Inspect before editing
The host reads the current table, list, outline, or flattened frontmatter structure and retains its content hash.
Published, reproducible evidence
480/480 correct. Zero harmful edits.
The Hermes auto profile completed all 480 trials correctly on MiniCPM, Ornith, and Gemma. Every routed request passed the exact tool-surface and host-resolved-argument audit.
| Hermes auto family | MiniCPM | Ornith | Gemma |
|---|---|---|---|
| Tables | 60/60 | 60/60 | 60/60 |
| Lists | 100/100 | 100/100 | 100/100 |
| Sections | 150/150 | 150/150 | 150/150 |
| Frontmatter | 110/110 | 110/110 | 110/110 |
| Table reads | 60/60 | 60/60 | 60/60 |
| Routed subset | 470/470 | 360/360 | 360/360 |
| Standard fallback | 10/10 | 120/120 | 120/120 |
Paired Ornith gain. Standard Hermes scored 460/480;
auto scored 480/480, with 20 treatment-only wins, no control-only wins, and exact McNemar p = 1.907×10⁻⁶. Both routed runs had zero harmful outcomes, framing errors, multiple mutations, or retries.Pi remains independently measured
| Ornith/Pi family | Correct | Outcome |
|---|---|---|
| Tables | 60/60 | Perfect |
| Lists | 100/100 | Perfect |
| Sections | 150/150 | Perfect |
| Frontmatter | 110/110 | Perfect |
| Table reads | 60/60 | Perfect |
| Routed subset | 360/360 | No routing, framing, or argument errors |
Scope matters. The preregistered run used Ornith 1.5 9B Q8 through llama.cpp and Pi with thinking off,
maxTokens: 2048, and parallel_tool_calls: false. One correct ordered-list trial took 184.4 seconds, so tail latency remains a caveat. A preceding holdout stopped at its first destructive result; the targeted repair and complete rerun are both preserved.Why the routed profile works
Route only exact intent
Unambiguous requests receive one small action-specific tool. Unsupported wording keeps the standard interface.
Own dangerous arguments
The host supplies resolved targets, exact predicates, typed values, and create/update preconditions, then stops after one successful routed mutation.
Gemma remains a strong measured option in Pi
Gemma 4 26B completed 475/479 usable Pi trials (99.2%) with zero harmful outcomes. Tables, lists, frontmatter, and table reads were perfect; sections reached 145/149 with four loud refusals. The later Hermes run reached 480/480 on seeds 40–49, but that cross-host comparison is descriptive rather than paired.
MiniCPM: the complete Pi profile reaches 480/480
| Measured path | Result | Disposition |
|---|---|---|
| Singleton-object table rows | 18/18, including 9/9 normalized calls | Shipped compatibility |
| Host-routed list pipeline, direct arm | 13/21 → 21/21, p = 0.0078 | Validates the architecture |
| MiniCPM list profile through Pi | 16/21; all 16 executed edits correct | Opt-in evaluation profile |
| Current Pi safe-routed profile | 480/480; zero harmful outcomes | Enabled by auto |
The final composition covered seeds 0–9: tables 60/60, lists 100/100, sections 150/150, frontmatter 110/110, and table reads 60/60. All 470 routed trials and 10 standard fallbacks passed exact surface audits. The result used MiniCPM5 2B Q8_0 through llama.cpp with thinking off, a 65,536-token context, 8,192 output tokens, temperature 0.7, and top-p 0.95.
The separate Hermes Agent 0.21.3 composition reproduced the same 480/480 family totals, 470 routed trials, and 10 standard fallbacks. Gemma and Ornith each retained 51/51 across the 17 changed tasks, and the post-adoption MiniCPM auto smoke passed 48/48.
Structured operations versus direct patching
The original controlled benchmark used identical tasks, seeds, model, and grader:
| Operation | Direct patch | Incise |
|---|---|---|
| Tables | 60.0% (36/60) | 100% (60/60) |
| Lists | 63.0% (63/100) | 91.0% (91/100) |
| Sections | 19.0% (19/100) | 74.0% (74/100) |
What the operations remove
Character arithmetic
The model no longer needs to count table padding or reproduce formatting exactly.
Boundary guesses
Section and list boundaries are determined by deterministic structure-aware code.
Whole-document rewrites
Targeted splices avoid recreating unrelated bytes and consuming them as model output.
Reproducible evidence
The complete tasks, prompts, raw calls, graders, preregistered plans, statistical comparisons, caveats, failed treatments, and reversed conclusions are versioned with the source. The Rust core is also compared byte-for-byte with an independent Python oracle across 110,406 generated cases over 54 fixtures.