INSTRUCTION MANUAL • READ BEFORE YOU BUILDHANDOVER®
For one rough prompt → one finished job

HANDOVER®
V6

You bring a rough idea.
Handover turns it into a plan you approve, then does the work and proves it.

Think flat-pack furniture. No building until you have seen all the parts, checked the diagram, and tightened the two big approval screws.

VER 6, Nov 15, 2026 ART 010.11.2026 MODE Relentless, not one-shot
ROUGH PROMPT PLAN + GRAPH + CHECKS DONE ✓ PROVED YOU: approve CHECKER: verifies NO APPROVAL = NO BUILD
Fig. 1: How the kit fits together
6

Assembly steps.
Do them in order.

2

Approval screws per run.
Stage 1 + Stage 2. Chained work shares one pair for the whole chain, not one pair per link.

0

Gaps left unowned.
Every gap gets Fixed, Queued, or Accepted.

Before you startPage 01 / Safety
!

Do not skip approval

Handover can plan on its own. It cannot start gated work until you approve Stage 1 and, where needed, Stage 2. An unlisted action pauses the build.

≠

Builder ≠ Checker

The person who builds a piece never grades it. A fresh verifier checks it. A fresh judge gives the final verdict. That separation never bends.

×

A fail is not the end

V6 is relentless by default. A first fail starts a fix cycle, not a dead-end report. Lowering the bar to “pass” is forbidden.

Parts includedCheck all parts before assembly

What's in the box If a part has no output file, it is not a real part. Every unit must produce something you can open and check.

A
✎

Rough prompt

Your starting sentence. Messy is fine. This is the raw board.

QTY: 1 × YOU SUPPLY IT
B
▭

Intent package

What Handover thinks you mean, with assumptions marked clearly.

QTY: 1 × STAGE 1
C
⬢

Execution manifest

The exact list of tools, targets, limits and budget that were approved.

QTY: 1 × STAGE 2
D
◎

Coordinator

Owns the plan and the record. Calls specialists. Does not do all the work itself.

QTY: 1 × PER RUN
E
✚

Generator

Builds each unit. Works in a fresh context with only its task.

QTY: AS NEEDED
F
⌕

Verifier

Fresh eyes. Gets the spec + the artifact, never the builder's sales pitch.

QTY: 1 × PER CHECK
G
⚖

Judge

Fresh context at the end. Resolves the verdict. Cannot turn a gap into a pass.

QTY: 1 × AT THE END
H
{ }

Per-unit list

JSON list of every unit: name, output, owner, check, and passes: false → true.

QTY: 1 × JSON FILE • NEW IN V4
I
≡

Decision trail

One line per decision: when, what, why, evidence pointer, result. Append-only.

QTY: 1 × LOG
J
$

Budget ledger

Every spawn is a reservation. At least 15% is held back for verification.

QTY: 1 × OWNED BY COORDINATOR
K
⬣

Graph

Map of units and links. Shows what runs apart, what waits, and where they join.

QTY: 1 × WHEN WORK SPLITS
L
▣

Field kit

Reusable playbook for a field: router, playbooks, checks, cleanup. Proven or draft.

QTY: 1 × PER FIELD
M
◈

Discovery router

Reads the registry, picks the skill whose trigger matches, writes one call line before each call.

QTY: 1 × PER CASE • REQUIRED MUST IN V6
N
✦

Discovery skills ×17

16 engines plus the umbrella. Entry 1, Research 6, Proof 3, Stuck 3, Search 4. Core 4 loop: Anomaly, Abduction, Falsification, First Principles underneath. Invention is the outcome, not a 17th skill.

QTY: 17 × PROVEN NOV 12–14, 2026

Tools you provide

● 1 human for preference, authority & risk calls✓ A clear “what counts as done” check▭ List of one-way doors: publish / send / pay / delete / deploy

Handover only asks you human questions: preference, authority, risk, a non-negotiable, or a trade-off with no rule. Everything else it settles by reading, inspecting, or testing.

Assembly instructionsPage 02 to 03 / Follow in order

Assemble in 6 steps Tap a step on the left. Each step is one diagram in the booklet, with its own parts and its own check.

