separation / static 8-category policy / ADR-0018 conformance / local-first deployment.

Architecture

operator writes Iskra-i-Piotr/* (plaintext)
    │
    ▼
agent issues read
    │
    ▼
vault_read_anonymized.py gatekeeper
  ├── path-policy (Tier 0 OK, Tier 1/2 DENY)
  ├── openai/privacy-filter (M1 local)
  ├── mask spans → [REDACTED:PERSON] etc.
  └── audit log
    │
    ▼
agent receives masked content

Tier: Trivial per ADR-0007.

Refs #177 #178 #180 #237 #64

## Summary Pre-implementation Spec Kit for local pseudo-anonymization filter between cloud cousins (Hermes, claude-cloud, codex-cloud) and operator's vault content (Tier 0). Uses `openai/privacy-filter` (Apache 2.0, 1.5B sparse-MoE, 8 PII categories). Defense-in-depth on top of physical separation (`vault-tier-3-architecture-v0`). BATCH G prebuild 4/4. Final prebuild in BATCH G. ## What's in - `docs/specs/pseudo-anonymization-readlayer-v0/` — 4-file Spec Kit - `prompts/codex-pseudo-anonymization-readlayer.md` — execution prompt ## Slices - **A** (Full, FIRST): ADR committing to openai/privacy-filter + arch - **B** (Full, canary 3+3): `privacy_filter.py` with `redact()` + CI perf gate (<500ms median) - **C** (Full, canary 3+3): `vault_read_anonymized.py` gatekeeper - **D** (Full, canary 3+3): Hermes integration as first cloud cousin through filter - **E** (Lite): `platformctl privacy-filter test` CLI for operator verification ## Constitution principles (P1-P8) Defense-in-depth NOT primary / default-deny on filter fail / <500ms perf budget / audit log per access / complementary to physical separation / static 8-category policy / ADR-0018 conformance / local-first deployment. ## Architecture ``` operator writes Iskra-i-Piotr/* (plaintext) │ ▼ agent issues read │ ▼ vault_read_anonymized.py gatekeeper ├── path-policy (Tier 0 OK, Tier 1/2 DENY) ├── openai/privacy-filter (M1 local) ├── mask spans → [REDACTED:PERSON] etc. └── audit log │ ▼ agent receives masked content ``` Tier: Trivial per ADR-0007. Refs #177 #178 #180 #237 #64

claude added 1 commit

2026-05-17 18:22:40 +02:00

docs(specs): pseudo-anonymization-readlayer-v0 prebuild (#177 )

base-is-main / guard (pull_request) Successful in 1s

Details

canary-required / collect-diff (pull_request) Successful in 3s

Details

patchwarden-pr-sanity / collect-diff (pull_request) Successful in 3s

Details

canary-required / canary (pull_request) Successful in 12s

Details

patchwarden-pr-sanity / sanity (pull_request) Successful in 21s

Details

cd82e620d3

Pre-implementation Spec Kit for adding local pseudo-anonymization
filter between cloud cousin agents and operator's vault content
(Tier 0). Defense-in-depth on top of physical separation established
by vault-tier-3-architecture-v0 (#178+#180).

Target model: openai/privacy-filter (Apache 2.0, 1.5B sparse-MoE,
128k context, 8 PII categories: account/address/email/person/phone/
url/date/secret). Local M1 CPU/MPS deployment default.

5 slices:
- A (Full, FIRST): ADR draft committing to openai/privacy-filter + arch
- B (Full, canary 3+3): privacy_filter.py with redact() API + perf gate
- C (Full, canary 3+3): vault_read_anonymized.py gatekeeper wrapper
- D (Full, canary 3+3): Hermes integration as first cloud cousin through filter
- E (Lite, parallel): platformctl privacy-filter test CLI for operator verification

Constitution (P1-P8): defense-in-depth NOT primary, default-deny on
fail, <500ms perf budget, audit log per access, complementary to
physical separation, static 8-category policy, ADR-0018 conformance,
local-first deployment.

Sibling: vault-tier-3-architecture-v0 (#178+#180) — complementary, may
share capabilities.yaml denylist when both land.

Tier: Trivial per ADR-0007 (Spec Kit + prompt, docs-only).

Refs #177 #178 #180 #237 #64 #142