# WO-22: DMN outreach, DM sharing, and outcome learning

## Incident

On 2026-09-03 at 17:07 PDT the bot posted in the OMNIUS group without being
addressed. The operator asked whether this was self-induced reflection.

It was not. The post was reactive to two inbound messages (5680, 5681) and
authorized as `direct_turn`. The stated router reason — *"Sender reacts to
bot's recent output in active thread"* — was false: the cited self-bearing
evidence was `turn:5678`, a bot reply from **09-01**, to a different user,
about a different topic.

Two independent defects produced that:

- `buildTelegramInteractionEvidencePacket` selects causal history with
  `.slice(-8)` and **no time bound**. In a quiet room the last 8 messages
  reach back days.
- The `direct_turn` validator requires a cited self-bearing evidence entry but
  checks only *kind* and *actor*, never *recency*. A two-day-old turn satisfied
  it.

Investigating that surfaced the larger finding below.

## The real finding: reflection could not speak

Omnius reflects constantly — **465 daydream artifacts on 2026-09-03 alone** —
and roughly two thirds of recent artifacts proposed a same-group follow-up.
None were ever sent. Across 639 committed routing decisions in this workspace,
`social_intervention` fired exactly **once**; the other 432 replies were all
`direct_turn`.

Root cause was ID plumbing, not judgement:

- The extraction schema advertised `reply_to_message_id: 1` and
  `source_message_ids: [1]` as examples. **78.6% of live follow-up proposals
  echoed a placeholder-like id**, `1` being the single most common value.
- Those ids never matched real Telegram ids, so `candidateMessageIds` was
  almost always empty.
- The discretion prompt then instructed the model to *"cite at least one
  current eligible Telegram message ID"* **without ever listing which ids were
  eligible**, while the host discarded any follow-up citing an id outside that
  undisclosed set.

Silence was the only reachable outcome.

## Changes

| Area | Change |
| --- | --- |
| Extraction prompt | Schema example ids replaced with non-numeric sentinels; explicit rule that ids must appear literally in the corpus; the valid id set is now stated. |
| Extraction result | `runTelegramReflectionExtraction` drops any `source_message_ids` / `reply_to_message_id` absent from the corpus, so a copied or hallucinated anchor cannot reach the follow-up gate. |
| Discretion prompt | Now supplies `Eligible evidence message IDs`, the exact set the host validates against. |
| DM reflection | Idle reflection and outreach admitted for private chats **only after that chat opts in** via `/reflect auto on`. Default remains off. |
| Outcome learning | New `telegram-outreach-outcomes.ts` records every autonomous send and settles it as engaged / ignored from subsequent room traffic. Settled history is fed back into the next discretion call. |

## Design notes

- **DMs are opt-in, not default.** A DM is one person's inbox. The capability
  exists; the host does not enable it on the recipient's behalf.
- **Silence is recorded as a real signal.** An unanswered outreach settles to
  `engaged: false` after 45 minutes rather than staying unknown, so the
  calibration block cannot mistake absence of data for success.
- **The bot's own traffic never counts as engagement.** Only a human message
  inside the window closes an outcome positively.
- **Calibration is neutral before it is negative.** With no settled history the
  prompt says so and asks the model to judge on merit; it raises the bar only
  once every settled attempt was ignored.

## Not addressed

The evidence-packet recency defect that produced the original false
`direct_turn` remains open. Bounding causal history by time would push honest
unaddressed interjections onto the `social_intervention` path, where they are
visible and tunable — but `social_intervention` currently has **no validation
rules at all** (compare five checks for `direct_turn`). Both should be done
together so interjection becomes a governed capability rather than an
unvalidated fallback.

## Verification

CLI suite: 2442 passed, 0 failed (261 files). Typecheck clean; `pnpm -r build`
clean. 14 new tests covering id anchoring, corpus allow-list enforcement,
outcome settlement, calibration wording, torn-ledger recovery, and bridge
wiring.
