v0.1 (Character Blocks, Workchain, Skill-Agent) + v0.2 (dissent block type, agent search) · §3 Character Blocks, §4.3 Vector-Space Proximity, §5.1 Skill-Agent, §14 Reputation

Walking the Tree

A managing director at a Tier-1 institution wants to ask three questions. Have we ever attempted this before? Who on the current project actually thinks we’re going to miss the deadline? What does the staff really think about our pricing change? Each is a question whose answer exists somewhere in the organisation. None of them is currently answerable without a quarter-long detective hunt across systems. HOP makes them into queries.

The situation

Every senior leader in every Tier-1 institution has the same problem. They need to know what their organisation actually thinks. Not the survey result with its 30% response rate and selection bias. Not the Slack consensus filtered through the dominant voices. Not the curated escalation that arrived by way of three lieutenants. The actual texture of belief and dissent across the people doing the work.

The information exists. There is no shortage of opinion in a 50,000-person institution. The substrate doesn’t let the MD query it honestly.

Confluence pages are unsigned. Anyone could have edited them; the version history doesn’t tell you who actually believed the content vs who was assigned to write it down. Slack threads are ephemeral, unstructured, scattered across hundreds of channels, and don’t survive the searcher’s lack of context. Anonymous surveys are honest but lose the trail back to the person you’d want to follow up with. Jira captures workflow, not opinion. Retros happen and then their findings disappear into a deck someone presented at one team’s all-hands and is now in a SharePoint folder no one can find. Email is private. One-on-ones are deniable.

A managing director who wants to know what the org actually thinks has to ask. Which means asking lieutenants. Which means getting a filtered, politically-conscious version of the answer, several days later, after the question has bounced off the org’s defensive surfaces.

The information exists. It is not retrievable.

The example

Three questions a managing director might want to ask on a Tuesday afternoon. None of them is exotic; all of them are currently impossible to answer in less than several weeks.

1. Have we ever tried building a payment integration with Stripe before? — a historical question. Existing systems answer this poorly: there’s no global index of “all the things we ever tried.” The MD has to ask people who might remember, hope they do, and trust that the people who actually objected to the previous attempt are surfaced rather than buried.

2. Who on Project Yellowhammer actually thinks we’re going to miss the Q3 deadline? — a current-state question. Existing systems give the MD the project manager’s status report, which is a rolling forecast that almost certainly does not contain dissent. The engineers who think the deadline is dead don’t put that in writing inside the project tracker; they say it at the pub.

3. What does branch staff actually think about the customer-segment pricing increase? — a distribution-of-opinion question. Existing systems give the MD: an engagement survey from six months ago, a regional manager’s read on the room, and a small handful of escalations from staff who felt strongly enough to write to corporate. The 90% in the middle is invisible.

What HOP does

Five steps. Each is a real protocol operation against the org’s Workchain, not a metaphor.

1. Every interaction signs a block

Code reviews, project decisions, post-mortems, retros, customer interactions, strategy discussions, status updates — each becomes a Character Block (§3), dual-signed by the participants. The block’s context field carries the structured metadata: which project, which customer segment, which decision, what confidence level. The block’s inventory field records what the participant had in their hands when they signed: which document they reviewed, which incident they were responding to, which question they were answering.

This is not new effort imposed on staff. The protocol-layer agent (§3.4) emits the block as a side-effect of existing tooling — the code-review tool already records who reviewed what; the project tool already records who approved which milestone; the customer-interaction system already records who handled which call. The change is that each of these emissions becomes a signed, queryable Character Block instead of an unsigned row in one of nine different proprietary databases.

2. Dissent is first-class

When a participant reviews work, they sign with a stance, not just an approval. The protocol supports approve, approve-with-reservation, dissent, and refuse-to-sign. A dissenting signature carries a dissent_reason field — structured enough to query, free enough to capture the actual objection. Dissent does not get suppressed; it gets attested.

This is the load-bearing affordance for the MD’s second question. Today, an engineer who thinks the project is doomed has no place to record that durably without political cost. Under HOP, they sign as approve-with-reservation or dissent on the milestone block, with a short signed reason, and the record survives.

3. The MD’s Skill-Agent walks the chain

The MD’s question lands in natural language: “Have we ever tried building a payment integration with Stripe?” The Skill-Agent (§5.1, §4.3) translates this into a structured query against the org’s Workchain: find blocks where context.work_class matches payment-integration patterns, filter to chains tagged Stripe, group by project, surface the verdict and the dissenting voices.

The agent does not return a single answer. It returns the structure of the answer: which projects ran, who ran them, what conclusion was signed at the end, which signatures carried dissent, who the dissenters were, what their stated reasons were.

