Methodology and limitations
Definitions before claims.
Observed characters, modelled_not_measured economics, privacy controls, and unsupported conclusions remain separate by construction.
Cohort and freeze protocol
The exact cohort is laconic_repository_development: closed OMP and Claude Code sessions whose allowlisted metadata resolves canonically inside the Laconic repository and whose timestamp falls in [2026-09-06T07:53:09Z, 2026-09-16T19:01:24Z). Discovery reads at most the first 50 records of each transcript and only the top-level cwd and timestamp keys.
The owner-local extractor seals transcript size, mtime, content digest, selected runtime-row digest, pricing provenance, and generator provenance before aggregation. It rechecks all selected sources before atomically replacing the public JSON. GitHub Actions never performs extraction.
Observed character equations
chars_avoided = raw_chars − visible_charsboundary_reduction_pct = 100 × chars_avoided ÷ raw_charsemission_rate_pct = 100 × emitted ÷ eligibleThese are character counts at the tool-result boundary. They are not token counts, provider billing, cache behavior, agent behavior, or task quality.
modelled_not_measured economics
The estimate converts observed avoided characters through explicit low/high characters-per-token assumptions, an explicit cache-reread multiplier, and pinned model list prices. The model names host-reported and modelled denominators separately. It withholds dollar figures when fallback-priced cost exceeds 25 percent or required inputs are absent.
This is a single arm. There is no disabled-Laconic counterfactual. A provider never billed the avoided amount.
Privacy boundary
The public schema is exact-key and aggregate-only. It excludes session identifiers and hashes, filesystem and repository paths, host folders, model and provider names, tool names, prompts, arguments, content, server names, credentials, and free-form notes. The closed cohort constant is the sole repository-identifying value.
Public reproduction
uv run python scripts/generate_github_pages.py
uv run python scripts/verify_github_pages.pyThese commands read only the committed canonical public JSON and deterministic demo input. They do not require private session stores. The private extractor is intentionally excluded.
modelled_not_measured
Ordered limitations.
Observed boundary counters remain reportable; broader economic and behavioral conclusions do not follow.
observed_characters_are_not_tokens_or_dollarsCharacter reduction is an observed tool-boundary effect, not a token or dollar count.single_arm_cohort_has_no_counterfactualThis self-use cohort has no untreated or disabled-Laconic counterfactual.modelled_avoided_cost_is_not_measured_savingsAvoided cost is modelled from assumptions; it is not measured or billed savings.token_density_and_cache_rereads_are_unmeasured_assumptionsToken density and cache rereads are explicit model inputs, not observed effects.host_reported_cost_covers_only_hosts_that_report_oneHost-reported cost covers only sessions whose host records a cost.fallback_pricing_withholds_dollars_above_25_percentDollar ranges are withheld when fallback-priced cost exceeds 25 percent.local_laconic_repository_cohort_is_not_generalizableThis local Laconic-repository cohort does not represent other users or workloads.