Open-model hardening (gated tracking issue)

[dasirra]

Pending human review. Edit freely before implementing.

Phase 4 · Part of #2

User story

As the forge maintainer, I want a single tracking issue for open-model hardening whose concrete sub-work is cut directly from the #6 findings, so that I build only the compensations the evidence shows are needed and file no speculative work I may immediately close.

Proposed behavior

This issue stays an empty placeholder until the #6 report lands; it opens no sub-work before then. When the report exists, sub-work is cut from it. The candidate compensations, each opened only if the report flags its risk as observed, are:

• (a) require the evaluator to paste verbatim command output per criterion before a pass;
• (b) split the long orchestrator into per-phase units and extract mechanical steps (grounding grep, gh sequences) into shell scripts;
• (c) embed a definition-of-done contract schema and validate it each round;
• (d) validate mermaid diagrams before filing;
• (e) apply the same verbatim-evidence rigor to the planning critic that (a) applies to the building evaluator.

Risks the report clears are dropped, not built. Any candidate that touches methodology must still round-trip on Claude Code. If the report clears everything, this issue closes with no sub-work.

Acceptance criteria

• No hardening sub-work is opened before the Empirically validate the pipeline on Pi/GLM-5.2 and set compensation depth #6 report exists.
• Each opened sub-item cites the specific report finding that justifies it; risks the report cleared are explicitly recorded as dropped, not built.
• Any methodology change round-trips on Claude Code: substituting the Claude binding into the changed prose still reconstructs today's tool names, model tiers, and surfaces.
• On a Pi re-run, each shipped compensation is observed to do its job (for example evaluator/critic pastes evidence; per-phase orchestrator retains its rules; contract/mermaid validation halts on bad input).

Out of scope

• Anything the Empirically validate the pipeline on Pi/GLM-5.2 and set compensation depth #6 report does not flag as an observed risk.

Grounding

• The Empirically validate the pipeline on Pi/GLM-5.2 and set compensation depth #6 findings report that scopes this work. NEW (produced by Empirically validate the pipeline on Pi/GLM-5.2 and set compensation depth #6).
• The evaluator evidence discipline candidate (a) targets: commands/building.md lines 288-293. EXISTS.
• The orchestrator prose candidate (b) restructures: the nine phases of commands/building.md. EXISTS.
• The contract structure candidate (c) would validate: commands/building.md lines 33-47. EXISTS.
• The mermaid outputs candidate (d) would validate: commands/planning.md lines 100-104, 195-198, 210-213, 240-244. EXISTS.
• The planning critic candidate (e) targets: commands/planning.md lines 118-140. EXISTS.
• Any new scripts or schema files. NEW.

Dependencies

Blocked by #6 (hard blocker). Also depends on #3 for the neutral prose.

---

Granular contract to be negotiated by generator/evaluator at /forge:building time and posted back to this issue as a comment.