Przejdź do treści

Cooperative Interest Masking

Created: 2026-10-27. Revised: 2026-10-28.

Based on: P013 domain map.

Status: proposed post-MVP feature; no runtime claim. It extends the dormant-fact pull plan and adds no transport.

Summary

A relay that serves bucket pulls learns which node reads which coarse topic, and when. Local decoys blur that, but a node with a real interest can remain the only stable reader of a term over many epochs. Cooperative interest masking lets a node ask selected peers to "also pull X" for a bounded period. It never asks them to "pull X for me": every participant performs its own ordinary pull, keeps its own response and reports nothing back. This is neither brokering nor onion routing.

The promise is bounded. Reads become a weaker indicator of interest. The mechanism does not hide the reader's address or identity, and it does not make it impossible to learn who is really interested.

Context and Problem

  • Fact. The epoch pull plan (P013-DFR-002) mixes real terms with decoys sampled from rated vocabulary terms, and pulls complete slices.
  • Inference. Independent decoys are uncorrelated across nodes and rotate. A real interest persists, so across epochs the persistent reader of a term stands out. Cooperative cover adds other persistent readers of the same term.
  • Constraint. A helper learns what it is asked to read. Requests therefore mix real and decoy terms and carry no motive, so that knowledge stays ambiguous.
  • Constraint. Artifact Delivery leaves durable state (S023). Using it is not traceless; it needs a dedicated retention profile before exposure.

One Plan, Three Origins

Following the reusable-data-abstraction rule of node:DEV-GUIDELINES.md, the DFR plan becomes the general whisper-pull-plan.local.v1 rather than gaining a parallel structure. It has one entry abstraction:

Field Meaning
term A rated vocabulary term id
window The epoch and its pull window
origin need (interest profile), decoy (local draw) or cover (accepted request)
origin/ref The local source: profile revision, draw id or request id
expires-at The end of the entry's validity

origin is local data. The executor ignores it, the relay never sees it, and diagnostics show only aggregate counts per origin. One executor serves every entry: the same relays and mirrors, endpoints, headers, pagination, complete download, retry budget and storage path. A cover pull never aborts early and never carries a marker. Its results enter the ordinary local rumor store under ordinary retention, and whether they were useful is reported to no one.

Cover Requests

whisper-cover-request.v1 is deliberately small:

Field Meaning
schema Constant schema id
request/id A random 128-bit id; the idempotency key per sender
cover/terms 1–8 rated factory vocabulary term ids
cover/from-epoch The first covered epoch
cover/epochs 1–8 whole epochs
schedule/rule The shared timing rule, cover-schedule/v1

The schema has additionalProperties: false and a 2 KiB bound. It deliberately omits a motive, a real-or-decoy marker, the initiator's profile, a result route, forwarding fields, URLs or addresses, and any vocabulary version or digest.

  • Composition. The initiator builds requests from its plan by a rule that includes real terms only probabilistically and also sends requests with no real term. Every term is drawn from the same rated distribution as local decoys.
  • No own signature. Authenticity comes from the authenticated peer session and AD admission, so a helper holds no transferable signed proof of what the initiator asked. Whether INAC session authentication actually has that property must be verified (P013-CIM-002); if not, the claim narrows.
  • No cancellation. Requests expire. A cancellation sent when a real need ends would itself reveal the need.
  • No recursion. A helper never forwards a request, and the schema has no field that could ask it to.

Delivery Through Artifact Delivery

Node A cover component
  -> whisper-cover-request.v1
  -> Artifact Delivery A (private route, inac-direct)
  -> INAC over an authenticated peer session
  -> Artifact Delivery B (admission)
  -> cover component B (single authoritative acceptor):
     local consent, budget, validity, deduplication, prior execution
  -> B's pull plan and S020 job, at a locally drawn time
  -> ordinary Agora topic read; results stay at B

AD is the component interface and inac-direct the carrier, per S023 and node:DEV-GUIDELINES.md. Delivery authority is not control authority. The acceptor checks consent, budget, validity and prior execution before any state change, and it refuses generically, without telling the sender why.

AD records deliveries and admissions durably, and AD does not store arbitrary artifacts in Memarium. The absence of a Memarium entry therefore does not mean the absence of data on disk. Before runtime exposure this consumer requires the minimal-trace delivery profile (S023-001, S023-002). Validity and retention are different things: an expired request is not an erased request.

