clawock investment-decision

You are the model side of a clawock decision run. DeepSeek Harness owns
conversation, memory and tools; clawock owns the decision contract. Your job
is one file in, one file out: read a request, research, write a decision, let
Python validate it. The model can never grade itself.

prepare ──► request.json ──► you: research + write decision.json ──► publish ──► receipt
(certified │                       (collect evidence,                │       (Python settles:
 context + │                        run the bull/bear debate,          │        evidence counts, debate
 gates)    │                        pick a bounded action)             │        links, money/FX to the cent)
           └───────────────────────────────────────────────────────────┘

Prerequisites

clawock must be installed (Python ≥ 3.11):

python -m pip install clawock

If a workspace is not initialized yet, create it with the workflow pinned —
that flag is what turns decision.json into the enforced contract:

clawock init ./book --workflow investment-decision
clawock workflow install investment-decision --workspace ./book

The second command copies the readable contract (including
references/decision-contract.md and the JSON Schema) into
./book/.agents/skills/investment-decision/ — a standard Agent Skill
location that DeepSeek Harness also discovers.

The loop

1. Prepare

clawock run prepare --workspace ./book

Read the emitted request file (the request_file path printed on stdout,
e.g. .clawock/work/<run_id>/request.json). It contains:

• task — the decision contract (evidence, opposing case, bounded action,
  reconciled amounts)
• context.documents — the certified context, per-file sha256 fingerprints
  (anything you cite must come from here or from your own research, and must
  be observed no later than as_of)
• workflow.parameters — gates: min_supporting_evidence,
  min_opposing_evidence, max_confidence_without_primary_source

2. Research, then debate, then decide

Research with the host's normal tools (web search, file reads, the live
dashboard at https://kcnyu.github.io/clawock/ as source_class: market
evidence). Collect both sides on purpose — the workflow has an opposing
minimum, and publish refuses a decision whose bear case is a strawman:

1. Gather supporting evidence (filings/financials → primary, reputable
   secondary analysis → secondary, market data → market).
2. Actively seek the strongest counterargument you can find or construct —
   a genuine bear case, not a token objection. Record it as opposing
   evidence.
3. State the thesis and its invalidation conditions before choosing an
   action.
4. If the action proposes an order, compute both money identities from the
   context (quantity × unit price; × FX) — never invent prices, FX or sizes.

Write decision.json at the workspace root — artifact paths are resolved
against the workspace root, not the request directory:

{
  "schema_version": 1,
  "workflow": {"id": "investment-decision", "version": "1.1.0"},
  "decision_id": "example-2026-08-08",
  "as_of": "2026-08-08T08:00:00+00:00",
  "subject": {"ticker": "EXAMPLE", "market": "US", "currency": "USD"},
  "evidence": [
    {"id": "filing-growth", "stance": "supporting",
     "summary": "The latest filed revenue figure grew year over year.",
     "source": "issuer filing", "source_class": "primary",
     "observed_at": "2026-08-08T07:30:00+00:00"},
    {"id": "valuation-risk", "stance": "opposing",
     "summary": "The current market multiple is above its stated range.",
     "source": "market data snapshot", "source_class": "market",
     "observed_at": "2026-08-08T07:45:00+00:00"}
  ],
  "debate": {
    "bull_case": {"summary": "Filed growth supports continued monitoring.",
                  "evidence_ids": ["filing-growth"]},
    "bear_case": {"summary": "Valuation leaves insufficient margin of safety.",
                  "evidence_ids": ["valuation-risk"]}
  },
  "thesis": {
    "statement": "Momentum is constructive, but price does not compensate for valuation risk.",
    "confidence": 0.7,
    "invalidation_conditions": ["Filed growth reverses"]
  },
  "decision": {
    "action": "watch",
    "rationale": "The opposing valuation evidence blocks an entry.",
    "evidence_ids": ["filing-growth", "valuation-risk"],
    "order": null
  }
}

Rules that are enforced by clawock run publish, not by this file:

1. debate.bull_case and debate.bear_case are both required and each cites
   evidence; the bear case must cite opposing-stance evidence — publish
   refuses without it. Python checks that linkage, not sincerity: make the
   bear case a genuine counterargument, never a strawman — that honesty is
   your contract.
2. Every evidence row needs a stance (supporting | opposing |
   context), a source_class (primary | secondary | market |
   agent) and an observed_at no later than as_of.
3. All evidence_ids references must exist (decision and both debate cases);
   conclusions in thesis.statement and decision.rationale should be
   backed by cited evidence.
4. If you propose an order, the currency amounts must reconcile to the cent;
   let the numbers come out of the context, never invent them.
5. Confidence above max_confidence_without_primary_source requires a cited
   primary source.

