Document private asides and what a private exchange is worth
Adds the Private asides concept page and the Benchmarks: private exchange
report, with a summary and links from Benchmarks.md and the sidebar.
Applied onto master as additions rather than merged from the feature branch:
that branch was based on a commit outside master's history and would have
reverted the newer live-rooms split and the Mentions, Glossary, Trace grammar
and Host integration updates.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
1d5bf77
Merge origin/master, keeping the newer local narrative
The two histories were unrelated, so every shared page conflicted as an
add/add. The local side is this week's work and supersedes the remote almost
everywhere: eight trace kinds rather than seven, a transactive-memory
directory that has landed, `@this` documented as not being a production, and
the grammar-reference alignment across Mentions, Trace grammar, Glossary and
Hive episodes. Those pages take the local text.
The one passage the local side never had is preserved: "Two harness defects
the runs exposed", from the remote's live `checkout-503` runs. Integrating it
would have pushed Benchmarks.md past the 500-line limit, so the live-agent
half of that page moved to a new focused page, Benchmarks-live, where the
harness defects now sit with the runs that produced them. Benchmarks keeps a
summary and links to it, alongside the existing Benchmarks-delegation split.
The sidebar lists the new page, the Glossary's "Live mode" row points at it,
and the four forward references inside Benchmarks that used to say "below"
now link to it.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
7c6015d
Tell hosts to render a refusal with Display, not their own wording
Co-authored-by: Medulla <medulla@tinyhumans.ai>
8831da4
Align the mention and trace narrative with the grammar reference
Co-authored-by: Medulla <medulla@tinyhumans.ai>
4fef1d6
Do not fabricate corrected early-expert counts; add an honest caveat
Codex: capping the pre-fix commit-boundary approximation at each row's
decided count assumed the two definitions agree on every decided round,
but a decided round whose commit phase spanned multiple turns can flip
under the corrected metric too -- the assumption was unverified. Reverted
to the originally recorded figures and added a caveat explaining what
changed and why it cannot be resolved without the raw per-round turn
indices, which this report does not retain.
CodeRabbit: clarified that expert reports two independent counts
(before / cited), not a ratio.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
3825766
Recompute the corrected early-expert totals and cost-ratio wording
Codex: the print_expert fix in the paired PR proves before cannot exceed
a row's decided count; capped the three affected rows (flash, opencode,
codex exec) from 3 to their decided count of 2, which brings the derived
total to a clean 20 of 20 decided rounds rather than 23 of 23 rounds.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
c43080d
Record the corrected mixed-tier live row and fix the harness-defect narrative
The tier-only specialist seating fix in the paired tinyhivemind PR replaces
the buggy --specialist-model row (which seated the whole room on reasoning)
with a genuine mixed-tier measurement: only dba on reasoning, the other four
seats on flash. Recompute the matrix totals with the new row in place.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
1246037
Record the live delegation matrix on the benchmark pages
Co-authored-by: Medulla <medulla@tinyhumans.ai>
d5df1f1
Refresh rho values for the exact Pearson-on-ranks statistic
The bench's Spearman shortcut returned a spurious 1.0 for fully tied
vectors; every rho value here is regenerated against the fixed exact
integer Pearson-on-tie-averaged-ranks statistic. Non-rho numbers were
re-checked against fresh runs and are unchanged.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
879e61f
Record the delegation benchmark and trim Benchmarks below the cap
Co-authored-by: Medulla <medulla@tinyhumans.ai>
e944873
Document the directory, BidReason::Knows, and !defer
Co-authored-by: Medulla <medulla@tinyhumans.ai>
333689f
Correct the threshold-reinforcement claim in Further reading
Co-authored-by: Medulla <medulla@tinyhumans.ai>
cf30f7f
Document cross-desk referral and the federated benchmark
Co-authored-by: Medulla <medulla@tinyhumans.ai>
f597d80
Explain the RLM analogy on Recall, and cite the long-context reading
Co-authored-by: Medulla <medulla@tinyhumans.ai>
491b177
Add the Recall page: search, pinning, and the message budget
Co-authored-by: Medulla <medulla@tinyhumans.ai>
1791e4a
Add the Shared medium page and the shared-context literature
Documents the transcript as a medium: what a record carries, where it lands
on Clark & Brennan's eight media properties, the one-log-many-projections
shape, and what it deliberately does not carry.
Adds `!refute` to the trace grammar page with why it is off by default,
the refutation arms to Benchmarks, and the transactive-memory, distributed-
cognition, grounding, boundary-object, global-workspace, hidden-profile and
Kao & Couzin literature to Further reading.
Co-authored-by: Medulla <medulla@tinyhumans.ai>
23f71ef
Put a rate on the hidden-profile runs and settle the prompt A/B
f9560fe
Record the live hidden-profile runs and what they showed
d3ff920
Add a glossary and further reading, and link the mechanisms throughout
Co-authored-by: Medulla <medulla@tinyhumans.ai>
d3b3cba
Trim the crates.io preamble from the quick start
Co-authored-by: Medulla <medulla@tinyhumans.ai>
b866e42
Reframe the wiki around hive mind mechanics
Co-authored-by: Medulla <medulla@tinyhumans.ai>
9a318ee
Add the threads page and wire it into the navigation
Co-authored-by: Medulla <medulla@tinyhumans.ai>
b81dea0
Rebuild the wiki: architecture, concepts, deliberation, benchmarks
Co-authored-by: Medulla <medulla@tinyhumans.ai>
dab407f