dos-enforce-tune

dos-enforce-tune — the loop where DOS tunes its OWN enforcement, from outcomes

DOS is a sound PDP with no feedback from the PEP. The kernel decides an intervention verdict, a host acts on it, the act is journaled — but nothing fed whether the act was right back into the policy that drove it. This loop closes that. It learns the enforcement thresholds from the journal's own ground truth: a deny the operator later overrode is a false-DENY (too aggressive); a deny that stood is a held catch. The loop tunes the policy to drive false-DENIES down while holding the catches — and keeps an edit only if the kernel, not the agent, measures that it helped.

This is [[dos-self-improve]] pointed at the enforcement policy, with one twist: the metric is not a generic count, it is the docs/143 net_task_delta of the policy over labelled cases the loop did not author.

A self-tuner's fatal failure mode is grading its own homework — relabelling its outcomes so its policy edit looks good. dos enforce-tune closes that hole the same way dos improve does: the metric is computed BY THE KERNEL from cases the loop did not author (a frozen corpus ∪ the live enforcement journal). The loop cannot keep a policy edit by claiming it is better. The only path to KEEP is to actually move net_task_delta.

What the kernel decides vs. what you do

Step	Who	What
Read the outcomes	dos enforce-outcomes	the live false-DENY / held-catch ledger — what to tune toward
Propose ONE policy-knob edit	YOU (a subagent)	the untrusted step — edit [intervention_policy] / [intervention] ranks / [improve] in a worktree
Measure	dos enforce-tune	score the candidate policy's net_task_delta over the corpus, on the worktree
Keep / revert / escalate	dos enforce-tune (rides dos improve)	the kernel's typed verdict over env-authored facts — NOT your opinion
Merge / discard / escalate	YOU (or the autonomous cadence)	carry out the kernel's verdict

Inputs

• --cases <corpus.jsonl> (required) — the labelled InterventionCase corpus the candidate policy is scored over (the docs/143 §13.2 ground truth). The reference corpus is the benchmark's intervention_cases.jsonl.
• the autonomous cadence (below) reads the dos enforce-tune EXIT CODE and merges on a KEEP. Run interactively WITHOUT auto-merge to inspect the verdict first (the safe default); let the cadence auto-merge once you trust it. Either way the verdict is the kernel's, and the runtime-logic rail still reverts a candidate that edited adjudication logic.
• --max-cycles <N> (default 5) — the backstop cap.
• --max-reverts <N> (default 3) — ESCALATE to a human after this many non-keeps in a row (the breaker).
• --lane <name> (optional) — the lane to take; a bare loop auto-picks a free one.

Step 0 — Read what enforcement is getting wrong

dos enforce-outcomes --workspace . --json

This folds the OP_ENFORCE journal into the false-DENY / held-catch ledger. A standing false_denies count is the signal: the policy is refusing edits the workspace genuinely makes. (The dos pulse heartbeat surfaces the same count when it crosses the threshold, so an always-on operator sees it without asking.) Note the targets being over-blocked — they tell you which way to tune (usually: loosen a rung or raise a confidence bar so a marginal mint earns WARN, not BLOCK).

Step 1 — Pre-flight: take a lane, record the GREEN baseline net_task_delta

dos doctor --workspace . --json
dos arbitrate --workspace . --lane <LANE>
RID=$(dos run-id mint dos-enforce-tune \
      | python -c "import sys,json;print(json.load(sys.stdin)['run_id'])")
python -m pytest -q                      # MUST be green to start
dos enforce-tune --cases <corpus> --suite-passed --truth-clean \
  --baseline-work 0 --json --workspace . # read measured_work = the BASELINE policy's score

Record the baseline measured_work B. You cannot measure improvement from a red baseline — if the suite is red, fix it first (an ordinary task, not a tuning cycle). RID is the loop's correlation run-id; pass it with --observe on every dos enforce-tune call so dos observe --loops can fold this always-on tuner's trajectory while it runs (docs/383) — otherwise a long-lived cadence is invisible between its commits.

Step 2 — Per cycle: propose ONE policy-knob edit in an ISOLATED worktree

The enforcement policy is the kernel's own config — editing it on the live tree is the SELF_MODIFY hazard ([[self-modification-hazard]]). Work in an isolated worktree:

git worktree add ../_et-candidate HEAD

Spawn ONE subagent with a tight brief, working in that worktree:

Improve DOS's enforcement policy to reduce false-DENIES without losing catches. Edit ONLY the policy KNOBS: the [intervention_policy] confidence-gating values (on_high_confidence / on_low_confidence / floor / ceiling), the [intervention] ladder ranks, or the [improve] thresholds — in dos.toml. Make the SMALLEST edit that should move net_task_delta. Do NOT touch enforcement LOGIC (self_modify.py, admission.py, arbiter.py, the ladder code) — that edit will be reverted regardless of its metric. Commit your change. Return what you changed and why. If you cannot find a real improvement, say so — do not invent one.

