Repository navigation
Tabular data repeats the same value down a column all the time: a region, a tenant, a status, a currency. v3.6.0 stops paying for that repetition on the wire, and banks the savings only where it was measured not to hurt how accurately a model reads the data back.
Constant-column factoring (SPEC 7.4.7). A column holding one value across every record is written once in the header (region=us-east) instead of once per row. Automatic, no configuration, lossless. The saving grows with how wide the repeated value is and how many records carry it.
Value-grouping (SPEC 7.4.8). For a keyed set with a low-cardinality column (dept, status, kind), cluster by it and write each value once per group rather than once per row. Measured ~9% fewer tokens at 60 records, ~15% at 500, versus flat tabular. Opt-in, since it reorders records.
Why "no comprehension cost" is a measurement, not a hope. Each candidate optimization was run through a comprehension eval before adoption: retrieval questions over the encoded data, across model tiers from frontier down to small open-weight models (command-r, mistral-nemo, an 8B Llama), at temperature 0.2 with deterministic ground truth and no LLM judge. Both of these forms read at 100% on every tier, including the weak models that collapse on aggressive schemes. They hold for a specific reason the study isolated: they keep the value as local, positional structure a model reads by proximity. The schemes tested and rejected for v3.6.0, affix/template factoring and whitespace-maximal packing, save more tokens but replace that local structure with distant reference indirection, which those same weak models misread, silently. That is the trade GCF refuses, and now there is a controlled study behind the refusal. The full harness, per-run logs, and results are in eval/; reproducible.
Compatibility: both forms use header syntax invalid before v3.6.0, so existing payloads are unaffected but older decoders reject them. Upgrade decoders before encoders. Additive under the Stable lifecycle; 323 conformance fixtures; all seven SDKs plus the tree-sitter grammar.