[jfrench9]

Summary

Two things, found in this order. (1) A harness defect in the published v0.2 run, disclosed and fixed. Every tool the ladder served from xbrlkit — 4c's nine, and the search_text / read_text instrument on 5c+doc, 7b+doc and 7c+doc — reached the model with an empty argument schema: the capture read the listing under the wire name inputSchema, which the mcp 2.x Tool object does not carry as an attribute, and silently fell back to {"type": "object"}. The 7a path reads input_schema and was whole. Measured from the published transcripts: 4c's 44 tool errors in 419 calls are all schema-validation refusals of names the model inferred from descriptions (33 concepts for elements, 4 phrase for query, 7 types and bounds); the three +doc rows show 6 refusals in 781 served calls. No score changes — the judge scores answers — but 4c's error, turn and cost figures carry a harness handicap, and the results README attributed the refusals to the model guessing against a schema it never saw. Recorded as a dated §9 erratum and a section in results/v0.2-sonnet-5-5/README.md, following the 2026-09-06 precedent; the capture now reads the attribute by its name and refuses a listing without a schema.

(2) Protocol v0.2.1 pre-registered (§12): two rows added to the Sonnet 5.5 column beside 4c and 7a under §8's last-number rule, each its base with one friction removed. 4c.1 is 4c with the schemas handed over, on xbrlkit 0.21.5 (RoboFinSystems/xbrlkit#105), whose fact_grid and resolve_element also take the names the model sent. 7a.1 is 7a with workflow steps 1 and 2 pointed at the rung's shaped tools instead of the Cypher prerequisites: 46 of 114 records opened with a 6,718-character get-example-queries fetch, and Cypher followed in 9. The v0.2 rows carry forward by identity.

Changes

• filing_ladder/representations/xbrlkit_serve.py — tool_defs_from(listed) / _schema(tool): the schema under input_schema (or the wire name), a listing without one raises instead of defaulting; ALIASES and accepts_aliases(tools) for the 4c.1 guard.
• PROTOCOL.md — §9 erratum dated 2026-11-09 (the finding, the footprint, no score changes, what it does to the claims, the fix); §12, the pre-registration: the two rows, the frictions measured, the exact change to each (old and new 7a step text quoted; 4c.1's two-part repair and why the row does not separate them), the fairness argument, two lock deltas, six expectations, cost ($15–18 + about $2 of judging; a k = 1 shakedown first, never scored), publication.
• results/v0.2-sonnet-5-5/README.md — the 4c bullet corrected (the model saw no schema) and a new section "What the served tools' schemas contained (added 2026-11-09)" with the per-row table; nothing else in the directory touched, hashes unchanged.
• filing_ladder/ladder.py — Rung.XBRLKIT_SERVE_V021 (4c.1), Rung.LPG_SHAPED_V021 (7a.1) and their specs; both built on their base via _BASES; V0_2_1_RUNGS; parse_rungs("v0.2.1").
• filing_ladder/prompts.py — LPG_SHAPED_SOURCE shared by 7a and 7a.1; SOURCES[4c.1] is SOURCES[4c]; SOURCES[7a.1] with the two new steps.
• filing_ladder/cli.py — 4c.1 is cannot_attempt when the served fact_grid schema lacks concepts; --rungs help lists v0.1.1 | v0.2 | v0.2.1.
• pyproject.toml, uv.lock — xbrlkit[lpg,mcp]>=0.21.5, locked to 0.21.5.
• tests/test_xbrlkit_serve.py — the served definitions carry the server's schemas (every SERVE_TOOLS schema has properties; fact_grid has elements and concepts, resolve_element query and phrase, search_text pattern, read_text offset / length); a listing without a schema is refused. tests/test_v0_2_1.py — the two rows' flags and bases; 4c.1's prompt is 4c's object; 7a.1's prompt equals 7a's outside steps 1 and 2; the alias guard.
• .gitignore — !results/v0.2.1-sonnet-5-5/. README.md — the status sentence on v0.2.1 and the disclosure.

Protocol impact

RUNG FIDELITY and PROTOCOL. Rung fidelity: in v0.2, what 4c and the three +doc rows handed the model differed from what §11 described — the served tools' argument schemas were empty. Disclosed in §9 and the results README with the footprint; no published number changes; the token table is unaffected (schemas are not in it). Protocol: §12 adds two rows under §8's last-number rule; 4c and 7a keep their records and hashes; 7a.1's prompt change is on a new row, not on 7a.

Testing

just test-all green (format, lint, typecheck, the suite with the new tests). The schema fix and the alias guard were run against the real server over the materialized package 0000012927-25-000015 on xbrlkit 0.21.5: all nine SERVE_TOOLS schemas carry their properties, accepts_aliases is true, and fact_grid with concepts returns byte-identical rows to elements. Nothing paid ran: the shakedown and the two rows wait on the protocol-v0.2.1 tag and a go.