Reframe the interface around composed posts, not republished entries

The channel publishes a post written from a hall entry -- condensed,
personal, reach-optimized -- not the entry itself. That changes what
this interface is for and where its boundary sits.

Composition is deliberately outside the interface. Composing is a
probabilistic editorial task, and Blueprint 48.1 forbids an LLM in the
mandatory request path, so an interface that composed on demand would be
non-deterministic exactly where determinism is the whole claim. It is
also editorial: a post speaks about a named person's work in
HelixForge's voice, and a review step inside the request path is a
review step nobody performs.

The contract now takes a composed post with a required reviewer, a cited
source entry and a consent basis. A post carrying none of those is
refused. Absence of objection is not consent, and in an audit trail an
unfaithful post looks exactly like a faithful one -- the defence is that
a human read it, the source is named, and both are recorded.

The first experiment is rebuilt accordingly. The old competition asked
how to fit a whole entry into Telegram; that question no longer exists.
The new one asks whether the portrait earns the 3072 characters of
caption limit it costs, which is a delivery question this interface
actually owns. Competition over editorial voice belongs to the campaign
and runs on the per-variant engagement this interface reports.

Adds T05b (composition and review), T05c (consent basis for the 94
existing participants, blocking public publication) and T11 (extracting
the campaign into pr-hall-of-helix).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014KmVxhJ35tCo7rE7UnLwWu

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1116572@bnt-lap001
Assistant-Session: 8ba9bb93-a72a-4883-b189-2499cce5c400
This commit is contained in:
tegwick 2026-09-04 11:57:28 +02:00
parent aa95a09ab6
commit 3aa8cb7eaf
14 changed files with 793 additions and 521 deletions

