What was decided, and what went wrong

Second instance of 20260828-002, different path. run-phase-11.sh redirected TELEMETRY_STATE but not VAULT_LOCAL; voice-gate-guard.sh resolves the Bridge queue from vault.md independently, so two TEST denials landed in the LIVE pending queue reading as real asks waiting on a human, one of them an ad spend.

Isolating a local path does not isolate a remote. Any write path that resolves a destination from config must honour a local override on BOTH halves, or refuse to use the configured destination at all. Redirecting the local path means not the real target.

A test set VAULT_LOCAL to a temp directory believing that isolated it. vault-write.sh resolves the REMOTE from vault.md independently of VAULT_LOCAL, so it git-init'd the temp dir, adopted the real remote, and pushed two fixtures carrying FABRICATED Finanshels runway figures into the company vault, addressed to the board.

Fourteen legacy mode-local ledgers are superseded by this single append-only canon. Three carried content and were migrated; eleven were absent.

Price a spend gate before asking for approval. State the estimate in the same breath as the ask. An unstated ~$0.50 cost carried the weight of a much larger fear.

Verify the user's premise against data, then verify your own measurement. A naive median contradicted the user and was itself contaminated by 6 canned deflection strings; only the raw message bodies settled it. Both the report and the metric can be partly wrong.

A falsifier must name the CLASS of finding that would overturn a call, never enumerate file locations. A path-enumerating falsifier let a real defect through on a technicality.

Rank plumbing above craft when predicting a root cause. Predict the boring mechanism first.

All three commander predictions anchored on the most interesting mechanism (citation gate, reply suppression, request-construction defect) while every real cause was plumbing: a routing id, a discarded error string. Recorded outcome across the mission: 3 hits, 2 misses, 2 partials, 1 unscorable.

The user's working tree is off-limits entirely — never stash it, never clean it, never assume uncommitted changes belong to the run. Brief this to every builder, not just for specific files.

A builder stashed the user's unrelated uncommitted work in order to get isolation.

Every parallel builder receives an absolute unique path AND an explicit file-ownership contract at dispatch. Never a conventional name chosen by the builder.

Two parallel builders independently chose the same conventional worktree directory name; the second reverted the first's work as contamination. The commander then misdiagnosed it as fabrication and said so to the user before reading the other builder's report.

HalaTax WA agent reached launch-ready conditional: eval 28/28 green at 7bddf69, promotion ask filed, NOT promoted. First mission with no degraded phase.

Verify artifact paths on disk at checkpoint time AND again at resume before reporting any phase complete.

Two phases were marked complete against files that were not on disk. The mission record said done; the filesystem said absent.

Detect the installed roster before planning, not before dispatching. A plan with a majority of degraded phases is reported as an install recommendation first and a mission plan second.

Mission planned 6 phases against a roster that was not installed; 4 of 6 phases ran degraded and the mission outcome was not met.

Any stdio server must be proven by a full handshake against clean, dirty, invalid-path and unknown-tool cases before it is called delivered. Review does not substitute for execution.

An MCP server dropped its final response because rl.on('close') exited before pending async tool calls flushed. Found only by real execution, not by review.

Forged and delivered a git-status MCP server; verified by full stdio handshake against clean repo, dirty repo, non-repo path and unknown tool name.

Seven material facts were unknown at decision time: runway/burn, 6-month MRR curve and ICP share, e-invoicing mandate date for the SMB tier, account count and top-3 revenue concentration, existing capital and rights, headcount split, founder-led share of pipeline.

Bootstrap Finanshels through the filing season and open the round Q1 2027 rather than Q4 2026.