Step 3 — Measure + decide (the kernel decides)

Gather the env-authored witnesses on the worktree, then ask the kernel:

cd ../_et-candidate && python -m pytest -q ; SUITE=$?
CHANGED=$(git diff --name-only HEAD~1 HEAD)   # the candidate's diff — for the rail

dos enforce-tune --cases <corpus> \
  --policy-toml ../_et-candidate/dos.toml \
  --baseline-work "$B" \
  $( [ "$SUITE" -eq 0 ] && echo --suite-passed ) \
  --truth-clean \
  --changed-files $CHANGED \
  --max-reverts <N> \
  --observe --run-id "$RID" --subject "cycle-$CYCLE" --json --workspace .

The --observe --run-id "$RID" --subject "cycle-$CYCLE" line records the cycle to the verdict journal so dos observe --loops folds this loop's trajectory live (docs/383).

dos enforce-tune re-scores the CANDIDATE policy (loaded from its worktree dos.toml), applies the runtime-logic rail to --changed-files, and rides dos improve. The verdict IS the exit code:

• 0 KEEP — suite green, truth clean, candidate net_task_delta strictly beat B, no runtime-logic edit. Witnessed. Go to Step 4-KEEP.
• 3 REVERT — a regression (suite red / a runtime-logic edit / truth dirty) or a no-op (the metric did not beat B). Go to Step 4-REVERT.
• 4 ESCALATE — the breaker is open (--max-reverts non-keeps in a row). STOP.

Step 4 — Carry out the verdict

• KEEP — merge the worktree commit onto the lane, RAISE the baseline to the candidate's measured_work (the ratchet), reset the revert count. The autonomous cadence does this on exit 0; by hand: git merge --ff-only <sha> then B=<measured_work>.
• REVERT — discard the worktree candidate; the live policy is untouched. Bump the revert count. git worktree remove --force ../_et-candidate.
• ESCALATE — discard, file a dos decisions entry with the kernel's reason, STOP. The loop has run dry of witnessed improvements; a person decides what matters next.

Re-create a fresh worktree off the (possibly advanced) lane HEAD and return to Step 2 until --max-cycles or an ESCALATE.

Always-on: the autonomous cadence

To run continuously, wire a durable scheduler (a cron / Windows Task via a gitignored .dos/ runner — the dos pulse runner sibling pattern). The runner does ONE bounded cycle per firing and merges on a KEEP by reading the exit code:

# per firing, in an isolated worktree the runner created off the lane HEAD:
dos enforce-tune --cases <corpus> --policy-toml <wt>/dos.toml \
  --baseline-work "$B" --suite-passed --truth-clean \
  --changed-files $(git -C <wt> diff --name-only HEAD~1 HEAD) \
  --observe --run-id "$RID" --subject "cycle-$N" --workspace .
case $? in
  0) git merge --ff-only <candidate-sha> ;;   # KEEP — auto-merge + ratchet B
  4) dos decisions ... ; exit 0 ;;            # ESCALATE — surface to a human, stop
  *) : ;;                                      # REVERT — discard, leave the policy
esac

The cadence keeps ONE $RID across all its firings (persist it in the runner's .dos/ state), so every cycle folds into the SAME trajectory — an always-on tuner that would otherwise be invisible between commits becomes a live curve in dos observe --loops.

Each firing is bounded (one cycle; over many firings the breaker + cap still hold), so it always terminates — never an unbounded run. dos pulse keeps its observe leg, so the heartbeat surfaces the false-DENY rate the loop acts on.

Watch the tuner live: dos observe --loops (every loop) or dos observe --loops --run "$RID" (this tuner's full per-cycle curve) — the metric high-water, the breaker N/max distance-to-escalate, and the ACTIVE/STALLED/ESCALATED band, all folded from the kernel's own verdicts (docs/383).

Why the autonomous merge is safe

• The keep-bit is non-forgeable. dos enforce-tune KEEPs only on a kernel-measured strict net_task_delta gain with the suite green and truth clean. A no-op or regression auto-REVERTs — the loop cannot narrate its way to a merge.
• The breaker escalates to a human exactly when the loop runs dry — the human-in-the-loop seed, preserved even when fully autonomous.
• The runtime-logic rail reverts ANY candidate whose diff touched the kernel's adjudication logic (self_modify._DISPATCH_RUNTIME_FILES), regardless of its metric. The loop owns the policy KNOBS; the SELF_MODIFY guard still owns the LOGIC blast radius. A tuner that rewrote the arbiter to make a number go up is the self-modification hazard DOS refuses — refused here too, before merge.

What this skill deliberately does NOT do (no silent gap)

• No logic edits. The loop tunes thresholds/ranks, never the enforcement code. A logic edit is auto-reverted; the brief forbids it.
• No metric the loop can game. net_task_delta is computed by the kernel from cases the loop did not author (the frozen corpus ∪ the live journal). A metric the subagent computes about itself is back to grading its own homework — refuse it.
• No unbounded run. Every firing terminates into the cap or a human ESCALATE.

Anti-patterns

• ❌ Believing the subagent's "it's better" — re-measure via dos enforce-tune; only the kernel verdict decides KEEP.
• ❌ Editing enforcement LOGIC to move the metric — the runtime-logic rail reverts it.
• ❌ Tuning from a red baseline — fix the suite first; you cannot measure a gain you cannot see.
• ❌ Loosening the policy until false-DENIES hit zero — that tanks coverage (catches held), which net_task_delta penalizes; the kernel will REVERT an over-loosened candidate because the metric balances both.