organvm-iv-taxis/session-governance-audit

Session Governance Audit

Turn a raw session transcript into an accountable record: what was touched, what it cost, what shipped against what was asked, and — most importantly — what is missing and which kind of missing it is.

Why this exists

Sessions produce two failure modes that look identical from the outside: work that was attempted and incomplete (a Gap) and work that was never represented anywhere at all (a Vacuum). Closeout rituals catch the first; only a deliberate sweep catches the second. This skill is the deliberate sweep — the "hall monitor" pass that compiles evidence rather than trusting the session's own self-report.

Core doctrine

• N/A is a vacuum, never a resting state. Any metric, deliverable, or task that resolves to "N/A / unknown / not done" must become a named item in the output — never silently dropped.
• Evidence over self-report. Claims in the transcript ("pushed", "fixed", "done") are hypotheses; verify each against disk/remote before recording it as shipped.
• The audit is itself an artifact. The index must be durable (committed or filed), or the audit becomes its own vacuum.

Workflow

1. Establish the tasked set

Collect everything the session was supposed to do:

• explicit user prompts (each imperative sentence = one tasked atom)
• plan documents authored or referenced in-session
• IRF rows / registry items the session claimed
• handoff documents the session inherited

Number them T-01 … T-NN.

2. Compile the annotated bibliography (the shipped set)

Walk the transcript chronologically and record every artifact event:

| ID | Artifact | Event | Evidence | Verified? | |----|----------|-------|----------|-----------| | A-01 | path/to/file | created / modified / deleted / committed / pushed | commit SHA, tool-use ref, or diff excerpt | disk/remote check result |

Rules:

• One row per artifact per event class (a file edited then committed = two events, one row each is acceptable but collapse when noise outweighs signal).
• Verify each row: git log/git show for commits, test -f/content grep for files, remote ref comparison for pushes. Mark unverifiable claims UNVERIFIED, not shipped.
• Include non-file artifacts: PRs opened, issues filed, branches created, external messages sent.

3. Account for internal energy

Estimate what the session consumed:

• Tool uses: count by class (read/search vs write/execute vs network).
• Tokens: estimate from transcript size (chars ÷ 4) split into input/output where derivable.
• Wall-clock: first-to-last timestamp.
• Subagents/parallel work: count and attribute their energy separately.

Energy data contextualizes the tally: a session that burned 400K tokens to ship 1 of 7 tasked atoms reads very differently from one that shipped 6 of 7.

4. Tally shipped vs tasked

Cross-join the tasked set against the shipped set:

• SHIPPED — tasked atom with verified artifact evidence
• PARTIAL — evidence of attempt, incomplete result
• UNSHIPPED — no artifact evidence at all
• UNTASKED-SHIPPED — artifacts with no originating task (scope additions; flag, don't condemn)

5. Classify the missing

For every PARTIAL and UNSHIPPED atom, classify:

• Gap — the item is represented somewhere durable (IRF row, plan, TODO, open PR) but unfinished. Remediation: finish or re-route the existing representation.
• Vacuum — the item exists nowhere durable; if this audit didn't name it, it would vanish. Remediation: file it (IRF row, plan, or registry entry) as part of this audit, then record the filed ID.

A vacuum that this audit names but does not file remains a vacuum. The audit is not complete until every vacuum has a durable home.

6. Emit the Session Governance Index

Assemble the report:

# Session Governance Index — {session-id} ({date})
## 1. Identity        — session id(s), agent, scope, duration
## 2. Bibliography    — the verified artifact table (step 2)
## 3. Energy          — tool-use counts, token estimate, subagent attribution
## 4. Tally           — shipped / partial / unshipped / untasked-shipped counts + per-atom table
## 5. Gaps            — each with its existing durable representation
## 6. Vacuums         — each with the durable home filed during this audit
## 7. Verdict         — one paragraph: did the session do what it said it did?

Route the index to the session's governance home (e.g. docs/evaluation/ or the session archive), commit it, and reference it from the closeout record.

Anti-patterns

• Trusting the transcript's own summary. Sessions over-report completion; always re-verify.
• Counting prompts instead of imperatives. One dense prompt can pack ten tasked atoms; decompose before tallying.
• Filing the index locally only. An uncommitted audit is unfinished work, not a finished audit.
• Treating untasked-shipped work as free. Scope additions consumed energy that tasked atoms didn't get; surface the trade.

Worked micro-example

Tasked: "fix the CI matrix and push" (T-01, T-02). Transcript shows an edit to .github/workflows/ci.yml (A-01, verified via git show abc1234) and a push claim with no remote evidence (A-02, git branch -r --contains empty → UNVERIFIED). Tally: T-01 SHIPPED, T-02 UNSHIPPED. Classification: T-02 is a Gap if the push failure is recorded in a handoff doc; a Vacuum if nothing durable mentions it — in which case file IRF-OPS-NNN: push abc1234 to origin now and record the ID in §6.