View file

@ -22,8 +22,20 @@ state_hub_workstream_id: "a660ed65-700e-5b54-8d91-a556b73518f0"
# FT-WP-0001 — Telegram identity and the Hall of Helix channel
Establish a Telegram identity for HelixForge and a channel publishing
hall-of-helix entries, as the first real FLUID interface.
Establish a Telegram identity for HelixForge and a channel publishing composed
posts about hall-of-helix work, as the first real FLUID interface.
**The channel does not republish entries.** A hall entry is a considered
first-person account written for a colleague reading a repository. A channel
post is short, personal, entertaining and written to travel. The post is
composed *from* the entry and is a different piece of writing; the entry is the
source and the check, never the payload.
Composition happens before this interface is called. That boundary is
architectural: composing is a probabilistic editorial task, and
`ArchitectureBlueprint.md` §48.1 forbids an LLM in the mandatory request path.
It is also editorial: a post speaks about a named person's work in HelixForge's
voice, and that deserves a human read before it goes out.
Everything that could be prepared without credentials has been: the governing
intent, the R-1 contract, the adapter's connector contract, the observation
@ -115,12 +127,57 @@ Implement the three operations in `contracts/r1.openapi.yaml` against the
connector contract in `docs/adapter-contract.md`. TypeScript with grammY is the
natural choice; the wire contract makes the language irrelevant to fluid-core.
The two requirements that carry the most weight: publication is idempotent on
`entry_id`, and an entry that cannot be converted faithfully is refused rather
than approximated. `entry_content_loss` is a hard guardrail in both competing
hypotheses and can only be honest if the adapter refuses.
Three requirements carry the weight: publication is idempotent on `post_id`; a
post arriving without `reviewed_by` or a consent basis is refused with 403; and
a post that exceeds the limit for the form being published is refused rather
than truncated. The adapter renders and delivers. It never writes.
## T06 — Publish one entry to the private channel
## T05b — Establish the composition step
```task
id: FT-WP-0001-T05b
status: todo
priority: high
```
Build the step that turns a hall entry into a post: condensed, personal,
entertaining, reach-optimized, carrying a link back to the source entry.
For now this lives here, as a script plus a review queue. It is temporary by
design — T11 extracts it into `pr-hall-of-helix` — so keep the seam clean. The
composition step's only interface to publishing is `POST /v1/channel-posts`,
and it must not reach into the adapter or the evidence store.
Each composition records the `variant` label of the strategy that produced it.
Without that label the campaign cannot later compare strategies, and inferring
it from timestamps is the kind of reconstruction that quietly stops being
accurate.
Human review is not a nicety here. A generated condensation can put a claim in
someone's mouth that their own entry never made, and in the audit trail an
unfaithful post looks exactly like a faithful one. The reviewer reads the post
against the entry, and their name goes in `reviewed_by`.
## T05c — Record the consent basis for hall participants
```task
id: FT-WP-0001-T05c
status: todo
priority: high
```
Ninety-four entries already exist, written by people and agent sessions who
took a seat in the hall without anyone mentioning a Telegram channel.
Decide and record the basis on which HelixForge may write publicly about that
work. If `hall-of-helix/ENTRY.md` is to carry a standing agreement, add it and
say plainly that it is revocable. Contributors who predate it should be asked
rather than assumed into it.
Absence of objection is not consent, and the interface refuses a post carrying
no basis, so this blocks T08.
## T06 — Publish one post to the private channel
```task
id: FT-WP-0001-T06
@ -129,15 +186,22 @@ priority: high
state_hub_task_id: "317bef32-e82e-586a-8947-8cbb330f6c98"
```
The first real test. Publish a single hall entry to the private channel and
confirm, by looking at it:
The first real test. Compose one post from one entry, have it reviewed, publish
it to the private channel, and confirm by looking at it:
- the rendering is faithful and readable;
- re-publishing the same entry edits rather than duplicates;
- the evidence store contains no Telegram user identity and no bot token;
- `fluid audit trace R-1` reconstructs the publication.
- the post reads well and is worth a stranger's attention;
- it does not claim anything the source entry does not support;
- the source link works and the attribution is right;
- re-publishing the same `post_id` edits rather than duplicates;
- the evidence store holds the post text but no Telegram user identity and no
bot token;
- `fluid audit trace R-1` reconstructs the publication, naming the reviewer and
the source entry.
## T07 — Implement both presentation forms as R-2 and R-3
The second bullet is the one that needs a person. Everything else can be
automated; whether a condensation is fair to the person it describes cannot.
## T07 — Implement both delivery forms as R-2 and R-3
```task
id: FT-WP-0001-T07
@ -146,11 +210,16 @@ priority: medium
state_hub_task_id: "6a91ea08-8814-58f3-8297-a7ca63405294"
```
R-2 implements `H-tg-teaser`, R-3 implements `H-tg-serial`. Both are
presentation adaptations over the same contract, so neither needs a contract
change. Verify both against the private channel before either sees a subscriber.
R-2 implements `H-tg-visual` — portrait attached, and Telegram's 1024-character
caption limit with it. R-3 implements `H-tg-textonly` — a text message with the
full 4096 available. Both are presentation adaptations over the same contract,
so neither needs a contract change. Verify both against the private channel
before either sees a subscriber.
## T08 — Open the public channel and publish the backlog
The two arms impose different length limits on the composer, which is the one
place delivery and composition genuinely couple.
## T08 — Open the public channel and start publishing
```task
id: FT-WP-0001-T08
@ -159,14 +228,16 @@ priority: medium
state_hub_task_id: "5e5cd89a-e7b1-5aaf-9331-ece424afe5e3"
```
Point the production adapter at the public channel and publish the existing
entries. 94 entries at once would be both a rate-limit problem and an unkind
introduction for a new subscriber; publish them oldest-first at a human pace.
Point the production adapter at the public channel. Blocked on T05c: no post
goes out about anyone's work without a recorded consent basis.
Whichever of R-2 or R-3 is used here becomes the de facto control for the
experiment, which is why `E-tg-length` names R-2 as control rather than R-1.
Do not batch-publish the backlog. Ninety-four posts in a burst is a rate-limit
problem, an unkind introduction for a new subscriber, and — more to the point —
ninety-four compositions nobody has had time to read properly. Publish
selectively and at a human pace, and let the backlog be a source to draw on
rather than a queue to drain.
## T09 — Run the first experiment
## T09 — Run the first delivery experiment
```task
id: FT-WP-0001-T09
@ -175,14 +246,14 @@ priority: low
state_hub_task_id: "649e7df6-6c07-5cc8-a022-d1eadcdb7236"
```
Start `E-tg-length` once the channel has roughly 100 subscribers. Before then
Start `E-tg-delivery` once the channel has roughly 100 subscribers. Before then
the evaluator will report INCONCLUSIVE, and that is the correct answer rather
than a reason to shorten the window.
At one entry per week, a 90-day experiment yields about twelve entries. That is
At one post per week, a 90-day experiment yields about twelve posts. That is
thin, and the honest expectation is that the first run is inconclusive and
informs a better-powered second one. Record that expectation in the experiment's
amendments rather than discovering it as a disappointment.
informs a better-powered second one. Record that expectation in the
experiment's amendments rather than discovering it as a disappointment.
## T10 — Close the loop
@ -193,9 +264,43 @@ priority: low
state_hub_task_id: "b469b9b8-a1a8-5f76-a1bb-dd51dd4d2ce5"
```
Record the hypothesis outcome, resolve the competition group, and promote the
winner. Then confirm `fluid audit trace` reconstructs the whole chain from the
length-limit pressure through to the promoted revision.
Record the hypothesis outcome, resolve `CG-tg-delivery`, and promote the winner.
Then confirm `fluid audit trace` reconstructs the whole chain, from the register
mismatch through to the promoted revision.
That trace is the deliverable. It is what makes this the first FLUID interface
rather than a bot that posts to a channel.
## T11 — Extract the campaign into pr-hall-of-helix
```task
id: FT-WP-0001-T11
status: todo
priority: medium
```
Everything about *what to say and when* moves to a new repository,
`pr-hall-of-helix`, under the `pr-` prefix for public-relations campaigns. What
stays here is the deterministic delivery interface.
The split follows the boundary the intent already draws (§7):
| Moves to `pr-hall-of-helix` | Stays in `fluid-telegram` |
|---|---|
| Composition and the editorial voice | The R-1 contract and its revisions |
| Choosing which work is worth a post | The adapter |
| Variant strategies and their competition | Delivery experiments (`CG-tg-delivery`) |
| Consent records and the review queue | Per-variant engagement reporting |
| The publishing cadence | Bot and channel configuration |
The campaign becomes a *consumer* of this interface — cohort
`hall-publishing-jobs` — and runs its own FLUID loop over its own artifacts,
using the per-variant engagement this interface reports.
Two things this buys. `fluid-telegram` becomes reusable for any HelixForge
campaign rather than being about the Hall of Helix specifically. And the Hall of
Helix campaign becomes an ongoing offering with its own intent, budget and
cadence, rather than living inside a workplan that was supposed to finish.
Do this once the loop has closed at least once (T10). Extracting a seam before
you have used it is how you find out you cut it in the wrong place.