4. The answer surfaces structure

For the three questions above:

Have we tried Stripe before? → Two past attempts found, both abandoned at architecture review. The agent surfaces: project leads, the architecture-review block where each was killed, the two engineers who signed dissent against the kill — their signed reasons preserved verbatim. The MD can read those reasons before greenlighting a third attempt.

Who thinks Project Yellowhammer will miss Q3? → Three engineers signed status blocks with context.confidence: low in the last two weeks. The agent surfaces: each one, the milestone they signed against, the signed reason. The MD reads what they actually wrote.

What does branch staff think about the pricing increase? → The agent walks blocks tagged with the pricing-decision context, retrieves stamps from the staff cohort, surfaces the distribution: how many endorsed, how many dissented, the dissent-reason cluster (extracted via the embedding space — §4.3), and any cross-referenced blocks where customer-success staff signed observations of the actual customer response in the field.

5. Privacy and audit travel together

Per §3.4 selective disclosure: the MD’s query only resolves against blocks the MD has authorisation to read. A junior engineer’s private musings about leaving the company are not in the queryable surface. A protected HR conversation is not. The MD sees what they have a right to see; the protocol enforces this cryptographically.

And: every query the MD runs is itself a Character Block. The CEO who searches who dissents on my strategy is on record as having searched it. Power is more legible inside the protocol than outside it. This asymmetry is intentional — the same affordance that lets a leader see the org’s honest answer also makes the leader’s pattern of inquiry visible to the org. If both directions are not legible, the protocol is doing something wrong.

What’s different

In the world-after, the MD opens their agent and asks a question. The agent walks the chain. The MD reads the structured answer with sources, signatures, dates, and preserved dissent. Total elapsed time: under a minute. The decision the MD makes downstream is grounded in what the organisation actually knows, not in what filtered up the political hierarchy.

Three durable consequences:

The bus factor on institutional knowledge becomes zero. When an engineer leaves the company, what they signed is still on the chain. The next leader wondering why a project failed five years ago can find the signed reasoning of the people who predicted the failure — even if those people are no longer at the company.

Dissent becomes durable. The engineer who said this is a bad idea two years ago can be re-found by the MD now wondering why the bad idea is failing. Their signature stays attached to the warning they gave. The organisation gets a memory that does not require the dissenter to still be present to defend the position they took.

The substrate becomes the source of truth, not the briefing. Today, the brief the MD reads is a representation of what the organisation knows, filtered by people. Under HOP, the brief is generated from the substrate on demand, and the MD can drill from the brief into the underlying signed blocks. The brief is no longer the authoritative artefact; the chain is.

What’s in v0.1 and what’s pending

In v0.1, today — works in the Python reference implementation:

Character Blocks with structured context and inventory fields, dual-signed by participants
Workchain accumulating signed interaction blocks
Per-chain ACLs enforcing selective disclosure
Basic search across blocks via embedding-space proximity

Pending v0.2:

The dissent block type as a first-class subclass of attestation, with dissent_reason and dissent_strength fields
The Skill-Agent’s natural-language query → structured chain-walk translation (the agent code exists in the reference implementation as an interface; the LLM-backed implementation is the v0.2 build target)
Audit-trail surfacing: making each query itself queryable so workers can see who has been looking at them

Pending v0.3:

Federation across peer institutions: an MD at one Tier-1 institution queries shared-industry blocks (e.g. have peer institutions tried this and failed?) under a federation treaty (§7.5), receiving anonymised but signed answers
Differential-privacy guarantees on aggregate queries to staff cohorts (so the MD can ask what does branch staff think without being able to deanonymise individual respondents)

A v0.1 implementer could ship the historical-question case (the MD’s first query above) as a one-month build over the existing code-review and project-tracker telemetry — the chain gets populated by adapters into the existing systems, and a basic Skill-Agent runs natural-language queries via an LLM API. The harder cases — durable dissent and distribution-of-opinion queries — depend on v0.2’s structured dissent block type, which is straightforward to specify and roughly a month of additional implementation work.

A note on the politics

This use case has obvious power-asymmetry implications. The MD who can walk the tree is also a more powerful MD than the one who cannot. Two structural protections matter:

Selective disclosure is per-block, not per-role. A junior engineer’s private blocks are not visible to the MD just because the MD has a senior role; they are visible only if the engineer signed them onto a chain the MD has read access to. The default is narrower than the org’s hierarchy, not broader.
Audit is bidirectional. The MD’s queries are themselves blocks. A worker can see who has been asking about them. The CEO running a who-dissents-on-my-strategy query is on record as having run it. This is intentional friction against the worst use of the affordance.

The protocol does not eliminate the power asymmetry between an MD and an engineer. It makes both sides of the relationship visible to each other. The MD gets the answer; the engineer gets to see that the MD asked. Both gains are real, and both gains are uncomfortable.