A curated, read-only discovery index for T81-ORI. Start with the architecture decision, manifest, and agent rules. This initial foundation exposes repository evidence and uncertainty. It does not certify findings or run experiments.
The active index pins merged PR608 at
fabc7aa90f70ead227ec257a1ae6833fa16ced40. PR610 is still an unmerged
draft, checked 2026-10-27. Its correction at
d1eee548d4658df37b35d0cc3781b5a77cbae96b is indexed only as proposed evidence,
not main or an accepted finding. PR609 remains draft pending human review/merge
of PR610 and a subsequent source-pin/receipt refresh. The canonical register
already supplies the exact dated H07 B/C and H08 C/E records; PR608’s stronger
summary dispositions cannot override them.
discrepancy_history retains PR607 identity/wording errors as superseded by
PR608, with resolution limits. PR608’s remaining H07/H08 error is unresolved
until the proposed correction actually merges. get_hypothesis exposes this
history and the integration dependency alongside the canonical record. No new
operation or execution capability was added.
Requires Python 3.9+, Git 2.46+ (a conservative minimum for documented GIT_NO_LAZY_FETCH support), and the dependency in
requirements.txt. Install into
an environment outside the repository. Commands below use python for that
environment’s interpreter; installation/fetch are explicit setup actions, outside
the discovery interface:
python -m pip install -r tools/research_commons/requirements.txt
# Needed in clones that do not contain PR605's separate merged proposal history:
git fetch origin 827d2accbf8ccb56afe3fe8d96dfb7251f249d7d
# Explicit setup for the UNMERGED correction snapshot, not an update of main:
git fetch origin d1eee548d4658df37b35d0cc3781b5a77cbae96b
python -B tools/research_commons/commons.py --validate
python -B tools/research_commons/commons.py --request '{"protocol":"t81.research-commons.v1","operation":"list_research_tracks","arguments":{}}'
Source paths are repository-relative at each source’s revision, not promises that every file exists in the working tree. PR605’s evidence is not on the pinned main tree. Each source has an immutable GitHub URL, commit and SHA-256. Git objects must be available locally; missing objects fail validation. No implicit network fallback occurs; Git lazy fetching, interactive prompts and optional locks are disabled. Working-tree edits to an evidence document do not alter pinned responses. New accepted evidence requires an explicitly reviewed index update; changing the checkout branch alone does not refresh this snapshot.
Request schema, response envelope, protocol t81.research-commons.v1:
| Operation | Required arguments |
Result |
|---|---|---|
list_research_tracks |
{} |
R, F, M, canonical IDs and evidence references |
list_hypotheses |
{} |
All eight complete historical records, evidence and questions |
get_hypothesis |
{"hypothesis_id":"H04"} |
Verbatim canonical record and supplementary references |
get_evidence |
{"evidence_id":"hanoi-negative"} |
Pinned UTF-8 document, role, uncertainty and identity |
get_reproduction_instructions |
{"hypothesis_id":"H02"} |
Curated instructions and requirements; never execution |
list_open_research_questions |
{} |
Research questions, discrepancies and limitations |
Use --request - for one JSON request on stdin. Success writes one canonical,
sorted-key ASCII JSON object to stdout and exits 0. Failure writes a structured
validation_failed error to stderr, no result to stdout, and exits 2. Unknown
operations, arguments, evidence IDs, protocol versions, duplicate keys and
non-finite numbers fail. No shell or command strings are executed from requests.
The CLI is a justified MCP equivalent for this phase; it is not an MCP server.
Responses carry protocol, operation, baseline revision, manifest SHA-256, scope
and result. manifest_sha256 hashes canonical JSON (sorted keys, compact
separators, ASCII), not the formatted manifest file. Every request checks all
indexed source commits, regular-file paths, hashes and references before returning
a result. Evidence responses return full curated text records; archive bytes and
arbitrary paths are not retrieval operations. Nested paths/URLs quoted inside a
historical source are preserved claims, not automatically indexed or verified
references. Follow source_refs/evidence_refs through get_evidence.
The original register is dated 2026-10-25. Its result, confidence, assumptions,
measurement and falsifier fields are returned unchanged. Later R02 and E1 records
are supplementary evidence; there is no computed global verdict. Evidence roles
(reported, negative, modeled, disputed, etc.) classify the indexed record,
not independent replication. The manifest indexes conflicts in PR606/607 and the
continuity hash packet so future agents cannot mistake repeated summaries for
verification. Hash identity proves bytes, not truth, authorization or authorship.
Self-contained Draft 2020-12 schemas cover:
All have status: proposed, explicit uncertainty, canonical hypothesis IDs,
pinned evidence references and authorization status. H-S1 is not H08; proposals
about integration should state the relationship to H07 without renaming it.
Noncanonical refinements belong in the research question/target fields. Schema
validation cannot prove reviewer independence, authentic authorization, adequate
statistics or a competent baseline. Contribution acceptance remains a human act.
python -B tools/research_commons/commons.py --validate-contribution docs/research/commons/examples/h04-proposal.json
python -B -m unittest discover -s tests/research_commons -v
The H08 proposal compares human-only, single-assistant
and adversarial-assistant work under matched budgets and blind independent
adjudication; authorization is not_requested.
The onboarding demonstration uses only these contracts and repository evidence. See implementation and limits for exactly what was tested.