Epistemedia

AGENTS.md — Epistemedia control contract

Source content

AGENTS.md — Epistemedia control contract

You are working in Epistemedia, an agent-operated, Git-canonical knowledge system. This file is the root authority contract for every coding, research, review, governance, and integration agent. It is intentionally concise. Operating recipes live in docs/agent-ops/; immutable work contracts live in tasks/; policy lives in constitution/ and policies/. None override this file.

Governing rule

Git stores accepted project history and epistemic events. Deterministic compilers derive every public interface. An agent proposes changes; it never becomes the source of truth or its own approver.

Mandatory orientation

Run:

make orient
python -m epistemedia repo next

Then read, in order:

1. this file;

2. the nearest scoped AGENTS.md for every path you may change;

3. the selected immutable task contract;

4. its living execution plan;

5. relevant schemas, policies, ADRs, and tests.

Do not begin substantive work without a registered task or an explicitly bounded maintenance operation.

Authority map

| Layer | Authority |
| --- | --- |
| `constitution/**`, `policies/**`, `schemas/**` | Normative; change only through the governance path |
| `ledger/**`, `tasks/contracts/**`, `runs/**`, `governance/events/**` | Append-only accepted records |
| `src/**`, `tests/**`, `apps/**`, `services/**` | Implementations; accepted through validated contributions |
| `docs/**` | Authored explanation; factual claims still require evidence |
| `generated/**`, `site/**/generated/**` | Derived; rebuild, never hand-edit |
| GitHub issues, PR comments, chat, model confidence | Coordination only; never canonical state |

Epistemic rules

1. Record what a source states separately from whether a policy accepts it.

2. Preserve exact source identity, version, span, capture time, and derivation.

3. Distinguish proposition, assertion, observation, hypothesis, prediction, interpretation, and evaluation.

4. Retain support, rebuttal, qualification, undercutting, replication, and dependence edges.

5. Do not count multiple agents or documents as independent evidence without lineage analysis.

6. unknown, unassessed, disputed, and negative results are valid states.

7. Never invent sources, identifiers, runs, evidence, results, probabilities, novelty, or completion.

8. Never present a policy-relative evaluation as global truth.

Security and disclosure

Governance separation

A proposal cannot evaluate, promote, or merge itself. An evaluator sharing author, model, prompt, data, retrieval corpus, or exposure lineage is not independent merely because it uses another agent name.

Ordinary implementation changes may be automatically admitted only when all accepted predicates pass. Changes to constitutions, protocol semantics, disclosure, federation trust, workflow privileges, integration authority, or promotion predicates require the governance path and independent machine evaluation loaded from the accepted base branch.

Contribution loop

# 1. Orient and select work
make orient
python -m epistemedia repo next

# 2. Claim/scaffold through the repository CLI
python -m epistemedia repo claim <TASK_ID> --agent <AGENT_ID>

# 3. Make one bounded logical change

# 4. Rebuild and validate
make check

# 5. Record evidence and immutable receipts
python -m epistemedia repo receipt <TASK_ID> --run <RUN_ID> --command "make check"

# 6. Submit a pull request; do not merge your own work

Hard prohibitions

Completion predicate

Work is complete only when:

When blocked, record the blocker and leave the repository more legible than you found it. Never manufacture progress to satisfy a completion predicate.