plat 289

← Pulse platform issue index

PLAT-289 — Substack research paths were fixed, but learning notes retained contradictory write/success contracts

Coordination	Value
State	Workflow learning corrections applied; retained artifacts verified; file-transport regression passes
Date	2026-09-05
Related	PLAT-175, PLAT-221, PLAT-257; research group in the full remaining-report audit

Actual state

The ten historical reports concern source steps told to write db/research/ while their managed-agent filesystem grant only permits db/assets/ below db/. Some attempted a per-date SQL table as a workaround; ordinary mutation tools correctly refuse DDL. Broadening db/ would expose raw SQLite and is not the repair.

Before this pass, all four source descriptions and their validation path regexes already used the final db/assets/research/YYYY-MM-DD-<source>.json location. The parent and db/README.md agreed. No promotion into db/research/ is required.

Three live learning instructions still contradicted that corrected contract:

• The shared SKILL index called legacy db/research/ canonical and described a deprecated flat mirror filename workaround.
• Hacker News claimed an orchestrator would promote the asset into the old path.
• Reddit directed the step to populate success literals after a denied write, assuming downstream tooling would materialize the packet, and named retired record_run_concern.

Corrected only these three statements in the workflow's learning files. Reddit now fails honestly on write/readback failure; success literals require actual readback validation. No plan, schedule, permissions, database schema, or business packet was changed. The retired CLI radar and separate growth/drafts groups are outside this repair.

Retained run verification (read-only)

The four 2026-09-05 durable packets parse, contain the expected source, and have matching finding_count == len(findings):

Source	Findings
Reddit	15
X/Twitter	5
Hacker News	29
Web search	8

All four forward outputs' findings match their durable packets exactly; receipt paths/sources/counts agree. The parent research_summary.json lists those four paths and totals 57 findings. File SHA-256 values and the checked plan hash are preserved in each SQLite closure event. This verifies retained files, not the truth/freshness of every external news story, and does not claim we launched a fresh producing run.

Regression

TestMessageSequenceResearchPacketsPersistInsideGrantedAssets constructs the real message-sequence guard, configures the managed DB session, and writes/reads JSON through the guarded workspace client and production HTTP document handlers in a temporary docs root. Each source packet is genuinely persisted and parsed. Writes to legacy db/research/, db/db.sqlite, and its WAL sidecar are rejected without creating those files.

Focused message-sequence, DB-session, spill-folder and path tests pass. The new test exercises the file API, not Linux Landlock or macOS shell execution. No platform permission implementation was changed.

Exact internal closures

PUL-AB711AC1, PUL-01F790F3, PUL-5903292F, PUL-137C5EDF, PUL-22656D6C, PUL-F7B23729, PUL-9CE79401, PUL-32C03E39, PUL-B8FC2F51, PUL-224A1885.

Resolved these superseded-path reports in SQLite and read-back verified, with prior concern/detail records retained in platform_tracking_resolved event metadata. This is not a claim that generic JSON regex validation independently checks referenced file existence: it does not. Current file presence was checked directly and the workflow's explicit readback contract remains necessary. Historical missing packets are not fabricated or backfilled. Learning/test changes remain uncommitted.

2026-09-09 local Pulse audit follow-up

The local Substack learning correction remains applied. Related contradictory guidance in other workflows belongs to builder/change-contract prevention under PLAT-257, not reopening this already-corrected Substack issue without a fresh reproduction.