TPI-013 eval-first monkey model goal-progress audit gate

notes/pr12_update_summary.md

@@ -0,0 +1,23 @@
+# PR #12 Update Summary
+
+## Original goal
+
+Reach one real unchanged TPI-004 baseline/candidate evidence pair.
+
+## Progress since TPI-001
+
+- eval-first mechanism fixed
+- execution environment path fixed
+- attach / provider blockers isolated
+- remote reflection recovered
+- execution-ready packet fixed
+- provider-facing request and dispatch surfaces fixed
+
+## Current bottleneck
+
+One returned filled provider block is still missing.
+
+## Course-correction decision
+
+Continue only if the next loop directly reduces distance to returned filled
+block or real attach.

notes/tpi_013_course_correction_gate.md

@@ -0,0 +1,27 @@
+# TPI-013 Course Correction Gate
+
+## Continue current line only if
+
+- the current bottleneck is still concrete and singular
+- the latest loop materially improves execution readiness
+- the next loop reduces distance to real attach or runtime evidence
+
+## Require correction if
+
+- the same absent-state is repeated without improving the handoff-to-execution boundary
+- the loop adds documentation without reducing execution distance
+- the loop reopens already-closed debates
+
+## Current decision
+
+Continue, but require the next loop to either obtain a returned filled block or reduce the distance to that returned block in one direct step.
+
+## Course-correction label
+
+`continue_sharpening`
+
+## Exact next action
+
+Send the fixed TPI-012 outbound dispatch packet if not yet sent, or obtain one
+returned filled provider block and check it against the TPI-012 transition rule
+before any attach begins.

notes/tpi_013_decision.md

@@ -0,0 +1,29 @@
+# TPI-013 Decision
+
+## Original goal
+
+Resume the unchanged TPI-004 evidence pass and obtain one real
+baseline/candidate evidence pair in the selected provider environment.
+
+## Current bottleneck
+
+The only blocker is the absence of one returned filled provider block.
+
+## Divergence check
+
+- no drift into Git auth recovery
+- no drift into environment reselection
+- no drift into tokenizer, training recipe, or model redesign
+- soft-drift risk exists if another loop adds documentation without reducing
+  distance to returned filled block or real attach
+
+## Course-correction decision
+
+`continue_sharpening`, but only under a stricter audit rule: the next loop must
+either obtain one returned filled provider block or reduce the distance to that
+reply in one direct step.
+
+## Exact next action
+
+Send the fixed outbound dispatch packet or obtain one returned filled provider
+block and run the TPI-012 transition check before any attach begins.

notes/tpi_013_divergence_check.md

@@ -0,0 +1,23 @@
+# TPI-013 Divergence Check
+
+## Current lead bottleneck
+
+The only lead blocker remains the absence of one returned filled provider block.
+
+## Drift check
+
+The loop family did not drift into model redesign, tokenizer change, or environment reselection.
+
+## Risk of soft drift
+
+There is a workflow-level risk of polishing absent-state artifacts without obtaining the one concrete returned block needed for execution resume.
+
+## Audit reading
+
+The loop sequence has improved clarity and execution readiness, but has not yet crossed the boundary into real attach or runtime evidence.
+
+## Conclusion
+
+There is no hard scope drift into closed debates, but there is a real risk of
+soft drift if another loop fails to reduce the distance to a returned filled
+provider block.

notes/tpi_013_goal_progress_audit.md

@@ -0,0 +1,26 @@
+# TPI-013 Goal Progress Audit
+
+## Original goal
+
+Resume the unchanged TPI-004 evidence pass in the selected provider environment and obtain one real baseline/candidate evidence pair.
+
+## Progress since TPI-001
+
+- eval-first comparison contract was fixed
+- runnable environment path was selected
+- attachability and provider metadata blockers were isolated
+- remote reflection was recovered
+- execution-ready packet shape was fixed
+- provider-facing request and dispatch forms were fixed
+
+## Not yet achieved
+
+- one accepted returned filled provider block
+- verified landing in /workspace/parameter-golf
+- one actual baseline/candidate evidence pair
+
+## Audit reading
+
+The loop family has reduced ambiguity around execution readiness, but has not
+yet crossed the external handoff boundary required to begin real attach or
+runtime evidence collection.

notes/tpi_loop_audit_rule.md

@@ -0,0 +1,15 @@
+# TPI Loop Audit Rule
+
+When the ping-pong PR count exceeds 10, insert an audit gate PR.
+
+The audit gate must include:
+- original goal
+- progress since the start of the loop family
+- current bottleneck
+- divergence check
+- course-correction decision
+- exact next action
+
+Continue the current line only if the audit shows that the newest loop still
+reduces distance to execution. Otherwise, record a course correction before
+starting another normal loop.