technical-memorandum

slug	technical-memorandum
name	Technical Memorandum
description	Canonical output. Dated, signed, witness-countersigned by a peer agent. Abstract → problem → prior TMs → method → result → open questions. ≤10 pages.
version	0.1.0
metadata	{"sources":[{"mode":"port-original","author":"Jannes Stubbemann","added_in":"0.1.0"}]}

Technical Memorandum

Port-original skill. Hand-authored for this Agent Company; not from upstream. Shared researcher core skill — referenced by every researcher agent.

When to fire

Any time a researcher produces a result worth recording. The default is always — write a TM whenever a cycle produces a conclusion, a negative finding, a working prototype, a partial result, or a reframing of a prior question. The cost of writing a TM is lower than the cost of an orphan claim floating in the Hallway or a researcher's head.

A "result worth recording" includes: a confirmed or refuted hypothesis; a working or broken implementation; a mathematical derivation; a literature summary with a new synthesis; a failed experiment with a clear explanation of why. If in doubt, write the TM.

Half-results that are not yet ready for a witnessed record go to the Hallway instead. A half-result that survives two more cycles becomes a TM regardless.

Inputs

• The result itself: data, code, derivation, conclusion, or partial finding.
• Prior TMs from memoranda/ that motivated or are superseded by this work. At least one prior TM or Hallway entry must be cited; see Invariants.
• The researcher's own Hallway entries from the current cycle — these form the living context for the Abstract.
• The current MISSION.md, to confirm the TM traces back to the north-star problem.

Outputs

A file at memoranda/TM-NNNN-<slug>.md where NNNN is the next sequential 4-digit integer (zero-padded: TM-0001, TM-0042, etc.) and <slug> is a 2–5 word kebab-case summary of the subject.

The file contains the following required sections in order:

1. Title — the TM's full title, the date (ISO 8601), and the author's role name.
2. Witness — the peer countersignature block (see Procedure). Must be non-empty before the TM is committed to memoranda/.
3. Abstract — ≤200 words. The single-paragraph answer to "what is this TM about and what did you find?"
4. Problem — the question or gap this TM addresses. One to three paragraphs.
5. Prior TMs cited — a list of TM-NNNN references and/or Hallway entry filenames. ≥1 required. Can include the researcher's own prior TMs.
6. Method — what the researcher did: the experiment, derivation, search strategy, or construction procedure.
7. Result — what was found, including negative or partial results. Do not suppress null results.
8. Open questions — things this TM does not settle. Every TM has open questions; that is the whole point of writing things down.

Procedure

1. Draft the TM with all eight sections populated. The Witness block should be marked [pending countersignature] at this stage. Do not leave any section empty, including Open questions.

2. Choose a peer agent for the witness countersignature. Default rule: choose a researcher from a different team. A Theorist's witness is preferentially a Bench or Network researcher, not the Mathematician (same team). An Experimentalist's witness is preferentially a Theory or Invention researcher. The Director can override this default during continuation review by naming a specific witness.

3. Send the draft to the chosen peer. The peer reads the draft in full and returns a structured critique with exactly these four required fields:

   • Claim under examination — restate, in the peer's own words, the central claim of the TM.
   • What I would do differently — one or more methodological suggestions.
   • What I think is wrong — identify gaps, errors, or over-interpretations. This field must be populated; "nothing is wrong" is not a valid entry — the peer must at least identify the weakest link.
   • Sign or refuse-to-sign — either SIGNED or REFUSED: <one-sentence reason>.

4. On REFUSED: the author incorporates the critique, updates the draft, and resubmits to the same peer. After three refusals on the same TM, escalate to the Director: write a one-paragraph escalation note at instigation/YYYY-MM-DD-<author>-tm-dispute-<slug>.md and wait for Director guidance before proceeding.

5. On SIGNED: fill in the Witness block with the peer's role name, the date of signing, and the structured critique verbatim. Commit the TM to memoranda/. Post a short Hallway entry noting the new TM's filename and one-sentence summary.

Invariants

• The Witness block must be non-empty. The block must contain the peer's role name, the countersignature date, and the structured critique verbatim (all four fields). A TM with an empty or [pending countersignature] Witness block may not be committed to memoranda/. Enforced by make check.

• Prior TMs cited must list ≥1 prior TM or Hallway entry — no orphan claims. Every TM must trace its lineage. If no prior TM exists, cite the Hallway entry that motivated the work. A TM with an empty Prior TMs section fails make check.

• TM filename matches TM-\d{4}-[a-z0-9-]+\.md. The four-digit zero-padded sequence number is mandatory. Files that do not match this pattern are not recognized by make check as valid memoranda.

• Author cannot witness their own TM. Self-witness is unconditionally forbidden. The peer listed in the Witness block must be a different agent than the author listed in the Title. Enforced by make check.

Anti-patterns

• "Self-witness." The author counter-signs their own TM to move faster. Forbidden; make check will catch it. If no peer is available, escalate to the Director — do not forge the witness.

• "Refuse-to-sign loop > 3 attempts." Three refusals on the same TM indicates a substantive dispute that the author and peer cannot resolve alone. Escalate to the Director; do not keep cycling indefinitely with the same peer.

• "TM with empty Open questions." A TM that claims to have no open questions is either the end of all science or a sign the author stopped thinking. Every TM has open questions; populate the section honestly.

• "Skip the Witness block to move faster." An unwitnessed TM is not a TM; it is a Hallway entry that overestimates itself. The witnessing step exists precisely because it is inconvenient — that friction is the quality gate.

• "Cite TMs that don't exist." References to TM-NNNN identifiers that do not correspond to actual files in memoranda/ are fabrications. Check that cited TMs exist before committing.

• "Suppress null results." A negative finding is a result. A TM that reports a failed experiment with a clear explanation of why is more valuable than silence. Write the TM.