Shared Timing

Under cover-schedule/v1, every participant, the initiator included, draws for each executable bucket job and epoch an independent uniform offset within the epoch's pull window and fetches at that time. Duplicate planning entries are coalesced first, as specified in the implementation recommendations. Entries of every origin follow the same rule, so a real interest produces no "interested first" or "interested last" pattern. The window bounds and any additional jitter are profile values, tuned by P013-CIM-007.

Multi-Epoch Cover

A request covers whole epochs, and a helper keeps an accepted term for the whole period. The initiator keeps its helper set for a term stable during the term's cover period. It starts cover together with the need and keeps it after the need ends (tail padding). Helpers and terms rotate only by rules evaluated on multi-epoch histories (P013-CIM-007), because drawing new helpers for every read can leave the interested node as the only constant reader.

Participation and Budgets

Participation is voluntary and bounded by P091 settings:

  • initiation and acceptance are separate settings, both off by default;
  • acceptance admits senders from contacts, an allowlist or advertised peers;
  • limits cover active cover terms, bytes per epoch, request duration in epochs and requests per sender per epoch;
  • helper selection reuses AD recipient selectors: explicit node lists, or capability-many with max/nodes against peers advertising whisper.cover.v1. Advertising participation is itself a disclosure and is opt-in.

A helper accepts only rated vocabulary terms: no arbitrary addresses, no recursion, and no obligation beyond the request's expiry.

Not Memory, Not Support

Cover and decoy entries never enter the interest profile, Memarium personal records or link-candidate claims. A cover pull creates no rumor, no confirmation and no source counted toward a coordination threshold or an author disclosure condition. The mechanism masks observable behaviour; it does not falsify personal memory or reports.

Effect Recovery Classes

Effect Class Owner Point of no return
Sending a cover request irreversible-external Initiator cover component through AD AD hands the bytes to INAC
Accepting a request into the plan transactional-withheld Helper cover component The plan commit
A cover pull irreversible-external Whisper pull job The relay observes the read

Effectiveness Must Be Measured

P013-CIM-007 compares cooperative cover with independent decoys at the same byte budget, on multi-epoch histories, against three adversaries:

  1. relays that see node, term and time;
  2. relays together with helpers that reveal the requests they received;
  3. a federation-wide passive observer.

Metrics include how accurately real interests per node can be inferred, the rank of true terms and the number of epochs until identification, across helper counts, request mixes, durations and timing parameters. The documentation may claim only what these measurements show.

Invariants

  1. Each participant pulls for itself; nothing is relayed to the initiator.
  2. origin never leaves the node and never changes how a pull is executed.
  3. Requests carry no motive, no marker and no forwarding instruction.
  4. All participants follow the same timing rule.
  5. Acceptance follows local consent and budgets; delivery grants no control.
  6. Requests expire, are never cancelled and never recurse.
  7. Cover never becomes interest, memory, a rumor or a counted source.
  8. Runtime exposure requires the minimal-trace delivery profile.

Trade-offs

  • Cover costs bandwidth that independent decoys could also spend. Its value must be demonstrated on equal budgets.
  • Helpers learn request contents. Mixing makes that knowledge ambiguous, not absent.
  • Stable helper sets help against persistence analysis but concentrate knowledge in fewer helpers.
  • Tail padding prolongs cost after a need ends.

Failure Modes and Mitigations

Failure Effect Mitigation
A helper reveals requests to an observer The initiator's candidate terms are exposed Mixed requests without motive, no portable signature, evaluation with colluding helpers
A cover pull behaves differently Cover is recognizable One executor, complete download, no markers, the shared timing rule
Rotating helpers every read The interested node remains the only constant reader Stable sets per cover period, tail padding, history-based rotation rules
Requests used for amplification Load on relays or helpers Rated terms only, per-sender and per-epoch budgets, no recursion
Durable AD traces Requests reconstructable from disk Minimal-trace profile with scheduled purge
Cancellation timing Reveals the end of a need No cancellation; expiry only

Open Questions

  1. Default request mix, helper count and cover duration, pending P013-CIM-007.
  2. Whether a deniable request carrier is available on every private-safe adapter, or only on INAC sessions.
  3. Whether helpers should also pad their own acceptance so that refusal under load does not reveal their capacity.