The sibling files next to this SKILL.md carry the full contract:
references/decision-contract.md (with assets/order.example.json, an
order-carrying decision whose two money identities reconcile). Introspect
the exact artifact shape any time: clawock workflow schema investment-decision decision.json.

3. Publish and report

clawock run publish \
  --workspace ./book \
  --request ./book/.clawock/work/<run_id>/request.json \
  --artifact decision.json=./book/decision.json

Report the receipt's status and run_id, and if status is published,
show the user a compact decision card:

Subject      EXAMPLE (US/USD)          decision_id example-2026-08-08
Bull         Filed growth supports continued monitoring.        [1 citing]
Bear         Valuation leaves insufficient margin of safety.    [1 citing]
Thesis       Momentum constructive, price does not compensate for risk.
Confidence   0.70  ·  Action watch  ·  Invalidation: filed growth reverses
Receipt      published  ·  run_id <run_id>  ·  certificate pinned

Never edit the receipt, never re-grade the decision, never "improve" the
artifact after publish — Python settles.

Before publish: self-check

Run this checklist before invoking publish — Python re-checks all of it
anyway, but a clean first pass beats a rejected receipt:

• both debate.bull_case and debate.bear_case present, each citing
  existing evidence_ids, the bear case citing opposing rows
• every cited evidence_id exists; no duplicates
• every evidence row has stance, source, source_class, and
  observed_at no later than as_of
• one primary row is cited whenever confidence exceeds the cap
• an order (if any) satisfies gross_amount_quote = quantity × unit_price
  and gross_amount_base = gross_amount_quote × fx_quote_to_base, to the
  cent

If publish rejects

The receipt lists validation_issues with code + message. Repair only
the named issue and retry against the same prepared request; re-run
prepare only if the context or workflow certificate changed.

typical code	meaning	fix
unknown_fields / missing_fields	a field name is wrong or extra	align with the JSON example above; no reasoning, no opposing_case
insufficient_supporting_evidence / insufficient_opposing_evidence	gate minimums not met	add more evidence rows with that stance
unsupported_bear_case / unsupported_bull_case	debate case cites the wrong stance	point bear_case at opposing rows, bull_case at supporting rows
missing_evidence_reference	evidence_ids names an unknown row	only cite ids that exist in evidence
gross_amount_mismatch / fx_amount_mismatch	order arithmetic wrong	recompute quantity × unit_price and × fx_quote_to_base to the cent
confidence_without_primary_source	confidence above the cap, no primary cited	add a primary row and cite it in decision.evidence_ids
future_evidence	observed_at after as_of	backdate or fix timestamps

Where things land

path	contents
<workspace>/decision.json	your artifact (read by publish)
<workspace>/.clawock/work/<run_id>/request.json	the certified request
<workspace>/.clawock/runs/<run_id>/	receipt + manifest (the settlement record)

Learn from outcomes

After the stated horizon, record price/FX outcome evidence and run
clawock workflow evaluate investment-decision --decision ... --outcome ....
The result reconciles price and FX before assigning a basis-point score — it
is evidence for the next decision, not portfolio P&L.

Record conversation verdicts

A conversation verdict (加仓/减仓/建仓/清仓/观望) is a decision — write it
to the shared decision-mind ledger, the same memory/decisions.jsonl the
daily OpenClaw brief writes to, so both sides can read it and the daily
settlement round-trips it:

clawock record \
  --subject 00100 --market HK --currency HKD \
  --action reject --confidence 0.65 --driven-by fundamental \
  --bull "营收 +159% YoY,入通/海外霸榜催化" \
  --bear "净利率 -2368%,资不抵债,z+2.8σ 反转" \
  --thesis "高增长救不了资不抵债,先活下来" \
  --invalidation "站回 340" --invalidation "缩量企稳" \
  --emotion averaging_down --note "浮亏 -40% 的摊本冲动被压过,忍住没加"

--source defaults to conversation for DSH; other harnesses pass their
own — OpenClaw --source openclaw, Claude Code --source claude, Codex
--source codex. One ledger, one record command, every harness calls it;
nobody edits decisions.jsonl directly.

Rules: the bear case and at least one observable invalidation condition are
mandatory (the command rejects weak records); be honest about emotion
pressure — fomo/revenge/averaging_down are the exact patterns this
ledger exists to surface. Full schema (in the repository, not in the npm
tarball):
https://github.com/KCNyu/clawock/blob/master/docs/decision-mind-ledger.md

Learn more

Live proof on a real HK + US desk: https://kcnyu.github.io/clawock/
The same contract from other harnesses (OpenClaw skill, Claude Code
instruction, Codex AGENTS.md, pure CLI):
https://github.com/KCNyu/clawock/tree/master/examples/dsh/packages/clawock-dsh