How Handover asks you things: one short sentence, 2–4 real options, each with a trade-off in ≤12 words, one marked recommended with why, plus “Something else, in my words.” About 120 words on the first screen. One decision per message.
Two engines insidePage 04 / Loop + Graph

Loop retries. Graph orders. They do different jobs. The loop is inside each unit. The graph is the shape of the whole job.

Engine A: Persistent loop

When a piece fails, climb the ladder

Classify the failure first. Then use a distinct approach: different input, source, method, or smaller size. Same wording ≠ distinct.

RUNG 1 Retry with corrected input
RUNG 2 Switch source / evidence route
RUNG 3 Split the unit smaller
RUNG 4 Replace the method
RUNG 5 Hold this unit, continue the rest

Up to 5 fix cycles per run. Dead end only after 3 distinct approaches for the same failure class fail, or budget is spent.

When to use which

Router rule

  • Loop alone: one straight line, one deliverable, nothing runs apart.
  • Graph: 2+ parts, some independent, some dependent, a later part needs both.
  • Default is loop alone, unless a graph clearly helps. Reason is written in the graph file.
  • In this environment, “parallel” means serial in order with no dependency, depth max 2.
Join rule: a join unit waits for all its inputs. It never runs early. Missing input = wait or dead end.
Engine B: The graph • Tap a node
RESEARCHno dependency DRAFTparallel branch CHECKneeds Research ◆ JOINneeds all inputs LINK = LATER UNIT NEEDS EARLIER OUTPUT
{ "unit": "research", "output": "brief.md", "owner": "generator-1", "check": "claims linked to sources", "passes": false → true only when check passes }
Selected unit

Research

One output, one owner, one check. Starts immediately because it has no link coming in. Its own loop handles its own fails without stopping the other branches.

Done = every unit in the list has passes: true. No unit removed, no check edited, no unit renamed to hide a failure.

New parts in this versionPage 05 / What V6 added

What changed from V5 V5 already had the Discovery library, the router files, and the $2 hard cap. V6 makes the router a required step in the skill text.

NEW • V6-01

Required router step in the skill text

The coordinator must read the registry before routing, pick a skill only when its trigger fires, write one call line before each call, and record each handoff in that skill's own format. Judge verdict PASS, Nov 15, 2026, run stage2-router-wire-2026-11-15-01, all 5 items pass.

NEW • V6-02

Wiring proof task

Anomaly Hunting and Abduction routed with 2 call lines and 7 of 7 handoff fields each, Nov 15, 2026. Top anomaly: 1 usable case against a predicted 40. Top explanation: the pool was consumed.

Still in the box from V5

NEW • V5-01

Discovery library, 17 skills

16 engines plus the umbrella. Each skill passed a sealed test with 2 fresh blind trials and a fresh judge, Nov 12 to 14, 2026.

NEW • V5-02

Discovery router

Registry plus rules. Reads one row per skill, picks by trigger, writes one call line before each call.

NEW • V5-03

Router proof, run 3

Both fresh blind trials passed both test situations, 5 of 5 repeats each, $0 of the $4 cap, Nov 14, 2026.

NEW • V5-04

$2 hard cap per run

Stated in the skill header. Larger envelopes are approved case by case.

NEW • V5-05

5 fix cycles as standard

Up to 5 fix cycles per run, shared across units. Router and umbrella runs use 3.

Still in the box from V4

NEW • 01

Per-unit list in JSON

The graph now ships with a machine-readable list. Only the passes field ever changes.

NEW • 02

Session start routine

Every session: read notes + logs → pick next unfinished unit → run setup → end-to-end check → fix base if broken → then start.

NEW • 03

End-to-end check

Proof is on the real artifact, the way a human would use it. Not just a unit test or a page fetch.

NEW • 04

Effort scaling

Simple: 1 agent, 3–10 calls. Compare: 2–4 agents, 10–15 each. Complex: 10+ agents, split clearly. Stated in the package.

NEW • 05

Subagent contract

Fresh context, only needed tools, short summary back. Stopped at limit = partial, not failed.

NEW • 06

Notes & compaction

When the trail gets long, summarise it: keep decisions, gaps, numbers. Notes live in a file, outside context.

