guard-rsi-score

name

guard-rsi-score

description

One repeatable pass that keeps the RSI loop for `fak guard` actually CLOSING on our own usage - the real, hash-chained decision journal `fak guard` writes - instead of a plan-mode scaffold that never closes. Runs the native guard-RSI scorecard (`fak guard-rsi-scorecard`) over the two guard RSI loops (the hardware-gated LATENCY loop tools/guard_hop_rsi.py and the hardware-free VERDICT loop `fak guard-verdict-rsi`), scores maturity (can it honestly close?) + realized value (does it run on our usage?), turns each HARD defect into a required affordance to ADD (a journal-grounded loop, a deterministic verdict-quality metric, a non-forgeable keep-bit, real adjudicated rows, control-pane membership, a paired honesty test), retires guard_rsi_debt worst-first, re-measures to PROVE the debt dropped, and commits only the guard-RSI lane by explicit path. The product-loop counterpart of rsi-maturity (which scores the generic internal/rsiloop engine). Use after a change to either guard RSI loop, after a guarded session adds journal rows, or on a /loop cadence to keep the guard learning from our workflow.

guard-rsi-score - the RSI loop for fak guard, scored against reality

An instance of the generic scorecard doctrine. Read that for the five laws and the RSI loop; this file is the per-surface specialization.

Why this scorecard exists

rsi_maturity_scorecard.py scores the generic RSI ENGINE (internal/rsiloop) - the demo self-improver. Nothing scored the loop attached to the PRODUCT a user actually runs: fak guard. That loop has a real signal - the default-on, hash-chained decision journal fak guard writes - and a real failure mode: the latency loop (tools/guard_hop_rsi.py, #733) is honestly hardware-gated (#734) and so runs in plan-mode forever, never closing on a normal box.

The fix is not to fake the latency number. It is the SIBLING loop the journal CAN close: the verdict-pattern loop (fak guard-verdict-rsi), which reads the real journal, scores the verdict distribution's honesty, and keeps a refinement only on a strict gain + a witness it did not author. See docs/fak/guard-verdict-rsi-loop.md.

The law, in one line: the guard's RSI loop must close on a re-measured number from OUR OWN usage journal + an external witness - never on a fabricated latency baseline, and never on an empty journal.

The KPIs (what it grades)

Two axes (the industry-scorecard shape: structure vs realized value). Each HARD failure is one unit of guard_rsi_debt.

MATURITY - can the loop honestly close?

• verdict_loop_present (HARD) - a journal-grounded loop exists (closes without hardware).
• deterministic_metric (HARD) - verdict-quality is a pure function of the journal bytes (no clock, no RNG), so a KEEP can't be a fluke.
• nonforgeable_keepbit (HARD) - the honesty gate rejects a keep lacking rows / a strict delta / a green witness.
• empty_journal_honesty (HARD) - refuses a keep on 0 rows; the row count IS the gate.
• latency_loop_honest (SOFT) - the latency loop discloses its hardware gate, not silently broken.

REALIZED - does it run on our own usage?

• loop_reads_real_journal (HARD) - a loop READS the real journal (not a dangling telemetry string no code consumes).
• registered_in_control_pane (HARD) - the scorecard is in SCORECARDS AND the baseline is pinned, so the ratchet folds + gates this debt.
• kept_iteration_on_real_rows (HARD) - real adjudicated rows exist, so the loop can bank a kept iteration on our usage.
• paired_honesty_test (HARD) - a test proves KEEP-on-gain, REVERT-on-no-gain, and the empty-journal refusal.
• documented (SOFT) - the real-usage loop has a doc + this skill.

The pass (the shared loop)

1. Run it - fak guard-rsi-scorecard (work-list), --json (payload), --compare baseline.json (prove the drop and the Nx verdict).
2. Retire guard_rsi_debt worst-first - fix the heaviest HARD KPI by ADDING the real thing: the journal-grounded loop, the deterministic metric, the control-pane row, the paired test. If kept_iteration_on_real_rows is red, the journal has no rows yet - run a guarded session (or seed via fak serve --policy ... + /v1/fak/adjudicate, then fak audit verify) so REAL verdicts land. Never fabricate rows.
3. Route what the session found (close the loop) - run fak guard-verdict-rsi route so the worst bucket of THIS run's journal becomes a pickable finding, not a printed-and-dropped line. An honesty-hole (blank-reason-on-deny / unknown-verdict) routes a P1 queue row AND a deduped gh issue; a recurring denial reason (>= --threshold) routes a P2 queue row only. It is idempotent (re-running the same finding is a noop/damped, never a duplicate) and fail-open (a sink failure never breaks the pass). Use --no-issues on a host without gh. This is the rung that makes a session actually feed the backlog; the scheduled fak garden tick runs it --no-issues for you.
4. Weigh the SOFT signals, then stop - add the doc/skill if missing; keep the latency loop's hardware-gate disclosure honest. Don't chase soft signals to zero.
5. Re-measure + prove - --compare prints the debt delta; the scorecard reads A (debt 0) on the disciplined tree once registered + pinned.
6. Commit only the guard-RSI lane, by explicit path - the scorecard tool/test, the verdict loop tool/test, the control-pane row + pinned baseline, the doc + this skill. Never git add -A. End the subject with (fak guard).

The anti-gaming rule (specific to this surface)

Close the loop by making it learn from real verdicts, never by lowering the bar. A red kept_iteration_on_real_rows is fixed by producing REAL rows (a guarded session), not by a fixture file checked into the journal path. A red deterministic_metric is fixed by removing the clock/RNG, not by deleting the check. A KEEP requires a strict verdict-quality gain on real bytes AND a green witness; weakening check_iteration to pass an unwitnessed or empty-journal keep is the one move that turns the loop into theater - exactly the plan-mode-forever failure this scorecard exists to catch.

ASCII gotcha

Keep the scorecard + loop source ASCII-only (no em-dashes / smart quotes). The repo's provenance pre-commit hook crashes on non-cp1252 UTF-8 before its own escape fires. Use - for a dash and straight quotes. (The .md docs may use UTF-8.)