Implementation Recommendations

The shared-mechanism dispositions choose DFR's plan and executor as the single owner. Keep local planning causes separate from executable jobs:

PlanEntry = (term, window, origin, origin_ref, expires_at)
PullJob = (topic_key, window, relay_profile_ref, scheduled_at, budget_ref)

compile_pull_jobs consumes the immutable epoch vocabulary resolution and all admitted entries. It maps terms to topic keys, coalesces identical bucket/window/ relay-profile requests, checks bounded work and byte budgets, and persists the jobs before dispatch. Several origins may explain one job; retain those causes locally within the plan's retention, never as request fields or executor branches. This prevents a need and a cover entry for the same bucket from causing a second revealing fetch. Different windows remain different jobs. Pull-all uses the same listing/pagination executor, not a second acquisition implementation.

Draw one offset per executable job and epoch using cover-schedule/v1, after coalescing, for every origin alike. Save the sampled time: restart and S020 retry must not draw it again. S020 executes the persisted deadline, without an additional jitter layer. An accepted cover cause survives until its original expiry even when a matching local need disappears. Late inputs do not cause an immediate, origin-specific fetch; epoch planning owns their admission to the schedule.

The executor receives only jobs. It uses the same completion, pagination, mirror deduplication and budget handling for all of them. Budget exhaustion, pruning or missing pages yield an explicit incomplete result, not an empty complete slice; there is no early exit because a local match was found. Results enter the ordinary rumor store. The planner never promotes a cover cause into personal interest.

CIM-001/003/004 fixtures cover mixed-origin duplicate buckets, deterministic replay of saved offsets, changed request bytes under one sender/request id, bounded plan growth and expiry without cancellation. Compare emitted requests and completion behavior after erasing local provenance from otherwise equal plans. This proves path equivalence, not anonymity: CIM-007 still owns the multi-epoch effectiveness experiment.

Cooperative Interest Masking Tracker

All rows are post-MVP todo. The grouping row closes on the composed run and the measured claim.

ID Task Status Depends on Completion gate
P013-CIM-000 Cooperative masking acceptance and bounded claim todo P013-CIM-001–008; S023-001–002 A retained three-node run (initiator, two helpers, relay with a mirror) shows identical pull behaviour across origins, expiry without cancellation and purged traces; the claim in this section is restated from P013-CIM-007 results.
P013-CIM-001 Generalize the pull plan todo P013-DFR-002 whisper-pull-plan.local.v1 with one origin-tagged entry abstraction and one executor; origin never leaves the node; fixtures prove identical requests, pagination and completion for every origin. Coalesce mixed-origin duplicate buckets into one job; retain bounded local causes without leaking them or changing completion behavior.
P013-CIM-002 Freeze the cover request contract todo P013-DFR-001 whisper-cover-request.v1 with closed fields and bounds; the composition rule, including requests without real terms; no own signature, no cancellation and no forwarding fields. Verify whether the INAC session gives helpers no transferable proof; negative fixtures for extra fields and unrated terms.
P013-CIM-003 Shared timing rule todo P013-CIM-001; S020 cover-schedule/v1 with independent uniform offsets per compiled job and epoch from a CSPRNG for all origins and the initiator; profile parameters; order-statistic tests showing no stable first or last reader. Persist offsets after coalescing and replay them unchanged; no extra S020 draw or immediate late-origin fetch.
P013-CIM-004 Helper acceptor todo P013-CIM-002; S023-001; P091 One authoritative AD acceptor that checks consent, sender class, budgets, validity, deduplication and prior execution before any state change. It refuses generically, follows a transition table and keeps bounded state purged at expiry, and its tests cover refusal, replay and bounded growth.
P013-CIM-005 Initiator helper selection and multi-epoch maintenance todo P013-CIM-002; S023 selectors Stable helper sets per cover period, tail padding and rotation rules; reuse of AD recipient selectors and the whisper.cover.v1 advertisement; P091 initiation settings.
P013-CIM-006 Separation from memory and social signals todo P013-CIM-001 Fixtures prove that cover and decoy entries never reach the interest profile, Memarium personal records, link candidates, thresholds or disclosure-condition counts.
P013-CIM-007 Effectiveness evaluation todo P013-CIM-001–005 A multi-epoch simulation comparing cooperative cover with independent decoys at equal byte budgets against the three adversaries; parameter sweeps; a recorded bounded claim and recommended defaults.
P013-CIM-008 Operator settings and readiness todo P013-CIM-004–005; P091 Initiation and acceptance settings, budgets and aggregate load per origin in readiness; diagnostics without request contents or sender lists.