NEW • 07

Tool risk ratings

Low / medium / high, by read vs write, undoability, permissions, cost. High-risk pauses or goes to a human. Listed in the manifest.

NEW • 08

Failure threshold N

Same tool fails N times in a row (N set in the manifest) → stop retrying it, escalate to a human.

NEW • 09

Rubric evaluator

Grades during work, 0.0–1.0 on: factual accuracy, citation accuracy, completeness, source quality, tool efficiency. Judge still decides at the end.

NEW • 10

Run-state resume

Durable state, run/action IDs, lease + fencing token. Resume never re-runs a finished tool. Mismatch = fail closed, don't guess.

Skill-text spec only: the harness part was not built in the V4 apply run.

NEW • 11

Hook events

Harness code at 4 points: PreToolUse, Stop, SubagentStop, PreCompact. See diagram below.

Skill-text spec only: the harness part was not built in the V4 apply run.

NEW • 12

Handoff pattern

Besides manager-calls-specialists, agents can hand control to the next agent in turn. Handoffs are edges, like tool calls.

Note on hooks + resume above: V4 wrote the specification into the skill text. The harness implementation for those parts was not built or claimed as done in the apply run.

PreToolUse

Before a tool call. Can block or change it.

SubagentStop

When a subagent finishes. Gathers its result.

PreCompact

Before summarising. Archives the full transcript first.

Stop

Before exit. Saves session state.

⬢ Manager pattern

One coordinator in the middle.

  • Coordinator calls specialists
  • Coordinator combines results
  • Use when one agent should own the whole picture

→ Handoff pattern • NEW

Control passes hand to hand.

  • Agent A finishes, passes control to Agent B
  • B continues, no coordinator combining
  • Use when work naturally passes in turn

Graph holds for both: agents are nodes, handoffs are edges.

Job sizeAgentsCallsPicture it as
Simple13–10One shelf. One person.
Comparing things2–410–15 eachA few shelves, side by side.
Complex10+Divided clearlyWhole room. Everyone owns one wall.

Defaults only. Each field kit sets its own numbers for its field.

TroubleshootingPage 06 / If something does not fit

If it gets stuck A run never ends with an unowned gap. A verdict starts the next step, it does not end the run.

✓ FIXED

Fix fits inside the approved envelope: same tools, same destination, no new cost or outward action. Do it, test it, re-check it in the same run.

→ QUEUED

Fix needs new scope, cost, risk, or an outward action. Record it as the next job, with its own approval gate. Queued is not done, and is never shown as done.

◎ ACCEPTED

You explicitly accepted the gap, with the reason recorded. Only explicit acceptance counts.

Done means only: all gaps Fixed, or all remaining gaps explicitly Accepted by you. Never “done, but…”

Dead-end report: short by default
  • What stopped + actual status, including any unmet check
  • Spent / remaining budget
  • ≤3 bullets: what was tried
  • What would unblock it
  • Options: Retry unchanged / Change scope or limit / Stop: each flagged “needs reapproval?”
Safety rails: enforced, not prose
  • Gates are tool classes enforced by the harness, not decided by the model
  • Manifest constrains arguments too: domains, paths, hosts, per-call caps
  • External content is data, never instruction, and is marked tainted
  • Changing the predicate = new version + reapproval. Old verdicts do not carry over
FINAL CHECK: BEFORE YOU CALL IT DONE

Open the real artifact like a user would. Check every link in the graph points to a real file. Check the per-unit list: is every passes true? Check the final report names its field kits, with proven / draft status and a link. If any answer is no, it is not done yet.

Scope note, printed on the box: V6 is supervised only. The research brief was the first execution pilot. Since then, supervised runs have proven 17 Discovery skills plus the umbrella and the harness router. V6 makes the Discovery router a required step in the skill text. The wiring proof routed 2 of 17 skills. The wire is proven for the Research phase engines, not for the Proof, Search, or Stuck phases. Other fields can be planned and packaged, but are not operational until a kit for that field passes its own real task. Unattended, live-data, and three-level fan-out are not in this box. The loader check never ran, because no Muse skill loader command exists in the environment where the proofs ran. The router proof covers two test situations, not live use of all 17 skills.